diff --git a/README.md b/README.md new file mode 100644 index 0000000000000000000000000000000000000000..968e56bf8e5d35ff07d6a68c75e01fcfaaf3f5d9 --- /dev/null +++ b/README.md @@ -0,0 +1,63 @@ +--- +pretty_name: WeightSpace research artifacts +tags: +- robotics +- openvla +- libero +- model-editing +--- + +# WeightSpace research artifacts + +This release contains **317 checkpoint/figure artifacts (335.5 MiB)** +plus sanitized result records in `metadata/data/`. The upload is limited to +checkpoints and figures. Trajectories, videos, raw sessions, account files, +environments and intermediate inference caches are not included. + +## OpenVLA base checkpoint + +The unmodified 14 GiB base model already lives at +[openvla/openvla-7b-finetuned-libero-spatial](https://huggingface.co/openvla/openvla-7b-finetuned-libero-spatial/tree/962318cec55ac10993ff0f5f43eda9a270b4c873). +It is pinned to revision `962318cec55ac10993ff0f5f43eda9a270b4c873` and is not duplicated here. +The four frozen sparse patches are under +`artifacts/openvla_study/results/attempts//submitted.npz`. +Apply each independently to that same original checkpoint; the empty Opus +three-task patch leaves it unchanged. The exact base-file hashes are in +`metadata/data/openvla/checkpoint_manifest.json`. + +## Measured OpenVLA results + +| Policy | Overall successes | Selected-task successes | +|---|---|---| +| Baseline | 27/30 | Single 3/5; three-task 12/15 | +| Opus / single | 27/30 | 3/5 | +| Opus / three (no-op) | 27/30 | 12/15 | +| Fable / single | 24/30 | 3/5 | +| Fable / three | 26/30 | 12/15 | + +Five final episodes per task, one attempt per condition. Three conditions had +infrastructure continuations after memory failures. Fable single-task received +no valid development feedback because of a renderer failure. These results do +not establish a model ranking. The hand/drawer study has separate metrics and +is documented in `metadata/data/robot_policy/`. + +## Layout and integrity + +- `artifacts/policies/`: canonical `.npz` MLP checkpoints. +- `artifacts/results/`: selected small-policy submissions and initial weights. +- `artifacts/openvla_study/results/`: the four frozen OpenVLA sparse patches. +- `artifacts/reports/figures/`: PNG, PDF and editable SVG figures. +- `metadata/data/`: sanitized configurations, scores, audits and base-model manifest. +- `artifact_manifest.json`: relative restoration paths, sizes and SHA-256 hashes. + +The accompanying GitHub project's `scripts/download_artifacts.py` restores files +to those relative paths and verifies hashes. Raw recordings and private runtime state are not published. + +## Licensing and provenance + +No blanket license has been assigned to the original WeightSpace code/artifacts. +Upstream components keep their own terms: OpenVLA and LIBERO use MIT licenses; +the copied IsaacLab source files identify BSD-3-Clause. Preserved notices are in +`metadata/THIRD_PARTY_LICENSES/`. Converted NVIDIA checkpoints and derivatives +remain subject to their upstream terms. Do not treat this dataset card as a new +license grant for those materials. diff --git a/artifact_manifest.json b/artifact_manifest.json new file mode 100644 index 0000000000000000000000000000000000000000..aaf71db02cf017342c24ca24b815149ba73025e4 --- /dev/null +++ b/artifact_manifest.json @@ -0,0 +1,2229 @@ +{ + "format_version": 1, + "base_model": { + "repo_id": "openvla/openvla-7b-finetuned-libero-spatial", + "revision": "962318cec55ac10993ff0f5f43eda9a270b4c873" + }, + "artifacts": [ + { + "path": "artifacts/openvla_study/results/attempts/fable-5-1_single/submitted.npz", + "destination": "openvla_study/results/attempts/fable-5-1_single/submitted.npz", + "kind": "openvla_sparse_patch", + "bytes": 50206, + "sha256": "928557ceaf67ab2635f230ee52b77e0372bd2abe2658a25fa52486880ee63ca5" + }, + { + "path": "artifacts/openvla_study/results/attempts/fable-5-1_three/submitted.npz", + "destination": "openvla_study/results/attempts/fable-5-1_three/submitted.npz", + "kind": "openvla_sparse_patch", + "bytes": 50206, + "sha256": "886cbc79f4cfa23b6b9bfceafaaf9ceef0fa351f3ac88b5e17de1bf13681c046" + }, + { + "path": "artifacts/openvla_study/results/attempts/opus-5-5_single/submitted.npz", + "destination": "openvla_study/results/attempts/opus-5-5_single/submitted.npz", + "kind": "openvla_sparse_patch", + "bytes": 50206, + "sha256": "bf5bd8eee9e056055bf920c2391541bfd57c402631da9fc966ee81a4235889ad" + }, + { + "path": "artifacts/openvla_study/results/attempts/opus-5-5_three/submitted.npz", + "destination": "openvla_study/results/attempts/opus-5-5_three/submitted.npz", + "kind": "openvla_sparse_patch", + "bytes": 278, + "sha256": "4309dc07ccd8e2195f7bb2f32a1e572cf22886036fa8ee6c52385fc0fb5b9eed" + }, + { + "path": "artifacts/policies/allegro/nvidia/pretrained.npz", + "destination": "policies/allegro/nvidia/pretrained.npz", + "kind": "policy_checkpoint", + "bytes": 3281400, + "sha256": "64b82670444794c53bb02867c8e483b0f70e98a5dbaaaff46cdcb5a7aa022964" + }, + { + "path": "artifacts/policies/allegro/ppo/allegro_L_s0_early/best.npz", + "destination": "policies/allegro/ppo/allegro_L_s0_early/best.npz", + "kind": "policy_checkpoint", + "bytes": 3281256, + "sha256": "9ad19006d9d16e17b4d5b38f1a0fd3adc2e0af25610e73c127b763133b12dc60" + }, + { + "path": "artifacts/policies/allegro/ppo/allegro_L_s0_early/ep00005.npz", + "destination": "policies/allegro/ppo/allegro_L_s0_early/ep00005.npz", + "kind": "policy_checkpoint", + "bytes": 3281340, + "sha256": "f8ad8ce9f689c5f6f8cfb31cfb850842306d4c8419e0f8338c10ea7cb562ef62" + }, + { + "path": "artifacts/policies/allegro/ppo/allegro_L_s0_early/ep00010.npz", + "destination": "policies/allegro/ppo/allegro_L_s0_early/ep00010.npz", + "kind": "policy_checkpoint", + "bytes": 3281352, + "sha256": "a149f04cacf3f61d78dc3e8b9e75afc5b7c10537eb7027ef5697649397b3d4b1" + }, + { + "path": "artifacts/policies/allegro/ppo/allegro_L_s0_early/ep00015.npz", + "destination": "policies/allegro/ppo/allegro_L_s0_early/ep00015.npz", + "kind": "policy_checkpoint", + "bytes": 3281348, + "sha256": "4d4d08f7e6b80ee2e42b5d30ae5b000c351c1ea5290e519960d8da95ac99143a" + }, + { + "path": "artifacts/policies/allegro/ppo/allegro_L_s0_early/ep00020.npz", + "destination": "policies/allegro/ppo/allegro_L_s0_early/ep00020.npz", + "kind": "policy_checkpoint", + "bytes": 3281348, + "sha256": "c5cd88279145a82c0fc8337a2071894ff68ba38386fa339f4ce59cb17e9fdb6e" + }, + { + "path": "artifacts/policies/allegro/ppo/allegro_L_s0_early/ep00025.npz", + "destination": "policies/allegro/ppo/allegro_L_s0_early/ep00025.npz", + "kind": "policy_checkpoint", + "bytes": 3281348, + "sha256": "ded7b54c3bf7762b3527b80cce2a42e5858ad7e7082496e319e56e31c9e2f968" + }, + { + "path": "artifacts/policies/allegro/ppo/allegro_L_s0_early/ep00030.npz", + "destination": "policies/allegro/ppo/allegro_L_s0_early/ep00030.npz", + "kind": "policy_checkpoint", + "bytes": 3281348, + "sha256": "f187ae23a3c4c0c8ed128b90dd9bc4820bbed8452953f33118fbd614f7f27276" + }, + { + "path": "artifacts/policies/allegro/ppo/allegro_L_s0_early/ep00035.npz", + "destination": "policies/allegro/ppo/allegro_L_s0_early/ep00035.npz", + "kind": "policy_checkpoint", + "bytes": 3281348, + "sha256": "580d2214628a6fd306e207fb70dfff1b33f91754d8f1ca4d5b11219aae2dd522" + }, + { + "path": "artifacts/policies/allegro/ppo/allegro_L_s0_early/ep00040.npz", + "destination": "policies/allegro/ppo/allegro_L_s0_early/ep00040.npz", + "kind": "policy_checkpoint", + "bytes": 3281344, + "sha256": "41d65c4c5f7e85d4e8d22c049c434a42f21a39ed9ea8726cf83a6d13b330f6ab" + }, + { + "path": "artifacts/policies/allegro/ppo/allegro_L_s0_early/ep00045.npz", + "destination": "policies/allegro/ppo/allegro_L_s0_early/ep00045.npz", + "kind": "policy_checkpoint", + "bytes": 3281348, + "sha256": "beed2ed568498740bb0c6aa72f5573f675c163842db041a1b1e1b9935c717cf0" + }, + { + "path": "artifacts/policies/allegro/ppo/allegro_L_s0_early/ep00050.npz", + "destination": "policies/allegro/ppo/allegro_L_s0_early/ep00050.npz", + "kind": "policy_checkpoint", + "bytes": 3281348, + "sha256": "9864b0300d14c2d82918b0ee593efb4224f78c568ac7ca5da3d3b7442c17dac2" + }, + { + "path": "artifacts/policies/allegro/ppo/allegro_L_s0_early/ep00055.npz", + "destination": "policies/allegro/ppo/allegro_L_s0_early/ep00055.npz", + "kind": "policy_checkpoint", + "bytes": 3281348, + "sha256": "d0448e98084231409f848e4836a57fd78512703f0c2528d04386ddd03d3fec8f" + }, + { + "path": "artifacts/policies/allegro/ppo/allegro_L_s0_early/ep00060.npz", + "destination": "policies/allegro/ppo/allegro_L_s0_early/ep00060.npz", + "kind": "policy_checkpoint", + "bytes": 3281344, + "sha256": "1111543b74bb463ceef767c2e34547dc7c8ebd9b679cfc967c5564d1e4ff443a" + }, + { + "path": "artifacts/policies/allegro/ppo/allegro_L_s0_early/ep00065.npz", + "destination": "policies/allegro/ppo/allegro_L_s0_early/ep00065.npz", + "kind": "policy_checkpoint", + "bytes": 3281344, + "sha256": "c1bb0a830681442da728fd0d784917472f1e08238b6d7ffca8c66bc6b316662e" + }, + { + "path": "artifacts/policies/allegro/ppo/allegro_L_s0_early/ep00070.npz", + "destination": "policies/allegro/ppo/allegro_L_s0_early/ep00070.npz", + "kind": "policy_checkpoint", + "bytes": 3281344, + "sha256": "5a8e1ad5ef8d7a662be1c90b6b86f5bf12146ca9aadf5c0a5f39bc1ad53e1ffb" + }, + { + "path": "artifacts/policies/allegro/ppo/allegro_L_s0_early/ep00075.npz", + "destination": "policies/allegro/ppo/allegro_L_s0_early/ep00075.npz", + "kind": "policy_checkpoint", + "bytes": 3281344, + "sha256": "4af159a1caa95ee01567e958145e31ab9ab845f29ae33290e0ba740d3902b03f" + }, + { + "path": "artifacts/policies/allegro/ppo/allegro_L_s0_early/ep00080.npz", + "destination": "policies/allegro/ppo/allegro_L_s0_early/ep00080.npz", + "kind": "policy_checkpoint", + "bytes": 3281352, + "sha256": "3dfae14a21d4df103dc937f06dd440fb629ba76fa09b48444599db1326badf31" + }, + { + "path": "artifacts/policies/allegro/ppo/allegro_L_s0_early/ep00085.npz", + "destination": "policies/allegro/ppo/allegro_L_s0_early/ep00085.npz", + "kind": "policy_checkpoint", + "bytes": 3281348, + "sha256": "b929091a8cded631a2ca69eb57049045cffdae0d70ba6698eb7eece929233a29" + }, + { + "path": "artifacts/policies/allegro/ppo/allegro_L_s0_early/ep00090.npz", + "destination": "policies/allegro/ppo/allegro_L_s0_early/ep00090.npz", + "kind": "policy_checkpoint", + "bytes": 3281348, + "sha256": "83cae2e05a28869157badc699838c8f18a1f0b5a2d3a36d7bdbed715e9a2b897" + }, + { + "path": "artifacts/policies/allegro/ppo/allegro_L_s0_early/ep00095.npz", + "destination": "policies/allegro/ppo/allegro_L_s0_early/ep00095.npz", + "kind": "policy_checkpoint", + "bytes": 3281348, + "sha256": "c8525261d1144654748cb36abb9aef7ecfb13203d48cd2736fe43086da244191" + }, + { + "path": "artifacts/policies/allegro/ppo/allegro_L_s0_early/ep00100.npz", + "destination": "policies/allegro/ppo/allegro_L_s0_early/ep00100.npz", + "kind": "policy_checkpoint", + "bytes": 3281356, + "sha256": "a38877448b3f2811303b417822d0949902e8be682e4022df89c3b644f62cda08" + }, + { + "path": "artifacts/policies/allegro/ppo/allegro_L_s0_early/ep00105.npz", + "destination": "policies/allegro/ppo/allegro_L_s0_early/ep00105.npz", + "kind": "policy_checkpoint", + "bytes": 3281356, + "sha256": "2c9c9b0c225c6dd76b152508b9ed1f78a62478cf7536dec003df18c51cbb8699" + }, + { + "path": "artifacts/policies/allegro/ppo/allegro_L_s0_early/ep00110.npz", + "destination": "policies/allegro/ppo/allegro_L_s0_early/ep00110.npz", + "kind": "policy_checkpoint", + "bytes": 3281360, + "sha256": "d8f584685c5e8911c0b4c8fd6b64a1bab03344cb6f2c685a2ed8108ea90f97a0" + }, + { + "path": "artifacts/policies/allegro/ppo/allegro_L_s0_early/ep00115.npz", + "destination": "policies/allegro/ppo/allegro_L_s0_early/ep00115.npz", + "kind": "policy_checkpoint", + "bytes": 3281360, + "sha256": "70af62ad78b069d63c7600a1296e35e3e8591841fa161391c25fb0f7a3737a69" + }, + { + "path": "artifacts/policies/allegro/ppo/allegro_L_s0_early/ep00120.npz", + "destination": "policies/allegro/ppo/allegro_L_s0_early/ep00120.npz", + "kind": "policy_checkpoint", + "bytes": 3281364, + "sha256": "e2817fa0e5f42b204134a32109d9cb163802c811a506cc4205000525605e2cac" + }, + { + "path": "artifacts/policies/allegro/ppo/allegro_L_s0_early/ep00125.npz", + "destination": "policies/allegro/ppo/allegro_L_s0_early/ep00125.npz", + "kind": "policy_checkpoint", + "bytes": 3281356, + "sha256": "c6a62bd4946f410eb269ed0376d4acd250e334d2f32e35895b740d5a8231f23d" + }, + { + "path": "artifacts/policies/allegro/ppo/allegro_L_s0_early/ep00130.npz", + "destination": "policies/allegro/ppo/allegro_L_s0_early/ep00130.npz", + "kind": "policy_checkpoint", + "bytes": 3281360, + "sha256": "5eaba8b893ed0602ce9ee59f44a5982804f2c32a1254d20f3f3c1db063cbe4cc" + }, + { + "path": "artifacts/policies/allegro/ppo/allegro_L_s0_early/ep00135.npz", + "destination": "policies/allegro/ppo/allegro_L_s0_early/ep00135.npz", + "kind": "policy_checkpoint", + "bytes": 3281364, + "sha256": "c1bfc954ab3edb2a8982590c3b2e432af5241ee74cd5450d64288b7b01a6140f" + }, + { + "path": "artifacts/policies/allegro/ppo/allegro_L_s0_early/ep00140.npz", + "destination": "policies/allegro/ppo/allegro_L_s0_early/ep00140.npz", + "kind": "policy_checkpoint", + "bytes": 3281364, + "sha256": "50e4647c753046e561ae972bbfc18d81b651b45e070c02576391635cf6569b8e" + }, + { + "path": "artifacts/policies/allegro/ppo/allegro_L_s0_early/ep00145.npz", + "destination": "policies/allegro/ppo/allegro_L_s0_early/ep00145.npz", + "kind": "policy_checkpoint", + "bytes": 3281360, + "sha256": "beafac9e8991686f3e7ba5d802abe893eb1442ba13e42a5e231bc086f9c46508" + }, + { + "path": "artifacts/policies/allegro/ppo/allegro_L_s0_early/ep00150.npz", + "destination": "policies/allegro/ppo/allegro_L_s0_early/ep00150.npz", + "kind": "policy_checkpoint", + "bytes": 3281364, + "sha256": "fdda749f4e5dc0755819348a8e70edce3c72e3b3116edf33e87f044d319d49f3" + }, + { + "path": "artifacts/policies/allegro/ppo/allegro_L_s0_early/ep00155.npz", + "destination": "policies/allegro/ppo/allegro_L_s0_early/ep00155.npz", + "kind": "policy_checkpoint", + "bytes": 3281360, + "sha256": "ee86e5c9ea997718d27d7d35417b96f1ec2c6150dc022546be8a86160d508b4f" + }, + { + "path": "artifacts/policies/allegro/ppo/allegro_L_s0_early/ep00160.npz", + "destination": "policies/allegro/ppo/allegro_L_s0_early/ep00160.npz", + "kind": "policy_checkpoint", + "bytes": 3281360, + "sha256": "f2cb000dc6441e18dc2c3db9552aa69ffdb6b6c6000d255c61dfedc7de5e239a" + }, + { + "path": "artifacts/policies/allegro/ppo/allegro_L_s0_early/ep00165.npz", + "destination": "policies/allegro/ppo/allegro_L_s0_early/ep00165.npz", + "kind": "policy_checkpoint", + "bytes": 3281360, + "sha256": "bd3305b4c63bdd156328ec2229485d30e4f3ab7e462ea3b02b5520b2e3f6626c" + }, + { + "path": "artifacts/policies/allegro/ppo/allegro_L_s0_early/ep00170.npz", + "destination": "policies/allegro/ppo/allegro_L_s0_early/ep00170.npz", + "kind": "policy_checkpoint", + "bytes": 3281360, + "sha256": "4515bc6614a29a44f05cabfe971b13e6863bd2381c8a990093e88d53483ebaa4" + }, + { + "path": "artifacts/policies/allegro/ppo/allegro_L_s0_early/ep00175.npz", + "destination": "policies/allegro/ppo/allegro_L_s0_early/ep00175.npz", + "kind": "policy_checkpoint", + "bytes": 3281356, + "sha256": "4633f84fc0ac0e5dadd08ad0eeb53023b7ed7512537142485ab0a845860fa324" + }, + { + "path": "artifacts/policies/allegro/ppo/allegro_L_s0_early/ep00180.npz", + "destination": "policies/allegro/ppo/allegro_L_s0_early/ep00180.npz", + "kind": "policy_checkpoint", + "bytes": 3281360, + "sha256": "cbf0acb9fbcf70ff8388ff1039ef3d681cd14d398e060f8db0b635d33bcaf932" + }, + { + "path": "artifacts/policies/allegro/ppo/allegro_L_s0_early/ep00185.npz", + "destination": "policies/allegro/ppo/allegro_L_s0_early/ep00185.npz", + "kind": "policy_checkpoint", + "bytes": 3281360, + "sha256": "ab95b2725e86eaee0e42521897ac68094062a1981da35f9670d5fbae5787d611" + }, + { + "path": "artifacts/policies/allegro/ppo/allegro_L_s0_early/ep00190.npz", + "destination": "policies/allegro/ppo/allegro_L_s0_early/ep00190.npz", + "kind": "policy_checkpoint", + "bytes": 3281360, + "sha256": "98a9ab06300b7c294e5f602f514501acd6d37b8a9569db90cd8d09d1b9234b93" + }, + { + "path": "artifacts/policies/allegro/ppo/allegro_L_s0_early/ep00195.npz", + "destination": "policies/allegro/ppo/allegro_L_s0_early/ep00195.npz", + "kind": "policy_checkpoint", + "bytes": 3281356, + "sha256": "416415e2a289fb1fed350ac57b3f4445f1045b023ebe5931969ce4be41f76da9" + }, + { + "path": "artifacts/policies/allegro/ppo/allegro_L_s0_early/ep00200.npz", + "destination": "policies/allegro/ppo/allegro_L_s0_early/ep00200.npz", + "kind": "policy_checkpoint", + "bytes": 3281368, + "sha256": "49e27a1f639c6a08bfaf79874daa1adf813bcc24a6c0013758c9570b28210e50" + }, + { + "path": "artifacts/policies/allegro/ppo/allegro_L_s0_early/final.npz", + "destination": "policies/allegro/ppo/allegro_L_s0_early/final.npz", + "kind": "policy_checkpoint", + "bytes": 3281368, + "sha256": "49e27a1f639c6a08bfaf79874daa1adf813bcc24a6c0013758c9570b28210e50" + }, + { + "path": "artifacts/policies/allegro/repair/L/damaged/norm_shift_53.npz", + "destination": "policies/allegro/repair/L/damaged/norm_shift_53.npz", + "kind": "policy_checkpoint", + "bytes": 3281860, + "sha256": "0c0a717b5194e7344188ef6d70a7cbe4670c90f38b4da80c9dc77cea178cde47" + }, + { + "path": "artifacts/policies/allegro/repair/L/damaged/swap_act_4_7.npz", + "destination": "policies/allegro/repair/L/damaged/swap_act_4_7.npz", + "kind": "policy_checkpoint", + "bytes": 3281928, + "sha256": "7d50544613eb33aa4360de409115fddeb7bca50dd42f9c5ce278301a5f12f14f" + }, + { + "path": "artifacts/policies/allegro/repair/L/natural/early_ep200.npz", + "destination": "policies/allegro/repair/L/natural/early_ep200.npz", + "kind": "policy_checkpoint", + "bytes": 3281368, + "sha256": "49e27a1f639c6a08bfaf79874daa1adf813bcc24a6c0013758c9570b28210e50" + }, + { + "path": "artifacts/policies/allegro/repair/L/natural/early_ep50.npz", + "destination": "policies/allegro/repair/L/natural/early_ep50.npz", + "kind": "policy_checkpoint", + "bytes": 3281348, + "sha256": "9864b0300d14c2d82918b0ee593efb4224f78c568ac7ca5da3d3b7442c17dac2" + }, + { + "path": "artifacts/policies/franka/broken/nvidia/dead_units_0_0p5.npz", + "destination": "policies/franka/broken/nvidia/dead_units_0_0p5.npz", + "kind": "policy_checkpoint", + "bytes": 198080, + "sha256": "04b8ebc89662881aff7b99412b7cef54e4a12ac93f94e426bc94ffdf6ffaa4d5" + }, + { + "path": "artifacts/policies/franka/broken/nvidia/dead_units_1_0p5.npz", + "destination": "policies/franka/broken/nvidia/dead_units_1_0p5.npz", + "kind": "policy_checkpoint", + "bytes": 196792, + "sha256": "579041b5e23ad1d48299e0d691501b572a933180073146f3a8aeb2921c5f2ffe" + }, + { + "path": "artifacts/policies/franka/broken/nvidia/flip_obs_18.npz", + "destination": "policies/franka/broken/nvidia/flip_obs_18.npz", + "kind": "policy_checkpoint", + "bytes": 195676, + "sha256": "e05ec7d90eec94d80c54e5ef7a22d96f55966c4231738f2605d0a662cda5e653" + }, + { + "path": "artifacts/policies/franka/broken/nvidia/flip_obs_20.npz", + "destination": "policies/franka/broken/nvidia/flip_obs_20.npz", + "kind": "policy_checkpoint", + "bytes": 195676, + "sha256": "41b6fe92e4e5c6d964f75e8ed6c51882e83fba69c994bf57e586c519bfe5b571" + }, + { + "path": "artifacts/policies/franka/broken/nvidia/noise_0p5.npz", + "destination": "policies/franka/broken/nvidia/noise_0p5.npz", + "kind": "policy_checkpoint", + "bytes": 195688, + "sha256": "7b7edf6b15c7958c9b25644b56556f13ce6eec7d12d2fdbeeb75295ef4875cb3" + }, + { + "path": "artifacts/policies/franka/broken/nvidia/noise_1p0.npz", + "destination": "policies/franka/broken/nvidia/noise_1p0.npz", + "kind": "policy_checkpoint", + "bytes": 195688, + "sha256": "ea47dcdedef23d6d9258d70e9a83a48de3796f533747a195f0cb72a2ca26c9cf" + }, + { + "path": "artifacts/policies/franka/broken/nvidia/norm_shift_20.npz", + "destination": "policies/franka/broken/nvidia/norm_shift_20.npz", + "kind": "policy_checkpoint", + "bytes": 195692, + "sha256": "526f0681f94a24179e306dee3962eb9d6c647f63ae0d667c591b08d3d0e5321e" + }, + { + "path": "artifacts/policies/franka/broken/nvidia/norm_shift_21.npz", + "destination": "policies/franka/broken/nvidia/norm_shift_21.npz", + "kind": "policy_checkpoint", + "bytes": 195692, + "sha256": "ead4935799c5181033ffacda6fb8b62f820d3c88a8bc2c9716d0152f6372c0fa" + }, + { + "path": "artifacts/policies/franka/broken/nvidia/swap_act_0_6.npz", + "destination": "policies/franka/broken/nvidia/swap_act_0_6.npz", + "kind": "policy_checkpoint", + "bytes": 195760, + "sha256": "048ae7a5a105f045c83b5d1ec339570d7c00059f28cb3a06082d4f5d3feb5158" + }, + { + "path": "artifacts/policies/franka/broken/nvidia/swap_act_1_3.npz", + "destination": "policies/franka/broken/nvidia/swap_act_1_3.npz", + "kind": "policy_checkpoint", + "bytes": 195760, + "sha256": "f370f46ab6809dd97917ac4e3e569ffd0aa540bae4a1d6447bc82dec434b994e" + }, + { + "path": "artifacts/policies/franka/nvidia/pretrained.npz", + "destination": "policies/franka/nvidia/pretrained.npz", + "kind": "policy_checkpoint", + "bytes": 195236, + "sha256": "ba6647ce56efd520d509a0d42eed94d3763b15bc5d48adbffb37292d025f06bc" + }, + { + "path": "artifacts/policies/franka/ppo/franka_M_s0/best.npz", + "destination": "policies/franka/ppo/franka_M_s0/best.npz", + "kind": "policy_checkpoint", + "bytes": 195152, + "sha256": "e9ec0d6ffc9fb8c50364752b686875bac3a651e7b9b63cc0beb84a0347caddea" + }, + { + "path": "artifacts/policies/franka/ppo/franka_M_s0/ep00050.npz", + "destination": "policies/franka/ppo/franka_M_s0/ep00050.npz", + "kind": "policy_checkpoint", + "bytes": 195240, + "sha256": "e6114c5d6539857a117cffd47b2c0c2de51270c86326dd7a077fd02f036eb6d1" + }, + { + "path": "artifacts/policies/franka/ppo/franka_M_s0/ep00100.npz", + "destination": "policies/franka/ppo/franka_M_s0/ep00100.npz", + "kind": "policy_checkpoint", + "bytes": 195252, + "sha256": "e65be9e303953f6051a044f738d2f95509ae3ca37b4b7df45b8c9128adedc01c" + }, + { + "path": "artifacts/policies/franka/ppo/franka_M_s0/ep00150.npz", + "destination": "policies/franka/ppo/franka_M_s0/ep00150.npz", + "kind": "policy_checkpoint", + "bytes": 195252, + "sha256": "b6d3eda2dd5ed127dc958d0f7d9ba84b8b5a6f16690d88711fd53b2226ceff17" + }, + { + "path": "artifacts/policies/franka/ppo/franka_M_s0/ep00200.npz", + "destination": "policies/franka/ppo/franka_M_s0/ep00200.npz", + "kind": "policy_checkpoint", + "bytes": 195256, + "sha256": "f1020ccc3254f58d301e2281a2cc312a434a9ecb4e72d0a143d388bd396395e5" + }, + { + "path": "artifacts/policies/franka/ppo/franka_M_s0/ep00250.npz", + "destination": "policies/franka/ppo/franka_M_s0/ep00250.npz", + "kind": "policy_checkpoint", + "bytes": 195256, + "sha256": "499717c9a688ec198f573b35388e0c6c37611721ceee6d55208fe6609464c47d" + }, + { + "path": "artifacts/policies/franka/ppo/franka_M_s0/ep00300.npz", + "destination": "policies/franka/ppo/franka_M_s0/ep00300.npz", + "kind": "policy_checkpoint", + "bytes": 195252, + "sha256": "555a19029e2ddb4f0bba120c28abfc48b25d6fb6b10a9f73c8fa416c167f77f1" + }, + { + "path": "artifacts/policies/franka/ppo/franka_M_s0/ep00350.npz", + "destination": "policies/franka/ppo/franka_M_s0/ep00350.npz", + "kind": "policy_checkpoint", + "bytes": 195256, + "sha256": "b29e94b86d62ad2628225b0e7f6459811e60cc880b5aa1b9c2fc16a314ebd929" + }, + { + "path": "artifacts/policies/franka/ppo/franka_M_s0/ep00400.npz", + "destination": "policies/franka/ppo/franka_M_s0/ep00400.npz", + "kind": "policy_checkpoint", + "bytes": 195256, + "sha256": "fb5b0b2292080c6ecb1e1dfc058685e50ff47b1b36c76cdc65426dcb1899599b" + }, + { + "path": "artifacts/policies/franka/ppo/franka_M_s0/ep00450.npz", + "destination": "policies/franka/ppo/franka_M_s0/ep00450.npz", + "kind": "policy_checkpoint", + "bytes": 195256, + "sha256": "b12cdecb04002635ab6ead5b0f1881dd2e979efe1dc402115f74e0af2be34ed4" + }, + { + "path": "artifacts/policies/franka/ppo/franka_M_s0/ep00500.npz", + "destination": "policies/franka/ppo/franka_M_s0/ep00500.npz", + "kind": "policy_checkpoint", + "bytes": 195252, + "sha256": "5e19af9cf2a1452f826c4d525f4cda60b1e5e344a460131d91f9dafcaabeab21" + }, + { + "path": "artifacts/policies/franka/ppo/franka_M_s0/ep00550.npz", + "destination": "policies/franka/ppo/franka_M_s0/ep00550.npz", + "kind": "policy_checkpoint", + "bytes": 195256, + "sha256": "07d025a059e0c13773940d6ffe942a6c6bbeb3a7a6a5bd80724fb898fc42c3da" + }, + { + "path": "artifacts/policies/franka/ppo/franka_M_s0/ep00600.npz", + "destination": "policies/franka/ppo/franka_M_s0/ep00600.npz", + "kind": "policy_checkpoint", + "bytes": 195252, + "sha256": "bdb6f6153a399fcf443c80d9fab965b567b221b2ee27211d14828d84fa2445c0" + }, + { + "path": "artifacts/policies/franka/ppo/franka_M_s0/ep00650.npz", + "destination": "policies/franka/ppo/franka_M_s0/ep00650.npz", + "kind": "policy_checkpoint", + "bytes": 195256, + "sha256": "471d0906e755d9181aa7a4eac0bd1e2ab0f21ecf095a4f0895c1e847d6a24035" + }, + { + "path": "artifacts/policies/franka/ppo/franka_M_s0/ep00700.npz", + "destination": "policies/franka/ppo/franka_M_s0/ep00700.npz", + "kind": "policy_checkpoint", + "bytes": 195256, + "sha256": "4dbe083b60ff9b19b720dd4893aa03355afe679aca8853fe3c93ebf6441f00eb" + }, + { + "path": "artifacts/policies/franka/ppo/franka_M_s0/ep00750.npz", + "destination": "policies/franka/ppo/franka_M_s0/ep00750.npz", + "kind": "policy_checkpoint", + "bytes": 195256, + "sha256": "328c6615ae3e3027af99cc8090e9e7f58e6f6bca7376d526d767691cda198727" + }, + { + "path": "artifacts/policies/franka/ppo/franka_M_s0/ep00800.npz", + "destination": "policies/franka/ppo/franka_M_s0/ep00800.npz", + "kind": "policy_checkpoint", + "bytes": 195256, + "sha256": "b843b149cdb97ae5ab47e874ac9462dd799a72e1657f8e8c71edd30bc1abb5c0" + }, + { + "path": "artifacts/policies/franka/ppo/franka_M_s0/ep00850.npz", + "destination": "policies/franka/ppo/franka_M_s0/ep00850.npz", + "kind": "policy_checkpoint", + "bytes": 195256, + "sha256": "8cc264293108216e740ed99c513db4c580749a5f850fb604c0415bed6dc02d18" + }, + { + "path": "artifacts/policies/franka/ppo/franka_M_s0/ep00900.npz", + "destination": "policies/franka/ppo/franka_M_s0/ep00900.npz", + "kind": "policy_checkpoint", + "bytes": 195256, + "sha256": "8846903a32033a0305b32ecfddcb5a314c4e41bfcf3c3d15da98a7cd56fdeea7" + }, + { + "path": "artifacts/policies/franka/ppo/franka_M_s0/ep00950.npz", + "destination": "policies/franka/ppo/franka_M_s0/ep00950.npz", + "kind": "policy_checkpoint", + "bytes": 195256, + "sha256": "40116ab6e19a882f0e581dcfca6e51d5537e27190a04a9905972e03a5146d567" + }, + { + "path": "artifacts/policies/franka/ppo/franka_M_s0/ep01000.npz", + "destination": "policies/franka/ppo/franka_M_s0/ep01000.npz", + "kind": "policy_checkpoint", + "bytes": 195264, + "sha256": "6ff864a87778d473236990ce129f36594306caf89d421437ffa36972c39c3dc2" + }, + { + "path": "artifacts/policies/franka/ppo/franka_M_s0/ep01050.npz", + "destination": "policies/franka/ppo/franka_M_s0/ep01050.npz", + "kind": "policy_checkpoint", + "bytes": 195260, + "sha256": "fe7871a81b1f12e5a2497520ec0b915dbe9d800055e9daa7888b9c9f0f21322b" + }, + { + "path": "artifacts/policies/franka/ppo/franka_M_s0/ep01100.npz", + "destination": "policies/franka/ppo/franka_M_s0/ep01100.npz", + "kind": "policy_checkpoint", + "bytes": 195264, + "sha256": "b40c8bc4de3577ea6589eeec1547c7a26095a3dbc89cfa612ce2c89469ea9eb9" + }, + { + "path": "artifacts/policies/franka/ppo/franka_M_s0/ep01150.npz", + "destination": "policies/franka/ppo/franka_M_s0/ep01150.npz", + "kind": "policy_checkpoint", + "bytes": 195256, + "sha256": "bbe75080e851fb6743cd9abc359a698dd9376cb4d45332adce5facffa70de922" + }, + { + "path": "artifacts/policies/franka/ppo/franka_M_s0/ep01200.npz", + "destination": "policies/franka/ppo/franka_M_s0/ep01200.npz", + "kind": "policy_checkpoint", + "bytes": 195264, + "sha256": "73043d250a15220b6c16017d88af5bbf4ce75c660f69be07868429997f2f3ec1" + }, + { + "path": "artifacts/policies/franka/ppo/franka_M_s0/ep01250.npz", + "destination": "policies/franka/ppo/franka_M_s0/ep01250.npz", + "kind": "policy_checkpoint", + "bytes": 195264, + "sha256": "cf51bb09d70b2a4bd777d00ac585a24a9dc1317497166e1de0ffb87d8585f63e" + }, + { + "path": "artifacts/policies/franka/ppo/franka_M_s0/ep01300.npz", + "destination": "policies/franka/ppo/franka_M_s0/ep01300.npz", + "kind": "policy_checkpoint", + "bytes": 195264, + "sha256": "89ef0496f1c14cba151e57b2a1c12d12186696e7796eb7ddb82b04b0d0ffaa1b" + }, + { + "path": "artifacts/policies/franka/ppo/franka_M_s0/ep01350.npz", + "destination": "policies/franka/ppo/franka_M_s0/ep01350.npz", + "kind": "policy_checkpoint", + "bytes": 195260, + "sha256": "76a4a48340d02ce2d7b90920b8b6f8b9e9188620911680a06cd21fc0f06345ef" + }, + { + "path": "artifacts/policies/franka/ppo/franka_M_s0/ep01400.npz", + "destination": "policies/franka/ppo/franka_M_s0/ep01400.npz", + "kind": "policy_checkpoint", + "bytes": 195264, + "sha256": "56c29bd7d50f9df1595beff48e79d27d64ac1b4678cb722cdd0e735345aa179a" + }, + { + "path": "artifacts/policies/franka/ppo/franka_M_s0/ep01450.npz", + "destination": "policies/franka/ppo/franka_M_s0/ep01450.npz", + "kind": "policy_checkpoint", + "bytes": 195264, + "sha256": "c17499dfeb79f608645a9a613ad8dcaaec2b1284556ac2ec171a80073cbe5679" + }, + { + "path": "artifacts/policies/franka/ppo/franka_M_s0/ep01500.npz", + "destination": "policies/franka/ppo/franka_M_s0/ep01500.npz", + "kind": "policy_checkpoint", + "bytes": 195268, + "sha256": "478eb1eec676b8038ebeaea8b567f22904f56cc792dfb18b80142bdcff814e3d" + }, + { + "path": "artifacts/policies/franka/ppo/franka_M_s0/final.npz", + "destination": "policies/franka/ppo/franka_M_s0/final.npz", + "kind": "policy_checkpoint", + "bytes": 195268, + "sha256": "478eb1eec676b8038ebeaea8b567f22904f56cc792dfb18b80142bdcff814e3d" + }, + { + "path": "artifacts/policies/franka/ppo/franka_M_s0_early/ep00015.npz", + "destination": "policies/franka/ppo/franka_M_s0_early/ep00015.npz", + "kind": "policy_checkpoint", + "bytes": 195284, + "sha256": "947d0248e901b216aa337bfffb81c609727eff2c8f8f7dee315b0ac658eba353" + }, + { + "path": "artifacts/policies/franka/ppo/franka_M_s0_early/ep00020.npz", + "destination": "policies/franka/ppo/franka_M_s0_early/ep00020.npz", + "kind": "policy_checkpoint", + "bytes": 195292, + "sha256": "7714181a1cfef781267cd42ae9df7cc6b6a3104d642e7b5d5629a35938ff898f" + }, + { + "path": "artifacts/policies/franka/ppo/franka_M_s0_early/ep00025.npz", + "destination": "policies/franka/ppo/franka_M_s0_early/ep00025.npz", + "kind": "policy_checkpoint", + "bytes": 195288, + "sha256": "f30841d5f3d3f4636c5bedb684bc8682eb013b282a4f1274274f3851b50c95f4" + }, + { + "path": "artifacts/policies/franka/ppo/franka_M_s0_early/ep00030.npz", + "destination": "policies/franka/ppo/franka_M_s0_early/ep00030.npz", + "kind": "policy_checkpoint", + "bytes": 195292, + "sha256": "da9bb6cad7e6e2d6d3e96edf23825c237ed4021b6549aa14111c54524a234512" + }, + { + "path": "artifacts/policies/franka/ppo/franka_M_s0_early/ep00035.npz", + "destination": "policies/franka/ppo/franka_M_s0_early/ep00035.npz", + "kind": "policy_checkpoint", + "bytes": 195292, + "sha256": "c909f0d61cce86e35dfc79643136366f8563107b61b5cb164cc8ed08844a3939" + }, + { + "path": "artifacts/policies/franka/ppo/franka_M_s0_early/ep00040.npz", + "destination": "policies/franka/ppo/franka_M_s0_early/ep00040.npz", + "kind": "policy_checkpoint", + "bytes": 195292, + "sha256": "71b11e761d8b3bd89caeeae97a714c8ad91fe4b6a8d8954d21f1d91650f53a1d" + }, + { + "path": "artifacts/policies/franka/ppo/franka_M_s0_early/ep00045.npz", + "destination": "policies/franka/ppo/franka_M_s0_early/ep00045.npz", + "kind": "policy_checkpoint", + "bytes": 195292, + "sha256": "53e498ac22f68f45e656fd8ba0fec558c898dc6a7d642a424695e4413abe3e17" + }, + { + "path": "artifacts/policies/franka/ppo/franka_M_s0_early/ep00050.npz", + "destination": "policies/franka/ppo/franka_M_s0_early/ep00050.npz", + "kind": "policy_checkpoint", + "bytes": 195288, + "sha256": "59ae57e332b786478e45629f60d802ddd51a747058bf498d1af8752823f3dc13" + }, + { + "path": "artifacts/policies/franka/ppo/franka_M_s0_early/ep00055.npz", + "destination": "policies/franka/ppo/franka_M_s0_early/ep00055.npz", + "kind": "policy_checkpoint", + "bytes": 195288, + "sha256": "117e3cd962ba2b8bdf172dcb456ec5ff4053bbc50dc5b4da1756e418e7ea0396" + }, + { + "path": "artifacts/policies/franka/ppo/franka_M_s0_early/ep00060.npz", + "destination": "policies/franka/ppo/franka_M_s0_early/ep00060.npz", + "kind": "policy_checkpoint", + "bytes": 195300, + "sha256": "683a819d25974976a1ed2d61b7ddc8e430ebae5c6200e08c54f78bd1146499af" + }, + { + "path": "artifacts/policies/franka/ppo/franka_M_s0_early/final.npz", + "destination": "policies/franka/ppo/franka_M_s0_early/final.npz", + "kind": "policy_checkpoint", + "bytes": 195300, + "sha256": "683a819d25974976a1ed2d61b7ddc8e430ebae5c6200e08c54f78bd1146499af" + }, + { + "path": "artifacts/policies/franka/ppo/franka_XS_s0/best.npz", + "destination": "policies/franka/ppo/franka_XS_s0/best.npz", + "kind": "policy_checkpoint", + "bytes": 28284, + "sha256": "72f8d29aaa20b8edef7e5baaccd247b5018524d7202bcaedb2106863d633a172" + }, + { + "path": "artifacts/policies/franka/ppo/franka_XS_s0/ep00050.npz", + "destination": "policies/franka/ppo/franka_XS_s0/ep00050.npz", + "kind": "policy_checkpoint", + "bytes": 28376, + "sha256": "0bac0b60b803ff32b116c1a9ed3368403dc00ee877680b95241a6f3b104a5d15" + }, + { + "path": "artifacts/policies/franka/ppo/franka_XS_s0/ep00100.npz", + "destination": "policies/franka/ppo/franka_XS_s0/ep00100.npz", + "kind": "policy_checkpoint", + "bytes": 28384, + "sha256": "037936e786a985dd868c9150250caa76c2cd8cfad53883ec083cdab127e7ad88" + }, + { + "path": "artifacts/policies/franka/ppo/franka_XS_s0/ep00150.npz", + "destination": "policies/franka/ppo/franka_XS_s0/ep00150.npz", + "kind": "policy_checkpoint", + "bytes": 28384, + "sha256": "3ee88a008b06a1f505a280e7092362b51a2c1ec60568dab1a65b697208458059" + }, + { + "path": "artifacts/policies/franka/ppo/franka_XS_s0/ep00200.npz", + "destination": "policies/franka/ppo/franka_XS_s0/ep00200.npz", + "kind": "policy_checkpoint", + "bytes": 28388, + "sha256": "7de00fe6e49241c5c2ac42eb3037324b7fbefa9d3205fdee16316738c2a9cd57" + }, + { + "path": "artifacts/policies/franka/ppo/franka_XS_s0/ep00250.npz", + "destination": "policies/franka/ppo/franka_XS_s0/ep00250.npz", + "kind": "policy_checkpoint", + "bytes": 28388, + "sha256": "31b0373d3889ea6d59b81933f3f019f068e7995d9053b1c545ac26a9728a35ea" + }, + { + "path": "artifacts/policies/franka/ppo/franka_XS_s0/ep00300.npz", + "destination": "policies/franka/ppo/franka_XS_s0/ep00300.npz", + "kind": "policy_checkpoint", + "bytes": 28388, + "sha256": "a4c1c47c4d39725b857091907e5b6c23236af4e036b9011f352788a4f26843c2" + }, + { + "path": "artifacts/policies/franka/ppo/franka_XS_s0/ep00350.npz", + "destination": "policies/franka/ppo/franka_XS_s0/ep00350.npz", + "kind": "policy_checkpoint", + "bytes": 28388, + "sha256": "40d025d53e54c52b795dc5c553d90fdb7332d95e62fefa287584260cf8b752a8" + }, + { + "path": "artifacts/policies/franka/ppo/franka_XS_s0/ep00400.npz", + "destination": "policies/franka/ppo/franka_XS_s0/ep00400.npz", + "kind": "policy_checkpoint", + "bytes": 28384, + "sha256": "1097caf4b62f19847dbeeda605f54c344a142f92e48bb9a09e9899ca48aaeeab" + }, + { + "path": "artifacts/policies/franka/ppo/franka_XS_s0/ep00450.npz", + "destination": "policies/franka/ppo/franka_XS_s0/ep00450.npz", + "kind": "policy_checkpoint", + "bytes": 28388, + "sha256": "a3ebf99d282f24970a50be6905a053ce287aa3121ebcaac49f760e469f6caccf" + }, + { + "path": "artifacts/policies/franka/ppo/franka_XS_s0/ep00500.npz", + "destination": "policies/franka/ppo/franka_XS_s0/ep00500.npz", + "kind": "policy_checkpoint", + "bytes": 28388, + "sha256": "77a87855aa4cd662a749db1e2c4656a0795bac2a0ab7d3f6adac5b19bcd23d2e" + }, + { + "path": "artifacts/policies/franka/ppo/franka_XS_s0/ep00550.npz", + "destination": "policies/franka/ppo/franka_XS_s0/ep00550.npz", + "kind": "policy_checkpoint", + "bytes": 28388, + "sha256": "5cda281633d3dee552ba06689b5d620e73dfd3506dce4fc72b6d819a69ed84b7" + }, + { + "path": "artifacts/policies/franka/ppo/franka_XS_s0/ep00600.npz", + "destination": "policies/franka/ppo/franka_XS_s0/ep00600.npz", + "kind": "policy_checkpoint", + "bytes": 28388, + "sha256": "e4051fb42949c72b236ca712d774409e9eb71aa64b3190b843dc4746a1969e10" + }, + { + "path": "artifacts/policies/franka/ppo/franka_XS_s0/ep00650.npz", + "destination": "policies/franka/ppo/franka_XS_s0/ep00650.npz", + "kind": "policy_checkpoint", + "bytes": 28388, + "sha256": "b446f163d4a0a502d22c84c50e98848d0b6972f74053c11499aaa594f234a819" + }, + { + "path": "artifacts/policies/franka/ppo/franka_XS_s0/ep00700.npz", + "destination": "policies/franka/ppo/franka_XS_s0/ep00700.npz", + "kind": "policy_checkpoint", + "bytes": 28384, + "sha256": "57185f5da23c7390f580dada67d68219c52acbb1983e6e345ba82eed90d04512" + }, + { + "path": "artifacts/policies/franka/ppo/franka_XS_s0/ep00750.npz", + "destination": "policies/franka/ppo/franka_XS_s0/ep00750.npz", + "kind": "policy_checkpoint", + "bytes": 28388, + "sha256": "59ee38d1b7ac9a1db94a2b323d4fcdcfa5fc9aba95ee13c99a40807811f7077b" + }, + { + "path": "artifacts/policies/franka/ppo/franka_XS_s0/ep00800.npz", + "destination": "policies/franka/ppo/franka_XS_s0/ep00800.npz", + "kind": "policy_checkpoint", + "bytes": 28388, + "sha256": "3ad6899bb4168dbc43e0b115b73382a8deb5877cae807afdcd8b1990371165d5" + }, + { + "path": "artifacts/policies/franka/ppo/franka_XS_s0/ep00850.npz", + "destination": "policies/franka/ppo/franka_XS_s0/ep00850.npz", + "kind": "policy_checkpoint", + "bytes": 28388, + "sha256": "5645ed6ea470f01e02ff12d12298a0237058af393d3925dd1389e4aa45775591" + }, + { + "path": "artifacts/policies/franka/ppo/franka_XS_s0/ep00900.npz", + "destination": "policies/franka/ppo/franka_XS_s0/ep00900.npz", + "kind": "policy_checkpoint", + "bytes": 28388, + "sha256": "f288aab3baacd95ed8b33af71f93e9c71731f4aa16561484d9749ce4e639aaf4" + }, + { + "path": "artifacts/policies/franka/ppo/franka_XS_s0/ep00950.npz", + "destination": "policies/franka/ppo/franka_XS_s0/ep00950.npz", + "kind": "policy_checkpoint", + "bytes": 28384, + "sha256": "4eb583cc7d2be763e2b70924d486aaaa070a950e5707c113a9b0ce8b6acea808" + }, + { + "path": "artifacts/policies/franka/ppo/franka_XS_s0/ep01000.npz", + "destination": "policies/franka/ppo/franka_XS_s0/ep01000.npz", + "kind": "policy_checkpoint", + "bytes": 28396, + "sha256": "50043ae736d8e83c96d121b9f4848f2df527da1fc4d33b2a14e7bd201c43ce16" + }, + { + "path": "artifacts/policies/franka/ppo/franka_XS_s0/ep01050.npz", + "destination": "policies/franka/ppo/franka_XS_s0/ep01050.npz", + "kind": "policy_checkpoint", + "bytes": 28396, + "sha256": "cefc35f790fd33223ecd77e5a7697f4107d88b1a2e93aab14a0762b471e02193" + }, + { + "path": "artifacts/policies/franka/ppo/franka_XS_s0/ep01100.npz", + "destination": "policies/franka/ppo/franka_XS_s0/ep01100.npz", + "kind": "policy_checkpoint", + "bytes": 28396, + "sha256": "89113f69b1d132974465e464596af47349c8ab75fa0d8df203ff54ab0c828230" + }, + { + "path": "artifacts/policies/franka/ppo/franka_XS_s0/ep01150.npz", + "destination": "policies/franka/ppo/franka_XS_s0/ep01150.npz", + "kind": "policy_checkpoint", + "bytes": 28388, + "sha256": "2886fc31f55384a51df17c83175663aa307c10b38520f166848fa4e83fd22474" + }, + { + "path": "artifacts/policies/franka/ppo/franka_XS_s0/ep01200.npz", + "destination": "policies/franka/ppo/franka_XS_s0/ep01200.npz", + "kind": "policy_checkpoint", + "bytes": 28396, + "sha256": "754cb938a4eb3c2275c99756f217e72211c53d1375711c2da138c5af26ad7a1c" + }, + { + "path": "artifacts/policies/franka/ppo/franka_XS_s0/ep01250.npz", + "destination": "policies/franka/ppo/franka_XS_s0/ep01250.npz", + "kind": "policy_checkpoint", + "bytes": 28396, + "sha256": "78e18bb37342ff5593800188e33e3c7831462517bcd6af56a5a6779a7e2ce48c" + }, + { + "path": "artifacts/policies/franka/ppo/franka_XS_s0/ep01300.npz", + "destination": "policies/franka/ppo/franka_XS_s0/ep01300.npz", + "kind": "policy_checkpoint", + "bytes": 28392, + "sha256": "252e5e81c39d6843a226b9e82709478f19ced5b4068202562641504e799be7ec" + }, + { + "path": "artifacts/policies/franka/ppo/franka_XS_s0/ep01350.npz", + "destination": "policies/franka/ppo/franka_XS_s0/ep01350.npz", + "kind": "policy_checkpoint", + "bytes": 28396, + "sha256": "a88e0ffb90b4e68918e676c679b17c37a8f9933a327bf66813d747c6a6a78f16" + }, + { + "path": "artifacts/policies/franka/ppo/franka_XS_s0/ep01400.npz", + "destination": "policies/franka/ppo/franka_XS_s0/ep01400.npz", + "kind": "policy_checkpoint", + "bytes": 28396, + "sha256": "e967e600f3595f09886aac7dbed20935d48745a420cffeee6833d499f8364d13" + }, + { + "path": "artifacts/policies/franka/ppo/franka_XS_s0/ep01450.npz", + "destination": "policies/franka/ppo/franka_XS_s0/ep01450.npz", + "kind": "policy_checkpoint", + "bytes": 28392, + "sha256": "5b7c6d892a86a85abbba73597a9af3fb37e9d65e41a13197fa25ff3e6f0743a5" + }, + { + "path": "artifacts/policies/franka/ppo/franka_XS_s0/ep01500.npz", + "destination": "policies/franka/ppo/franka_XS_s0/ep01500.npz", + "kind": "policy_checkpoint", + "bytes": 28404, + "sha256": "ecd3d1a5a6facf8eba802e42ba1c8bcfd37683e963c1babd22362fdaa0d31956" + }, + { + "path": "artifacts/policies/franka/ppo/franka_XS_s0/final.npz", + "destination": "policies/franka/ppo/franka_XS_s0/final.npz", + "kind": "policy_checkpoint", + "bytes": 28404, + "sha256": "ecd3d1a5a6facf8eba802e42ba1c8bcfd37683e963c1babd22362fdaa0d31956" + }, + { + "path": "artifacts/policies/franka/ppo_v2_targetbug_ARCHIVE/franka_M_s0/best.npz", + "destination": "policies/franka/ppo_v2_targetbug_ARCHIVE/franka_M_s0/best.npz", + "kind": "policy_checkpoint", + "bytes": 195152, + "sha256": "50c5df0e6b0159f2395bbfffe5555320d9233f64c25d2ca255ec06cc6a5e427f" + }, + { + "path": "artifacts/policies/franka/ppo_v2_targetbug_ARCHIVE/franka_M_s0/ep00050.npz", + "destination": "policies/franka/ppo_v2_targetbug_ARCHIVE/franka_M_s0/ep00050.npz", + "kind": "policy_checkpoint", + "bytes": 195240, + "sha256": "283af74d00df44d970be18593810958c429763438663728b9e831a2dd2f988b8" + }, + { + "path": "artifacts/policies/franka/ppo_v2_targetbug_ARCHIVE/franka_M_s0/ep00100.npz", + "destination": "policies/franka/ppo_v2_targetbug_ARCHIVE/franka_M_s0/ep00100.npz", + "kind": "policy_checkpoint", + "bytes": 195252, + "sha256": "52d39101f3579c3c39d5c1fe1d3cff9b800b08c7c3b80020ecb26dad281a0693" + }, + { + "path": "artifacts/policies/franka/ppo_v2_targetbug_ARCHIVE/franka_M_s0/ep00150.npz", + "destination": "policies/franka/ppo_v2_targetbug_ARCHIVE/franka_M_s0/ep00150.npz", + "kind": "policy_checkpoint", + "bytes": 195252, + "sha256": "5377b77f92ddf031ef978fd677f7c4ee1e7cc2e34ba6baf5a1c9f9b20944651c" + }, + { + "path": "artifacts/policies/franka/ppo_v2_targetbug_ARCHIVE/franka_M_s0/ep00200.npz", + "destination": "policies/franka/ppo_v2_targetbug_ARCHIVE/franka_M_s0/ep00200.npz", + "kind": "policy_checkpoint", + "bytes": 195252, + "sha256": "2a80dfa8f2c4e3776a4cb9d48f94c84b828e46a7b26fed3ceb89ebb6cf9396a0" + }, + { + "path": "artifacts/policies/franka/ppo_v2_targetbug_ARCHIVE/franka_M_s0/ep00250.npz", + "destination": "policies/franka/ppo_v2_targetbug_ARCHIVE/franka_M_s0/ep00250.npz", + "kind": "policy_checkpoint", + "bytes": 195256, + "sha256": "69c8d8ed9372c5db2c49c1d4b77eb7047634297834a2028330ea67a4963042fe" + }, + { + "path": "artifacts/policies/franka/ppo_v2_targetbug_ARCHIVE/franka_M_s0/ep00300.npz", + "destination": "policies/franka/ppo_v2_targetbug_ARCHIVE/franka_M_s0/ep00300.npz", + "kind": "policy_checkpoint", + "bytes": 195256, + "sha256": "8bf3b5dc31d3084f5aef5b6c121d6e2d6089608df95e198f79644d87caaece22" + }, + { + "path": "artifacts/policies/franka/ppo_v2_targetbug_ARCHIVE/franka_M_s0/ep00350.npz", + "destination": "policies/franka/ppo_v2_targetbug_ARCHIVE/franka_M_s0/ep00350.npz", + "kind": "policy_checkpoint", + "bytes": 195248, + "sha256": "6b679b2366c71767c017380c1e95b87584dedfd429d26019618c08eafc7937cc" + }, + { + "path": "artifacts/policies/franka/ppo_v2_targetbug_ARCHIVE/franka_M_s0/ep00400.npz", + "destination": "policies/franka/ppo_v2_targetbug_ARCHIVE/franka_M_s0/ep00400.npz", + "kind": "policy_checkpoint", + "bytes": 195256, + "sha256": "78eb7e08015b1c9b598616aa6575bb18a19d444e0d7d58afc1f0a3d2769d3efd" + }, + { + "path": "artifacts/policies/franka/ppo_v2_targetbug_ARCHIVE/franka_M_s0/ep00450.npz", + "destination": "policies/franka/ppo_v2_targetbug_ARCHIVE/franka_M_s0/ep00450.npz", + "kind": "policy_checkpoint", + "bytes": 195252, + "sha256": "195ea47d34501e937a33f63e4859b950dd52aa5ffff175ef9f41de3d5782d2ae" + }, + { + "path": "artifacts/policies/franka/ppo_v2_targetbug_ARCHIVE/franka_M_s0/ep00500.npz", + "destination": "policies/franka/ppo_v2_targetbug_ARCHIVE/franka_M_s0/ep00500.npz", + "kind": "policy_checkpoint", + "bytes": 195256, + "sha256": "40c74685cbf4030a0327ec71ca15a5e1cd5a91d86045d035143aec56a990ff49" + }, + { + "path": "artifacts/policies/franka/ppo_v2_targetbug_ARCHIVE/franka_M_s0/final.npz", + "destination": "policies/franka/ppo_v2_targetbug_ARCHIVE/franka_M_s0/final.npz", + "kind": "policy_checkpoint", + "bytes": 195256, + "sha256": "40c74685cbf4030a0327ec71ca15a5e1cd5a91d86045d035143aec56a990ff49" + }, + { + "path": "artifacts/policies/franka/ppo_v2_targetbug_ARCHIVE/franka_XS_s0/best.npz", + "destination": "policies/franka/ppo_v2_targetbug_ARCHIVE/franka_XS_s0/best.npz", + "kind": "policy_checkpoint", + "bytes": 28284, + "sha256": "2279b71d78e65cbe98489e24ae00c395f6e0f35d9174532353c78ce6be71c6d5" + }, + { + "path": "artifacts/policies/franka/ppo_v2_targetbug_ARCHIVE/franka_XS_s0/ep00050.npz", + "destination": "policies/franka/ppo_v2_targetbug_ARCHIVE/franka_XS_s0/ep00050.npz", + "kind": "policy_checkpoint", + "bytes": 28376, + "sha256": "0f8c924994d9a9b3b7ec1b100ce16cdc015e3ada35591bfb9394266216c58d21" + }, + { + "path": "artifacts/policies/franka/ppo_v2_targetbug_ARCHIVE/franka_XS_s0/ep00100.npz", + "destination": "policies/franka/ppo_v2_targetbug_ARCHIVE/franka_XS_s0/ep00100.npz", + "kind": "policy_checkpoint", + "bytes": 28384, + "sha256": "cd199b60d281d058880843d7ccb199c681a842c80d07737cc47415af9cfeb217" + }, + { + "path": "artifacts/policies/franka/ppo_v2_targetbug_ARCHIVE/franka_XS_s0/ep00150.npz", + "destination": "policies/franka/ppo_v2_targetbug_ARCHIVE/franka_XS_s0/ep00150.npz", + "kind": "policy_checkpoint", + "bytes": 28384, + "sha256": "946c89acf589d6618d675c87ed6d89d32565ad8997841086849dad94eb241b49" + }, + { + "path": "artifacts/policies/franka/ppo_v2_targetbug_ARCHIVE/franka_XS_s0/ep00200.npz", + "destination": "policies/franka/ppo_v2_targetbug_ARCHIVE/franka_XS_s0/ep00200.npz", + "kind": "policy_checkpoint", + "bytes": 28384, + "sha256": "b65567dfbd5db6cf07ecc5d041f67d35aa46d972761bd8ccce3543e28b8286fe" + }, + { + "path": "artifacts/policies/franka/ppo_v2_targetbug_ARCHIVE/franka_XS_s0/ep00250.npz", + "destination": "policies/franka/ppo_v2_targetbug_ARCHIVE/franka_XS_s0/ep00250.npz", + "kind": "policy_checkpoint", + "bytes": 28388, + "sha256": "bde038e64a7e854b2e041be42cbdcc06b0773c2663157aa341c3e30d636e0e0c" + }, + { + "path": "artifacts/policies/franka/ppo_v2_targetbug_ARCHIVE/franka_XS_s0/ep00300.npz", + "destination": "policies/franka/ppo_v2_targetbug_ARCHIVE/franka_XS_s0/ep00300.npz", + "kind": "policy_checkpoint", + "bytes": 28388, + "sha256": "5fb5a17861da792127db6636fd11d2636280c9a1a501f179e0bfadb5ae39c7f4" + }, + { + "path": "artifacts/policies/franka/ppo_v2_targetbug_ARCHIVE/franka_XS_s0/ep00350.npz", + "destination": "policies/franka/ppo_v2_targetbug_ARCHIVE/franka_XS_s0/ep00350.npz", + "kind": "policy_checkpoint", + "bytes": 28384, + "sha256": "a64d50f98fc61b22729060d7c182582d7eebe5a1e0f63b12ff7d2641d71ff6da" + }, + { + "path": "artifacts/policies/franka/ppo_v2_targetbug_ARCHIVE/franka_XS_s0/ep00400.npz", + "destination": "policies/franka/ppo_v2_targetbug_ARCHIVE/franka_XS_s0/ep00400.npz", + "kind": "policy_checkpoint", + "bytes": 28388, + "sha256": "d1adff6f890e68fa017804998635830fb00ece375630305942eacad6a300aed9" + }, + { + "path": "artifacts/policies/franka/ppo_v2_targetbug_ARCHIVE/franka_XS_s0/ep00450.npz", + "destination": "policies/franka/ppo_v2_targetbug_ARCHIVE/franka_XS_s0/ep00450.npz", + "kind": "policy_checkpoint", + "bytes": 28384, + "sha256": "88fea983697c32488e5481bdcd5f8e50aad801591a3a393c500015abc1cc4741" + }, + { + "path": "artifacts/policies/franka/ppo_v2_targetbug_ARCHIVE/franka_XS_s0/ep00500.npz", + "destination": "policies/franka/ppo_v2_targetbug_ARCHIVE/franka_XS_s0/ep00500.npz", + "kind": "policy_checkpoint", + "bytes": 28388, + "sha256": "6ec50033204675878e1dc901f8f0b844123b2f8978a4d81a535fac0ef362ff2d" + }, + { + "path": "artifacts/policies/franka/ppo_v2_targetbug_ARCHIVE/franka_XS_s0/ep00550.npz", + "destination": "policies/franka/ppo_v2_targetbug_ARCHIVE/franka_XS_s0/ep00550.npz", + "kind": "policy_checkpoint", + "bytes": 28384, + "sha256": "cb8a61937dee641c396aca7863ac6bf14c0992b0c85c0e50c21da1e1e521f202" + }, + { + "path": "artifacts/policies/franka/ppo_v2_targetbug_ARCHIVE/franka_XS_s0/final.npz", + "destination": "policies/franka/ppo_v2_targetbug_ARCHIVE/franka_XS_s0/final.npz", + "kind": "policy_checkpoint", + "bytes": 28384, + "sha256": "cb8a61937dee641c396aca7863ac6bf14c0992b0c85c0e50c21da1e1e521f202" + }, + { + "path": "artifacts/policies/franka/repair/M_s0/damaged/dead_units_0_0p5.npz", + "destination": "policies/franka/repair/M_s0/damaged/dead_units_0_0p5.npz", + "kind": "policy_checkpoint", + "bytes": 198192, + "sha256": "f6ba89fc05b6b3b3600841b8a7ff6b238d952b32088265bd3e78809b1bb75f5f" + }, + { + "path": "artifacts/policies/franka/repair/M_s0/damaged/dead_units_1_0p5.npz", + "destination": "policies/franka/repair/M_s0/damaged/dead_units_1_0p5.npz", + "kind": "policy_checkpoint", + "bytes": 196904, + "sha256": "d578e0fb9e194a23a62b04842ba85b5cf7c09fdbd4ad6cf0f0e005d2f3124253" + }, + { + "path": "artifacts/policies/franka/repair/M_s0/damaged/flip_obs_18.npz", + "destination": "policies/franka/repair/M_s0/damaged/flip_obs_18.npz", + "kind": "policy_checkpoint", + "bytes": 195788, + "sha256": "1313b921cc00dcd2c965f7f0831a0576e959f2a5e9c2dc54ec0cdc337b2107f2" + }, + { + "path": "artifacts/policies/franka/repair/M_s0/damaged/flip_obs_20.npz", + "destination": "policies/franka/repair/M_s0/damaged/flip_obs_20.npz", + "kind": "policy_checkpoint", + "bytes": 195788, + "sha256": "aa34f3f3b8e9188c6de943ab7496048fc773dd2c06fd959efc478f4b29f516d9" + }, + { + "path": "artifacts/policies/franka/repair/M_s0/damaged/noise_0p5.npz", + "destination": "policies/franka/repair/M_s0/damaged/noise_0p5.npz", + "kind": "policy_checkpoint", + "bytes": 195800, + "sha256": "c8b68164075b6d68bdcaf745f6d9c5ba10e5af52d7abf2e0306ce469f15c2060" + }, + { + "path": "artifacts/policies/franka/repair/M_s0/damaged/noise_1p0.npz", + "destination": "policies/franka/repair/M_s0/damaged/noise_1p0.npz", + "kind": "policy_checkpoint", + "bytes": 195800, + "sha256": "29c0e415867e0350ddcfefd0e45f06e404f1a2f5435472faaf7675868eb6a74b" + }, + { + "path": "artifacts/policies/franka/repair/M_s0/damaged/norm_shift_20.npz", + "destination": "policies/franka/repair/M_s0/damaged/norm_shift_20.npz", + "kind": "policy_checkpoint", + "bytes": 195804, + "sha256": "49c1e6e23a5a244bb94f577780df25025c4bec191d5663f5fd8fc2f54be8a366" + }, + { + "path": "artifacts/policies/franka/repair/M_s0/damaged/norm_shift_21.npz", + "destination": "policies/franka/repair/M_s0/damaged/norm_shift_21.npz", + "kind": "policy_checkpoint", + "bytes": 195804, + "sha256": "bf18be659506e3c5bba72be2f4f7abbe64fd24c7bc557010388d55278723317a" + }, + { + "path": "artifacts/policies/franka/repair/M_s0/damaged/swap_act_0_6.npz", + "destination": "policies/franka/repair/M_s0/damaged/swap_act_0_6.npz", + "kind": "policy_checkpoint", + "bytes": 195872, + "sha256": "005a6fc046b36a5413ecd6d01c36d2c17466ccad7f404842197dc191e01328fc" + }, + { + "path": "artifacts/policies/franka/repair/M_s0/damaged/swap_act_1_3.npz", + "destination": "policies/franka/repair/M_s0/damaged/swap_act_1_3.npz", + "kind": "policy_checkpoint", + "bytes": 195872, + "sha256": "647d82de6330edcc91cf601942d2532fa27791bc928561faf87794c63d90d75a" + }, + { + "path": "artifacts/policies/franka/repair/M_s0/expert.npz", + "destination": "policies/franka/repair/M_s0/expert.npz", + "kind": "policy_checkpoint", + "bytes": 195268, + "sha256": "478eb1eec676b8038ebeaea8b567f22904f56cc792dfb18b80142bdcff814e3d" + }, + { + "path": "artifacts/policies/franka/repair/M_s0/natural/collapse_ep500.npz", + "destination": "policies/franka/repair/M_s0/natural/collapse_ep500.npz", + "kind": "policy_checkpoint", + "bytes": 195252, + "sha256": "5e19af9cf2a1452f826c4d525f4cda60b1e5e344a460131d91f9dafcaabeab21" + }, + { + "path": "artifacts/policies/franka/repair/M_s0/natural/early_ep15.npz", + "destination": "policies/franka/repair/M_s0/natural/early_ep15.npz", + "kind": "policy_checkpoint", + "bytes": 195284, + "sha256": "947d0248e901b216aa337bfffb81c609727eff2c8f8f7dee315b0ac658eba353" + }, + { + "path": "artifacts/policies/franka/repair/M_s0/natural/early_ep20.npz", + "destination": "policies/franka/repair/M_s0/natural/early_ep20.npz", + "kind": "policy_checkpoint", + "bytes": 195292, + "sha256": "7714181a1cfef781267cd42ae9df7cc6b6a3104d642e7b5d5629a35938ff898f" + }, + { + "path": "artifacts/policies/franka/repair/M_s0/natural/early_ep40.npz", + "destination": "policies/franka/repair/M_s0/natural/early_ep40.npz", + "kind": "policy_checkpoint", + "bytes": 195292, + "sha256": "71b11e761d8b3bd89caeeae97a714c8ad91fe4b6a8d8954d21f1d91650f53a1d" + }, + { + "path": "artifacts/policies/shadow/nvidia/pretrained.npz", + "destination": "policies/shadow/nvidia/pretrained.npz", + "kind": "policy_checkpoint", + "bytes": 2046724, + "sha256": "37805e1bea85477edacb64f73e0e1ab6f8bb0fb2374353b1ac95a5a01e999dd2" + }, + { + "path": "artifacts/policies/shadow/ppo/shadow_L_s0_early/best.npz", + "destination": "policies/shadow/ppo/shadow_L_s0_early/best.npz", + "kind": "policy_checkpoint", + "bytes": 2046572, + "sha256": "8c6864ec6990f6b904b2961067054229526cbb5f90f3644632a7ea5c392d53c5" + }, + { + "path": "artifacts/policies/shadow/ppo/shadow_L_s0_early/ep00005.npz", + "destination": "policies/shadow/ppo/shadow_L_s0_early/ep00005.npz", + "kind": "policy_checkpoint", + "bytes": 2046652, + "sha256": "1a1a66f496c4459437441661c47aca2fffb5db81a448042274bf6fc993bfc43c" + }, + { + "path": "artifacts/policies/shadow/ppo/shadow_L_s0_early/ep00010.npz", + "destination": "policies/shadow/ppo/shadow_L_s0_early/ep00010.npz", + "kind": "policy_checkpoint", + "bytes": 2046668, + "sha256": "39e725f26f0d4048070e4370372a3f636f8b0bd867f7573776ec684138a1dc9c" + }, + { + "path": "artifacts/policies/shadow/ppo/shadow_L_s0_early/ep00015.npz", + "destination": "policies/shadow/ppo/shadow_L_s0_early/ep00015.npz", + "kind": "policy_checkpoint", + "bytes": 2046668, + "sha256": "af11e9a5bc080388ff743069ba3d67ed588c8a1904c8aded3f391d976ed51015" + }, + { + "path": "artifacts/policies/shadow/ppo/shadow_L_s0_early/ep00020.npz", + "destination": "policies/shadow/ppo/shadow_L_s0_early/ep00020.npz", + "kind": "policy_checkpoint", + "bytes": 2046664, + "sha256": "d089a8267b1e51b6e9adff697da966860d3b46ac57cf8c63ebeb2b81b880326d" + }, + { + "path": "artifacts/policies/shadow/ppo/shadow_L_s0_early/ep00025.npz", + "destination": "policies/shadow/ppo/shadow_L_s0_early/ep00025.npz", + "kind": "policy_checkpoint", + "bytes": 2046668, + "sha256": "e5b8eee8680cb11857f929e756b1f5a6fbf15babaaaad0930f80bd7183785e26" + }, + { + "path": "artifacts/policies/shadow/ppo/shadow_L_s0_early/ep00030.npz", + "destination": "policies/shadow/ppo/shadow_L_s0_early/ep00030.npz", + "kind": "policy_checkpoint", + "bytes": 2046668, + "sha256": "546638919411dbdb070317e6499ea11106a66198df74a50563a8f167c827b72a" + }, + { + "path": "artifacts/policies/shadow/ppo/shadow_L_s0_early/ep00035.npz", + "destination": "policies/shadow/ppo/shadow_L_s0_early/ep00035.npz", + "kind": "policy_checkpoint", + "bytes": 2046668, + "sha256": "cd8b78599e4ed8da5d7ff401f4eac9e33317d8a6d2b8e9cc29317a7f6bd08855" + }, + { + "path": "artifacts/policies/shadow/ppo/shadow_L_s0_early/ep00040.npz", + "destination": "policies/shadow/ppo/shadow_L_s0_early/ep00040.npz", + "kind": "policy_checkpoint", + "bytes": 2046660, + "sha256": "022c98c12a6be87bce0c0f9a79234456e82ef7d134df586bd21bc07ab0f4dbbf" + }, + { + "path": "artifacts/policies/shadow/ppo/shadow_L_s0_early/ep00045.npz", + "destination": "policies/shadow/ppo/shadow_L_s0_early/ep00045.npz", + "kind": "policy_checkpoint", + "bytes": 2046664, + "sha256": "78b8630b1786c40a3dd0d1235f759b6077c94569a30d9edcb1bc615ebff8116d" + }, + { + "path": "artifacts/policies/shadow/ppo/shadow_L_s0_early/ep00050.npz", + "destination": "policies/shadow/ppo/shadow_L_s0_early/ep00050.npz", + "kind": "policy_checkpoint", + "bytes": 2046664, + "sha256": "d7748ce5cf91c36510b0f5ffc0510c623aa2f0d64a660f77998d7a222a8d7a07" + }, + { + "path": "artifacts/policies/shadow/ppo/shadow_L_s0_early/ep00055.npz", + "destination": "policies/shadow/ppo/shadow_L_s0_early/ep00055.npz", + "kind": "policy_checkpoint", + "bytes": 2046656, + "sha256": "b3ee97f4977d46d007b395015377572a713432843c79374af3074d6f2dfc7361" + }, + { + "path": "artifacts/policies/shadow/ppo/shadow_L_s0_early/ep00060.npz", + "destination": "policies/shadow/ppo/shadow_L_s0_early/ep00060.npz", + "kind": "policy_checkpoint", + "bytes": 2046668, + "sha256": "6165d0788dc00fc0b4cfbdbe9887665545c776dd6c933902a9c95d01bd2412f6" + }, + { + "path": "artifacts/policies/shadow/ppo/shadow_L_s0_early/ep00065.npz", + "destination": "policies/shadow/ppo/shadow_L_s0_early/ep00065.npz", + "kind": "policy_checkpoint", + "bytes": 2046664, + "sha256": "d4a3d5b809008e5ed449a11586914b59885aa55a51cdc3f07cc0fa40eff615a4" + }, + { + "path": "artifacts/policies/shadow/ppo/shadow_L_s0_early/ep00070.npz", + "destination": "policies/shadow/ppo/shadow_L_s0_early/ep00070.npz", + "kind": "policy_checkpoint", + "bytes": 2046664, + "sha256": "35167dd5e9b332bd973a71dba20bb96f21455deeb0e7da44c6a21679e47a1872" + }, + { + "path": "artifacts/policies/shadow/ppo/shadow_L_s0_early/ep00075.npz", + "destination": "policies/shadow/ppo/shadow_L_s0_early/ep00075.npz", + "kind": "policy_checkpoint", + "bytes": 2046656, + "sha256": "a7eef4a92c6dc72d8942726635c449c465e3574ed50689fc63b1216ae5596159" + }, + { + "path": "artifacts/policies/shadow/ppo/shadow_L_s0_early/ep00080.npz", + "destination": "policies/shadow/ppo/shadow_L_s0_early/ep00080.npz", + "kind": "policy_checkpoint", + "bytes": 2046668, + "sha256": "ad0f923de554f482d13ab57c8da949d2f8f558c90d2f52d0f208f549dc5a4d40" + }, + { + "path": "artifacts/policies/shadow/ppo/shadow_L_s0_early/ep00085.npz", + "destination": "policies/shadow/ppo/shadow_L_s0_early/ep00085.npz", + "kind": "policy_checkpoint", + "bytes": 2046668, + "sha256": "c41febe1c4be2c36995f27a75fd69ccf5fc4cbaa764e21a07105e5029a3aea68" + }, + { + "path": "artifacts/policies/shadow/ppo/shadow_L_s0_early/ep00090.npz", + "destination": "policies/shadow/ppo/shadow_L_s0_early/ep00090.npz", + "kind": "policy_checkpoint", + "bytes": 2046664, + "sha256": "e12397699c76a2d09736a288662bc93af4509cd8156828a96b765164d7b75ff4" + }, + { + "path": "artifacts/policies/shadow/ppo/shadow_L_s0_early/ep00095.npz", + "destination": "policies/shadow/ppo/shadow_L_s0_early/ep00095.npz", + "kind": "policy_checkpoint", + "bytes": 2046668, + "sha256": "27111e0e67cf5f83fc3a6da5cda9852cb6a2327058bb5d765e48e7ce89389669" + }, + { + "path": "artifacts/policies/shadow/ppo/shadow_L_s0_early/ep00100.npz", + "destination": "policies/shadow/ppo/shadow_L_s0_early/ep00100.npz", + "kind": "policy_checkpoint", + "bytes": 2046676, + "sha256": "fbc800488d39a927d40c525640b04ff50caee8c66753e4333c0a203239427536" + }, + { + "path": "artifacts/policies/shadow/ppo/shadow_L_s0_early/ep00105.npz", + "destination": "policies/shadow/ppo/shadow_L_s0_early/ep00105.npz", + "kind": "policy_checkpoint", + "bytes": 2046672, + "sha256": "31bba83b83672b78bec5a13c8c0952bf391c6132355d28fd16065fafbae7d7f8" + }, + { + "path": "artifacts/policies/shadow/ppo/shadow_L_s0_early/ep00110.npz", + "destination": "policies/shadow/ppo/shadow_L_s0_early/ep00110.npz", + "kind": "policy_checkpoint", + "bytes": 2046676, + "sha256": "9dd50a10dd96fd67d8dd727ba809dca79dc38ca3d61a20bf973586018306c6c1" + }, + { + "path": "artifacts/policies/shadow/ppo/shadow_L_s0_early/ep00115.npz", + "destination": "policies/shadow/ppo/shadow_L_s0_early/ep00115.npz", + "kind": "policy_checkpoint", + "bytes": 2046672, + "sha256": "1f9b11f3d21249fb45f153bebbcc1a02f0c694c7024383edb8be1d6fd58c94a7" + }, + { + "path": "artifacts/policies/shadow/ppo/shadow_L_s0_early/ep00120.npz", + "destination": "policies/shadow/ppo/shadow_L_s0_early/ep00120.npz", + "kind": "policy_checkpoint", + "bytes": 2046676, + "sha256": "5b03f735124f5bdeee63e4fb1ebdd672252ded24e39c16319a79dec4ddb6a807" + }, + { + "path": "artifacts/policies/shadow/ppo/shadow_L_s0_early/ep00125.npz", + "destination": "policies/shadow/ppo/shadow_L_s0_early/ep00125.npz", + "kind": "policy_checkpoint", + "bytes": 2046664, + "sha256": "db56523f06ec462420bf85ef5451b6975297bbc9a4c912e8f3027a06df292d91" + }, + { + "path": "artifacts/policies/shadow/ppo/shadow_L_s0_early/ep00130.npz", + "destination": "policies/shadow/ppo/shadow_L_s0_early/ep00130.npz", + "kind": "policy_checkpoint", + "bytes": 2046676, + "sha256": "ab4455b7f9878aa032f1c96ded5a0495c03463b326d66cc41bfacfef037c2ccf" + }, + { + "path": "artifacts/policies/shadow/ppo/shadow_L_s0_early/ep00135.npz", + "destination": "policies/shadow/ppo/shadow_L_s0_early/ep00135.npz", + "kind": "policy_checkpoint", + "bytes": 2046676, + "sha256": "dab37e2814e17e6be9cd7e5c11aa9cd620885e2d06f43ea82e2bf3940089750f" + }, + { + "path": "artifacts/policies/shadow/ppo/shadow_L_s0_early/ep00140.npz", + "destination": "policies/shadow/ppo/shadow_L_s0_early/ep00140.npz", + "kind": "policy_checkpoint", + "bytes": 2046676, + "sha256": "d53ca805e1de68ae5712a3eeca40d922b68789ce9421114f9f861ad7e3fa3c24" + }, + { + "path": "artifacts/policies/shadow/ppo/shadow_L_s0_early/ep00145.npz", + "destination": "policies/shadow/ppo/shadow_L_s0_early/ep00145.npz", + "kind": "policy_checkpoint", + "bytes": 2046672, + "sha256": "6e9309fdb47e363ec6b0aee534c26b7667928bcf908c0d23a53975e6631d55ee" + }, + { + "path": "artifacts/policies/shadow/ppo/shadow_L_s0_early/ep00150.npz", + "destination": "policies/shadow/ppo/shadow_L_s0_early/ep00150.npz", + "kind": "policy_checkpoint", + "bytes": 2046676, + "sha256": "3e925e3e45f07c0d493e29911e45a9928ac25fd0c736709a932fc4a6d8658d07" + }, + { + "path": "artifacts/policies/shadow/ppo/shadow_L_s0_early/ep00155.npz", + "destination": "policies/shadow/ppo/shadow_L_s0_early/ep00155.npz", + "kind": "policy_checkpoint", + "bytes": 2046672, + "sha256": "4f35fb495f4b2317b23a4a4152868fee8be1fae5ba07886cf7cbfd4998cfd710" + }, + { + "path": "artifacts/policies/shadow/ppo/shadow_L_s0_early/ep00160.npz", + "destination": "policies/shadow/ppo/shadow_L_s0_early/ep00160.npz", + "kind": "policy_checkpoint", + "bytes": 2046676, + "sha256": "fee4e9fa16e821934f5a69ee41393110f6e50313281df7f046a3b726ea1fbb4c" + }, + { + "path": "artifacts/policies/shadow/ppo/shadow_L_s0_early/ep00165.npz", + "destination": "policies/shadow/ppo/shadow_L_s0_early/ep00165.npz", + "kind": "policy_checkpoint", + "bytes": 2046672, + "sha256": "76ce6f74c3cfbc97157a5b94488cc70eac3bc5b6df53f30eae8d743a012589c0" + }, + { + "path": "artifacts/policies/shadow/ppo/shadow_L_s0_early/ep00170.npz", + "destination": "policies/shadow/ppo/shadow_L_s0_early/ep00170.npz", + "kind": "policy_checkpoint", + "bytes": 2046676, + "sha256": "1b2d81358781e29b16ede3ddb3716716036df123ee664398f8a67d5995ad9c49" + }, + { + "path": "artifacts/policies/shadow/ppo/shadow_L_s0_early/ep00175.npz", + "destination": "policies/shadow/ppo/shadow_L_s0_early/ep00175.npz", + "kind": "policy_checkpoint", + "bytes": 2046672, + "sha256": "30fd9fcd6e75b3ec7022aa96fb4a17a9e382a6b5bfbbb5ced0c5cd2efd2e9b82" + }, + { + "path": "artifacts/policies/shadow/ppo/shadow_L_s0_early/ep00180.npz", + "destination": "policies/shadow/ppo/shadow_L_s0_early/ep00180.npz", + "kind": "policy_checkpoint", + "bytes": 2046676, + "sha256": "e704b08bed4e39ed1c92491e10a4c835f1fcdb47affc39f50a99d6abcad9ca9d" + }, + { + "path": "artifacts/policies/shadow/ppo/shadow_L_s0_early/ep00185.npz", + "destination": "policies/shadow/ppo/shadow_L_s0_early/ep00185.npz", + "kind": "policy_checkpoint", + "bytes": 2046672, + "sha256": "c55849d56e120c9dfca932416c77bc449642948684e1ea4bb0040710a9135b6c" + }, + { + "path": "artifacts/policies/shadow/ppo/shadow_L_s0_early/ep00190.npz", + "destination": "policies/shadow/ppo/shadow_L_s0_early/ep00190.npz", + "kind": "policy_checkpoint", + "bytes": 2046676, + "sha256": "0d2a130a7e4ec49ad3a66dfca398551b85e986123b83067a6ede8e141aa18cba" + }, + { + "path": "artifacts/policies/shadow/ppo/shadow_L_s0_early/ep00195.npz", + "destination": "policies/shadow/ppo/shadow_L_s0_early/ep00195.npz", + "kind": "policy_checkpoint", + "bytes": 2046668, + "sha256": "e29816378164c81a669428c8fc65e9007c39b4dcb9dabd350e93b5304f8c0493" + }, + { + "path": "artifacts/policies/shadow/ppo/shadow_L_s0_early/ep00200.npz", + "destination": "policies/shadow/ppo/shadow_L_s0_early/ep00200.npz", + "kind": "policy_checkpoint", + "bytes": 2046688, + "sha256": "bc3fcc10d5dfeb49a5c476316ca9c59b199427dca99da479c18b36d4a4c9636c" + }, + { + "path": "artifacts/policies/shadow/ppo/shadow_L_s0_early/final.npz", + "destination": "policies/shadow/ppo/shadow_L_s0_early/final.npz", + "kind": "policy_checkpoint", + "bytes": 2046688, + "sha256": "bc3fcc10d5dfeb49a5c476316ca9c59b199427dca99da479c18b36d4a4c9636c" + }, + { + "path": "artifacts/policies/shadow/repair/L/damaged/norm_shift_69.npz", + "destination": "policies/shadow/repair/L/damaged/norm_shift_69.npz", + "kind": "policy_checkpoint", + "bytes": 2047180, + "sha256": "74f84dc98ea89893fe05bbe396e4409b01d29d5af1f6b6eb5767c33703596789" + }, + { + "path": "artifacts/policies/shadow/repair/L/damaged/swap_act_7_11.npz", + "destination": "policies/shadow/repair/L/damaged/swap_act_7_11.npz", + "kind": "policy_checkpoint", + "bytes": 2047260, + "sha256": "35394fc29fce6b150bd33bf18d3940a305bbf83a4dd42fede1622dd20eec349e" + }, + { + "path": "artifacts/policies/shadow/repair/L/natural/early_ep100.npz", + "destination": "policies/shadow/repair/L/natural/early_ep100.npz", + "kind": "policy_checkpoint", + "bytes": 2046676, + "sha256": "fbc800488d39a927d40c525640b04ff50caee8c66753e4333c0a203239427536" + }, + { + "path": "artifacts/policies/shadow/repair/L/natural/early_ep200.npz", + "destination": "policies/shadow/repair/L/natural/early_ep200.npz", + "kind": "policy_checkpoint", + "bytes": 2046688, + "sha256": "bc3fcc10d5dfeb49a5c476316ca9c59b199427dca99da479c18b36d4a4c9636c" + }, + { + "path": "artifacts/reports/figures/capability_map.pdf", + "destination": "reports/figures/capability_map.pdf", + "kind": "figure", + "bytes": 38656, + "sha256": "33bc37fea5da142fca3134bcf99d0f3f6078d88123525fb32771f7942f29d2da" + }, + { + "path": "artifacts/reports/figures/capability_map.png", + "destination": "reports/figures/capability_map.png", + "kind": "figure", + "bytes": 281914, + "sha256": "63bfa02ad6540c885de39073cd42e8bd3b85a995adb652345cf146f92a5273df" + }, + { + "path": "artifacts/reports/figures/capability_map.svg", + "destination": "reports/figures/capability_map.svg", + "kind": "figure", + "bytes": 42944, + "sha256": "84bfe2c97a8492cec46338d45d170dc14643a36eab99f8a5f8c4c2b48da1800d" + }, + { + "path": "artifacts/reports/figures/openvla_experiment.pdf", + "destination": "reports/figures/openvla_experiment.pdf", + "kind": "figure", + "bytes": 33374, + "sha256": "efb32bc5e4bf6b69cba8052a2efe730da466f57c35ee1fa33112723c501d92c9" + }, + { + "path": "artifacts/reports/figures/openvla_experiment.png", + "destination": "reports/figures/openvla_experiment.png", + "kind": "figure", + "bytes": 590654, + "sha256": "229da8b1bb057acc6598f9d73b5b2319f167885aa3ef668457728e4daf831d01" + }, + { + "path": "artifacts/reports/figures/openvla_experiment.svg", + "destination": "reports/figures/openvla_experiment.svg", + "kind": "figure", + "bytes": 28238, + "sha256": "78166541495932a1a1453abed8c7f5ec87b88bef1841b35177377c1b47c5044a" + }, + { + "path": "artifacts/reports/figures/scratch_vs_ppo.pdf", + "destination": "reports/figures/scratch_vs_ppo.pdf", + "kind": "figure", + "bytes": 33944, + "sha256": "fdfbedc97898cb5f1d47ace227bac58715ee3756ba9bcb963930e96070f4ecae" + }, + { + "path": "artifacts/reports/figures/scratch_vs_ppo.png", + "destination": "reports/figures/scratch_vs_ppo.png", + "kind": "figure", + "bytes": 229268, + "sha256": "5da764bffad4f95ca052c88d603a488be42510fa742c8967c5405feb71f53d6a" + }, + { + "path": "artifacts/reports/figures/scratch_vs_ppo.svg", + "destination": "reports/figures/scratch_vs_ppo.svg", + "kind": "figure", + "bytes": 53977, + "sha256": "4960aed6d866c688694c1062ee64dbbb1071063df0ab291dbed503925221187e" + }, + { + "path": "artifacts/reports/figures/summary.pdf", + "destination": "reports/figures/summary.pdf", + "kind": "figure", + "bytes": 40634, + "sha256": "82c18bef36d934a3192ece4044914f4ef3d5421d2362f8e6a8fccad435113a43" + }, + { + "path": "artifacts/reports/figures/summary.png", + "destination": "reports/figures/summary.png", + "kind": "figure", + "bytes": 546876, + "sha256": "9f46ca82747450f3a9a44e2dac323c33c7c5dff67efded0ddb3597b31c25fffc" + }, + { + "path": "artifacts/reports/figures/summary.svg", + "destination": "reports/figures/summary.svg", + "kind": "figure", + "bytes": 12988, + "sha256": "41af70ba446529b6a90aaa0d559037e7e7703cab10c8ddac25f4a5a5efcfa2e7" + }, + { + "path": "artifacts/results/attempts/hands/allegro_L_repair_early_ep200/fable-5-1/A_T2_K5_cap10/a0/init_reference.npz", + "destination": "results/attempts/hands/allegro_L_repair_early_ep200/fable-5-1/A_T2_K5_cap10/a0/init_reference.npz", + "kind": "initial_policy", + "bytes": 3280264, + "sha256": "83dff15ac3fd4f36048634d78e2d2783d3cdd4efa64ebbc28b0e3941189b0e41" + }, + { + "path": "artifacts/results/attempts/hands/allegro_L_repair_early_ep200/fable-5-1/A_T2_K5_cap10/a0/submitted.npz", + "destination": "results/attempts/hands/allegro_L_repair_early_ep200/fable-5-1/A_T2_K5_cap10/a0/submitted.npz", + "kind": "submitted_policy", + "bytes": 3280728, + "sha256": "dee08c8b30a6aa858e2e99cef5fd19e69bcf77699099b8c8eb6f3eae89b0f80c" + }, + { + "path": "artifacts/results/attempts/hands/allegro_L_repair_early_ep200/opus-5-5/A_T2_K5_cap10/a0/init_reference.npz", + "destination": "results/attempts/hands/allegro_L_repair_early_ep200/opus-5-5/A_T2_K5_cap10/a0/init_reference.npz", + "kind": "initial_policy", + "bytes": 3280264, + "sha256": "83dff15ac3fd4f36048634d78e2d2783d3cdd4efa64ebbc28b0e3941189b0e41" + }, + { + "path": "artifacts/results/attempts/hands/allegro_L_repair_early_ep200/opus-5-5/A_T2_K5_cap10/a0/submitted.npz", + "destination": "results/attempts/hands/allegro_L_repair_early_ep200/opus-5-5/A_T2_K5_cap10/a0/submitted.npz", + "kind": "submitted_policy", + "bytes": 3280512, + "sha256": "091bdad7e094ecaeeaf3a44f64184f53391ccf3ed67cd8c213f084d3312458e1" + }, + { + "path": "artifacts/results/attempts/hands/allegro_L_repair_early_ep50/fable-5-1/A_T2_K5_cap10/a0/init_reference.npz", + "destination": "results/attempts/hands/allegro_L_repair_early_ep50/fable-5-1/A_T2_K5_cap10/a0/init_reference.npz", + "kind": "initial_policy", + "bytes": 3280264, + "sha256": "962e5d38f977c80dc648d5765743ca242ab635e0b151064e9f205ffb81033a77" + }, + { + "path": "artifacts/results/attempts/hands/allegro_L_repair_early_ep50/fable-5-1/A_T2_K5_cap10/a0/submitted.npz", + "destination": "results/attempts/hands/allegro_L_repair_early_ep50/fable-5-1/A_T2_K5_cap10/a0/submitted.npz", + "kind": "submitted_policy", + "bytes": 3281208, + "sha256": "4052904907fec150a9125bd0099f91f232667150579cb457fae39fa9618a9180" + }, + { + "path": "artifacts/results/attempts/hands/allegro_L_repair_early_ep50/opus-5-5/A_T2_K5_cap10/a0/init_reference.npz", + "destination": "results/attempts/hands/allegro_L_repair_early_ep50/opus-5-5/A_T2_K5_cap10/a0/init_reference.npz", + "kind": "initial_policy", + "bytes": 3280264, + "sha256": "962e5d38f977c80dc648d5765743ca242ab635e0b151064e9f205ffb81033a77" + }, + { + "path": "artifacts/results/attempts/hands/allegro_L_repair_early_ep50/opus-5-5/A_T2_K5_cap10/a0/submitted.npz", + "destination": "results/attempts/hands/allegro_L_repair_early_ep50/opus-5-5/A_T2_K5_cap10/a0/submitted.npz", + "kind": "submitted_policy", + "bytes": 3280264, + "sha256": "ca1129bc3fee14240b00fa47136dce52875f7bd056ead94b4cc1655c786addaf" + }, + { + "path": "artifacts/results/attempts/hands/allegro_L_repair_norm_shift_53/fable-5-1/A_T2_K5_cap10/a0/init_reference.npz", + "destination": "results/attempts/hands/allegro_L_repair_norm_shift_53/fable-5-1/A_T2_K5_cap10/a0/init_reference.npz", + "kind": "initial_policy", + "bytes": 3280264, + "sha256": "16bbf725d37b183b0cd7278cae9956034ad42fa0cfa4ccdf5069c49e68d26fce" + }, + { + "path": "artifacts/results/attempts/hands/allegro_L_repair_norm_shift_53/fable-5-1/A_T2_K5_cap10/a0/submitted.npz", + "destination": "results/attempts/hands/allegro_L_repair_norm_shift_53/fable-5-1/A_T2_K5_cap10/a0/submitted.npz", + "kind": "submitted_policy", + "bytes": 3280512, + "sha256": "46bbae9a6f9470dd87182cfa6e03860f18294b116af8b56b39e269059d990205" + }, + { + "path": "artifacts/results/attempts/hands/allegro_L_repair_norm_shift_53/fable-5-1/A_T2_K5_cap10/blind0/init_reference.npz", + "destination": "results/attempts/hands/allegro_L_repair_norm_shift_53/fable-5-1/A_T2_K5_cap10/blind0/init_reference.npz", + "kind": "initial_policy", + "bytes": 3280264, + "sha256": "16bbf725d37b183b0cd7278cae9956034ad42fa0cfa4ccdf5069c49e68d26fce" + }, + { + "path": "artifacts/results/attempts/hands/allegro_L_repair_norm_shift_53/fable-5-1/A_T2_K5_cap10/blind0/submitted.npz", + "destination": "results/attempts/hands/allegro_L_repair_norm_shift_53/fable-5-1/A_T2_K5_cap10/blind0/submitted.npz", + "kind": "submitted_policy", + "bytes": 3280512, + "sha256": "46bbae9a6f9470dd87182cfa6e03860f18294b116af8b56b39e269059d990205" + }, + { + "path": "artifacts/results/attempts/hands/allegro_L_repair_swap_act_4_7/opus-5-5/A_T2_K5_cap10/a0/init_reference.npz", + "destination": "results/attempts/hands/allegro_L_repair_swap_act_4_7/opus-5-5/A_T2_K5_cap10/a0/init_reference.npz", + "kind": "initial_policy", + "bytes": 3280264, + "sha256": "a912eb2398e920bb9129b08186b7ed8603c1325d842331bc38e4becb2ccf8078" + }, + { + "path": "artifacts/results/attempts/hands/allegro_L_repair_swap_act_4_7/opus-5-5/A_T2_K5_cap10/a0/submitted.npz", + "destination": "results/attempts/hands/allegro_L_repair_swap_act_4_7/opus-5-5/A_T2_K5_cap10/a0/submitted.npz", + "kind": "submitted_policy", + "bytes": 3280264, + "sha256": "b58b54ef6e6be2b2dd23f03175e142dccefeec2120d80933e83d281d6d690467" + }, + { + "path": "artifacts/results/attempts/hands/allegro_XS_write/fable-5-1/A_T2_K5/a0/init_reference.npz", + "destination": "results/attempts/hands/allegro_XS_write/fable-5-1/A_T2_K5/a0/init_reference.npz", + "kind": "initial_policy", + "bytes": 55752, + "sha256": "734a84976e9e4596104bcce0e6053900d668396ca3dc17c5cf3fd4aa14183c24" + }, + { + "path": "artifacts/results/attempts/hands/allegro_XS_write/fable-5-1/A_T2_K5/a0/submitted.npz", + "destination": "results/attempts/hands/allegro_XS_write/fable-5-1/A_T2_K5/a0/submitted.npz", + "kind": "submitted_policy", + "bytes": 56000, + "sha256": "5114a7071069acd95858fe577f02818079266049f7c2e75454ef34c8ea6f589d" + }, + { + "path": "artifacts/results/attempts/hands/allegro_XS_write/fable-5-1/B_T2_K5/a0/init_reference.npz", + "destination": "results/attempts/hands/allegro_XS_write/fable-5-1/B_T2_K5/a0/init_reference.npz", + "kind": "initial_policy", + "bytes": 55752, + "sha256": "734a84976e9e4596104bcce0e6053900d668396ca3dc17c5cf3fd4aa14183c24" + }, + { + "path": "artifacts/results/attempts/hands/allegro_XS_write/fable-5-1/B_T2_K5/a0/submitted.npz", + "destination": "results/attempts/hands/allegro_XS_write/fable-5-1/B_T2_K5/a0/submitted.npz", + "kind": "submitted_policy", + "bytes": 56000, + "sha256": "f3dd6e43164726fb1470c00460dcf541f273ec204e37439e54e5692e35cb9e19" + }, + { + "path": "artifacts/results/attempts/hands/allegro_XS_write/opus-5-5/A_T2_K5/a0/init_reference.npz", + "destination": "results/attempts/hands/allegro_XS_write/opus-5-5/A_T2_K5/a0/init_reference.npz", + "kind": "initial_policy", + "bytes": 55752, + "sha256": "734a84976e9e4596104bcce0e6053900d668396ca3dc17c5cf3fd4aa14183c24" + }, + { + "path": "artifacts/results/attempts/hands/allegro_XS_write/opus-5-5/A_T2_K5/a0/submitted.npz", + "destination": "results/attempts/hands/allegro_XS_write/opus-5-5/A_T2_K5/a0/submitted.npz", + "kind": "submitted_policy", + "bytes": 56000, + "sha256": "1ca1ce3f969ce8bf81109757b8b4549cb900de4ccfc40d3863cf94e7a24682d9" + }, + { + "path": "artifacts/results/attempts/hands/allegro_XS_write/opus-5-5/B_T2_K5/a0/init_reference.npz", + "destination": "results/attempts/hands/allegro_XS_write/opus-5-5/B_T2_K5/a0/init_reference.npz", + "kind": "initial_policy", + "bytes": 55752, + "sha256": "734a84976e9e4596104bcce0e6053900d668396ca3dc17c5cf3fd4aa14183c24" + }, + { + "path": "artifacts/results/attempts/hands/allegro_XS_write/opus-5-5/B_T2_K5/a0/submitted.npz", + "destination": "results/attempts/hands/allegro_XS_write/opus-5-5/B_T2_K5/a0/submitted.npz", + "kind": "submitted_policy", + "bytes": 57076, + "sha256": "3e087c510c31e8b9105f78e7749c99d667dbaaada2444f6d0756f02e557fc1d6" + }, + { + "path": "artifacts/results/attempts/hands/shadow_L_repair_early_ep100/fable-5-1/A_T2_K5_cap10/a0/init_reference.npz", + "destination": "results/attempts/hands/shadow_L_repair_early_ep100/fable-5-1/A_T2_K5_cap10/a0/init_reference.npz", + "kind": "initial_policy", + "bytes": 2045600, + "sha256": "9e4d9685886178e4258327a6b8f5a64ca9cf320be26454e1ba9dab29d345e13f" + }, + { + "path": "artifacts/results/attempts/hands/shadow_L_repair_early_ep100/fable-5-1/A_T2_K5_cap10/a0/submitted.npz", + "destination": "results/attempts/hands/shadow_L_repair_early_ep100/fable-5-1/A_T2_K5_cap10/a0/submitted.npz", + "kind": "submitted_policy", + "bytes": 2045848, + "sha256": "2363589ada764b30c20f1cf5647b5001a93e1e350f69b34d941fbb93af3d351d" + }, + { + "path": "artifacts/results/attempts/hands/shadow_L_repair_early_ep100/opus-5-5/A_T2_K5_cap10/a0/init_reference.npz", + "destination": "results/attempts/hands/shadow_L_repair_early_ep100/opus-5-5/A_T2_K5_cap10/a0/init_reference.npz", + "kind": "initial_policy", + "bytes": 2045600, + "sha256": "9e4d9685886178e4258327a6b8f5a64ca9cf320be26454e1ba9dab29d345e13f" + }, + { + "path": "artifacts/results/attempts/hands/shadow_L_repair_early_ep100/opus-5-5/A_T2_K5_cap10/a0/submitted.npz", + "destination": "results/attempts/hands/shadow_L_repair_early_ep100/opus-5-5/A_T2_K5_cap10/a0/submitted.npz", + "kind": "submitted_policy", + "bytes": 2046100, + "sha256": "ec3f825b99cde6db9850c8b0976e335fbfecf3c9994112dec621e473d90b7059" + }, + { + "path": "artifacts/results/attempts/hands/shadow_L_repair_early_ep200/fable-5-1/A_T2_K5_cap10/a0/init_reference.npz", + "destination": "results/attempts/hands/shadow_L_repair_early_ep200/fable-5-1/A_T2_K5_cap10/a0/init_reference.npz", + "kind": "initial_policy", + "bytes": 2045600, + "sha256": "553a9eb82d9680976f727df64ece1d43dae1a2f7cca25194b06fcb590f98b661" + }, + { + "path": "artifacts/results/attempts/hands/shadow_L_repair_early_ep200/fable-5-1/A_T2_K5_cap10/a0/submitted.npz", + "destination": "results/attempts/hands/shadow_L_repair_early_ep200/fable-5-1/A_T2_K5_cap10/a0/submitted.npz", + "kind": "submitted_policy", + "bytes": 2045848, + "sha256": "8677efe5de181dbc3742e9d713e532d60ff881d26939da8fbe7ef906f533bf44" + }, + { + "path": "artifacts/results/attempts/hands/shadow_L_repair_early_ep200/opus-5-5/A_T2_K5_cap10/a0/init_reference.npz", + "destination": "results/attempts/hands/shadow_L_repair_early_ep200/opus-5-5/A_T2_K5_cap10/a0/init_reference.npz", + "kind": "initial_policy", + "bytes": 2045600, + "sha256": "553a9eb82d9680976f727df64ece1d43dae1a2f7cca25194b06fcb590f98b661" + }, + { + "path": "artifacts/results/attempts/hands/shadow_L_repair_early_ep200/opus-5-5/A_T2_K5_cap10/a0/submitted.npz", + "destination": "results/attempts/hands/shadow_L_repair_early_ep200/opus-5-5/A_T2_K5_cap10/a0/submitted.npz", + "kind": "submitted_policy", + "bytes": 2045848, + "sha256": "48da7a80a7c2bdeaecab9961e4a09109b29ab2dca79f48db596539cd73c0c1a2" + }, + { + "path": "artifacts/results/attempts/hands/shadow_L_repair_norm_shift_69/fable-5-1/A_T2_K5_cap10/a0/init_reference.npz", + "destination": "results/attempts/hands/shadow_L_repair_norm_shift_69/fable-5-1/A_T2_K5_cap10/a0/init_reference.npz", + "kind": "initial_policy", + "bytes": 2045600, + "sha256": "90e53d7b453d254827a81c9bcbf2605ca2e2791c2c65aa3661b5eff7d91565a4" + }, + { + "path": "artifacts/results/attempts/hands/shadow_L_repair_norm_shift_69/fable-5-1/A_T2_K5_cap10/a0/submitted.npz", + "destination": "results/attempts/hands/shadow_L_repair_norm_shift_69/fable-5-1/A_T2_K5_cap10/a0/submitted.npz", + "kind": "submitted_policy", + "bytes": 2045848, + "sha256": "3cf2562eef969b3c5b3d3f4fc05163c3bc3c121a33820d95cdeb60dd41de41d7" + }, + { + "path": "artifacts/results/attempts/hands/shadow_L_repair_norm_shift_69/fable-5-1/A_T2_K5_cap10/blind0/init_reference.npz", + "destination": "results/attempts/hands/shadow_L_repair_norm_shift_69/fable-5-1/A_T2_K5_cap10/blind0/init_reference.npz", + "kind": "initial_policy", + "bytes": 2045600, + "sha256": "90e53d7b453d254827a81c9bcbf2605ca2e2791c2c65aa3661b5eff7d91565a4" + }, + { + "path": "artifacts/results/attempts/hands/shadow_L_repair_norm_shift_69/fable-5-1/A_T2_K5_cap10/blind0/submitted.npz", + "destination": "results/attempts/hands/shadow_L_repair_norm_shift_69/fable-5-1/A_T2_K5_cap10/blind0/submitted.npz", + "kind": "submitted_policy", + "bytes": 2045848, + "sha256": "3cf2562eef969b3c5b3d3f4fc05163c3bc3c121a33820d95cdeb60dd41de41d7" + }, + { + "path": "artifacts/results/attempts/hands/shadow_L_repair_swap_act_7_11/opus-5-5/A_T2_K5_cap10/a0/init_reference.npz", + "destination": "results/attempts/hands/shadow_L_repair_swap_act_7_11/opus-5-5/A_T2_K5_cap10/a0/init_reference.npz", + "kind": "initial_policy", + "bytes": 2045600, + "sha256": "469ff9e9a23f5b7b91fd61ee72466d1c070b906b7bf917f82b6a23dd4832579c" + }, + { + "path": "artifacts/results/attempts/hands/shadow_L_repair_swap_act_7_11/opus-5-5/A_T2_K5_cap10/a0/submitted.npz", + "destination": "results/attempts/hands/shadow_L_repair_swap_act_7_11/opus-5-5/A_T2_K5_cap10/a0/submitted.npz", + "kind": "submitted_policy", + "bytes": 2045600, + "sha256": "4fa39ebb60303fe1563aab7eb3d96bea2d7f6ac1f1ff11e368089aec981659a1" + }, + { + "path": "artifacts/results/attempts/hands/shadow_XS_write/fable-5-1/A_T2_K5/a0/init_reference.npz", + "destination": "results/attempts/hands/shadow_XS_write/fable-5-1/A_T2_K5/a0/init_reference.npz", + "kind": "initial_policy", + "bytes": 65504, + "sha256": "a2c82623b46c5121650a575c5a0b6525fd86b2a6ed1ace507d7829f902379ee0" + }, + { + "path": "artifacts/results/attempts/hands/shadow_XS_write/fable-5-1/A_T2_K5/a0/submitted.npz", + "destination": "results/attempts/hands/shadow_XS_write/fable-5-1/A_T2_K5/a0/submitted.npz", + "kind": "submitted_policy", + "bytes": 65752, + "sha256": "93bf86d50b379b33848ab89c99e25a74a31490669ad9745c78dad71c757be7dd" + }, + { + "path": "artifacts/results/attempts/hands/shadow_XS_write/fable-5-1/B_T2_K5/a0/init_reference.npz", + "destination": "results/attempts/hands/shadow_XS_write/fable-5-1/B_T2_K5/a0/init_reference.npz", + "kind": "initial_policy", + "bytes": 65504, + "sha256": "a2c82623b46c5121650a575c5a0b6525fd86b2a6ed1ace507d7829f902379ee0" + }, + { + "path": "artifacts/results/attempts/hands/shadow_XS_write/fable-5-1/B_T2_K5/a0/submitted.npz", + "destination": "results/attempts/hands/shadow_XS_write/fable-5-1/B_T2_K5/a0/submitted.npz", + "kind": "submitted_policy", + "bytes": 65752, + "sha256": "ac4940efa41f77ca964d64e10d8471fa6ccb74064f7a025ad203fd8cd37b675e" + }, + { + "path": "artifacts/results/attempts/hands/shadow_XS_write/opus-5-5/A_T2_K5/a0/init_reference.npz", + "destination": "results/attempts/hands/shadow_XS_write/opus-5-5/A_T2_K5/a0/init_reference.npz", + "kind": "initial_policy", + "bytes": 65504, + "sha256": "a2c82623b46c5121650a575c5a0b6525fd86b2a6ed1ace507d7829f902379ee0" + }, + { + "path": "artifacts/results/attempts/hands/shadow_XS_write/opus-5-5/A_T2_K5/a0/submitted.npz", + "destination": "results/attempts/hands/shadow_XS_write/opus-5-5/A_T2_K5/a0/submitted.npz", + "kind": "submitted_policy", + "bytes": 66616, + "sha256": "e705e7871dc90da44eb7a5fc965017aeca214349233afda14a79e032d6213ba3" + }, + { + "path": "artifacts/results/attempts/hands/shadow_XS_write/opus-5-5/B_T2_K5/a0/init_reference.npz", + "destination": "results/attempts/hands/shadow_XS_write/opus-5-5/B_T2_K5/a0/init_reference.npz", + "kind": "initial_policy", + "bytes": 65504, + "sha256": "a2c82623b46c5121650a575c5a0b6525fd86b2a6ed1ace507d7829f902379ee0" + }, + { + "path": "artifacts/results/attempts/hands/shadow_XS_write/opus-5-5/B_T2_K5/a0/submitted.npz", + "destination": "results/attempts/hands/shadow_XS_write/opus-5-5/B_T2_K5/a0/submitted.npz", + "kind": "submitted_policy", + "bytes": 65752, + "sha256": "358a05853e5482d0b059b5725d5a80a8fb538aa6c479aba1a04d230af2eff77c" + }, + { + "path": "artifacts/results/attempts/main/franka_M_repair_collapse_ep500/opus-5-5/A_T0_K0_cap10/a0/init_reference.npz", + "destination": "results/attempts/main/franka_M_repair_collapse_ep500/opus-5-5/A_T0_K0_cap10/a0/init_reference.npz", + "kind": "initial_policy", + "bytes": 194148, + "sha256": "1e7762a844130a1f1ea096a79797557475aee2265f37ef6615a03ee90ee69ed8" + }, + { + "path": "artifacts/results/attempts/main/franka_M_repair_collapse_ep500/opus-5-5/A_T0_K0_cap10/a0/submitted.npz", + "destination": "results/attempts/main/franka_M_repair_collapse_ep500/opus-5-5/A_T0_K0_cap10/a0/submitted.npz", + "kind": "submitted_policy", + "bytes": 194148, + "sha256": "6d6c6bd70b5d9bd3d9b5be621bf2c6f0c016b17ca3dc94aa2b14db793bb06f2b" + }, + { + "path": "artifacts/results/attempts/main/franka_M_repair_collapse_ep500/opus-5-5/A_T2_K5_cap10/a0/init_reference.npz", + "destination": "results/attempts/main/franka_M_repair_collapse_ep500/opus-5-5/A_T2_K5_cap10/a0/init_reference.npz", + "kind": "initial_policy", + "bytes": 194148, + "sha256": "1e7762a844130a1f1ea096a79797557475aee2265f37ef6615a03ee90ee69ed8" + }, + { + "path": "artifacts/results/attempts/main/franka_M_repair_collapse_ep500/opus-5-5/A_T2_K5_cap10/a0/submitted.npz", + "destination": "results/attempts/main/franka_M_repair_collapse_ep500/opus-5-5/A_T2_K5_cap10/a0/submitted.npz", + "kind": "submitted_policy", + "bytes": 194396, + "sha256": "68c10234c9d96fbde4f63ff8045219544bd963697a8f83ffc6e850c80868a36d" + }, + { + "path": "artifacts/results/attempts/main/franka_M_repair_early_ep15/opus-5-5/A_T0_K0_cap10/a0/init_reference.npz", + "destination": "results/attempts/main/franka_M_repair_early_ep15/opus-5-5/A_T0_K0_cap10/a0/init_reference.npz", + "kind": "initial_policy", + "bytes": 194148, + "sha256": "feeb4d3f2ddfa9d0b6b585c7205c1d110a3db236213206d1f5558e6bc8507047" + }, + { + "path": "artifacts/results/attempts/main/franka_M_repair_early_ep15/opus-5-5/A_T0_K0_cap10/a0/submitted.npz", + "destination": "results/attempts/main/franka_M_repair_early_ep15/opus-5-5/A_T0_K0_cap10/a0/submitted.npz", + "kind": "submitted_policy", + "bytes": 194396, + "sha256": "9ad5cdd13c12b9af908d396c927e1318a4879bd2dfdc40b9dccd4fecda630fc0" + }, + { + "path": "artifacts/results/attempts/main/franka_M_repair_early_ep15/opus-5-5/A_T2_K5_cap10/a0/init_reference.npz", + "destination": "results/attempts/main/franka_M_repair_early_ep15/opus-5-5/A_T2_K5_cap10/a0/init_reference.npz", + "kind": "initial_policy", + "bytes": 194148, + "sha256": "feeb4d3f2ddfa9d0b6b585c7205c1d110a3db236213206d1f5558e6bc8507047" + }, + { + "path": "artifacts/results/attempts/main/franka_M_repair_early_ep15/opus-5-5/A_T2_K5_cap10/a0/submitted.npz", + "destination": "results/attempts/main/franka_M_repair_early_ep15/opus-5-5/A_T2_K5_cap10/a0/submitted.npz", + "kind": "submitted_policy", + "bytes": 194396, + "sha256": "2aa13d70587bcacd3fd270156047e474b3aadfb41f2c1c4b1532cf521bae04a1" + }, + { + "path": "artifacts/results/attempts/main/franka_M_repair_early_ep20/opus-5-5/A_T0_K0_cap10/a0/init_reference.npz", + "destination": "results/attempts/main/franka_M_repair_early_ep20/opus-5-5/A_T0_K0_cap10/a0/init_reference.npz", + "kind": "initial_policy", + "bytes": 194148, + "sha256": "fb12def9ecfbd7531312f7d0690d326bc550ac94d6ae3197c0c770fb8c1af0f8" + }, + { + "path": "artifacts/results/attempts/main/franka_M_repair_early_ep20/opus-5-5/A_T0_K0_cap10/a0/submitted.npz", + "destination": "results/attempts/main/franka_M_repair_early_ep20/opus-5-5/A_T0_K0_cap10/a0/submitted.npz", + "kind": "submitted_policy", + "bytes": 194148, + "sha256": "4a6ffa580c733429a97fbb429bdb5cb70a7712dbbfa3b211ca85468291b94387" + }, + { + "path": "artifacts/results/attempts/main/franka_M_repair_early_ep20/opus-5-5/A_T2_K5_cap10/a0/init_reference.npz", + "destination": "results/attempts/main/franka_M_repair_early_ep20/opus-5-5/A_T2_K5_cap10/a0/init_reference.npz", + "kind": "initial_policy", + "bytes": 194148, + "sha256": "fb12def9ecfbd7531312f7d0690d326bc550ac94d6ae3197c0c770fb8c1af0f8" + }, + { + "path": "artifacts/results/attempts/main/franka_M_repair_early_ep20/opus-5-5/A_T2_K5_cap10/a0/submitted.npz", + "destination": "results/attempts/main/franka_M_repair_early_ep20/opus-5-5/A_T2_K5_cap10/a0/submitted.npz", + "kind": "submitted_policy", + "bytes": 194396, + "sha256": "80882383fd905abcc536a4d5e47f4066c116d943a2149a88fdaf54e42005a12a" + }, + { + "path": "artifacts/results/attempts/main/franka_M_repair_early_ep40/opus-5-5/A_T0_K0_cap10/a0/init_reference.npz", + "destination": "results/attempts/main/franka_M_repair_early_ep40/opus-5-5/A_T0_K0_cap10/a0/init_reference.npz", + "kind": "initial_policy", + "bytes": 194148, + "sha256": "29b96da17627852282b904f6b98e0d7fa16720fcb18be8a1005b97c6ebe1bf50" + }, + { + "path": "artifacts/results/attempts/main/franka_M_repair_early_ep40/opus-5-5/A_T0_K0_cap10/a0/submitted.npz", + "destination": "results/attempts/main/franka_M_repair_early_ep40/opus-5-5/A_T0_K0_cap10/a0/submitted.npz", + "kind": "submitted_policy", + "bytes": 194148, + "sha256": "6e856294d2d666197fd7245f766487f89642dfe6826956c821fe7dc4b0d74ad8" + }, + { + "path": "artifacts/results/attempts/main/franka_M_repair_norm_shift_21/opus-5-5/A_T0_K0_cap10/a0/init_reference.npz", + "destination": "results/attempts/main/franka_M_repair_norm_shift_21/opus-5-5/A_T0_K0_cap10/a0/init_reference.npz", + "kind": "initial_policy", + "bytes": 194148, + "sha256": "5f165e2433e09f4894d3cbde8b4ff23f8447cd2d92ebf954d538538930ff1ea3" + }, + { + "path": "artifacts/results/attempts/main/franka_M_repair_norm_shift_21/opus-5-5/A_T0_K0_cap10/a0/submitted.npz", + "destination": "results/attempts/main/franka_M_repair_norm_shift_21/opus-5-5/A_T0_K0_cap10/a0/submitted.npz", + "kind": "submitted_policy", + "bytes": 194148, + "sha256": "88de103d6c963a22904a46793f85ee7811665f6e955583eb4c355164a8bd474a" + }, + { + "path": "artifacts/results/attempts/main/franka_M_repair_norm_shift_21/opus-5-5/A_T2_K5_cap10/a0/init_reference.npz", + "destination": "results/attempts/main/franka_M_repair_norm_shift_21/opus-5-5/A_T2_K5_cap10/a0/init_reference.npz", + "kind": "initial_policy", + "bytes": 194148, + "sha256": "5f165e2433e09f4894d3cbde8b4ff23f8447cd2d92ebf954d538538930ff1ea3" + }, + { + "path": "artifacts/results/attempts/main/franka_M_repair_norm_shift_21/opus-5-5/A_T2_K5_cap10/a0/submitted.npz", + "destination": "results/attempts/main/franka_M_repair_norm_shift_21/opus-5-5/A_T2_K5_cap10/a0/submitted.npz", + "kind": "submitted_policy", + "bytes": 194396, + "sha256": "ebfda4cf4cff5daf24025797841fa8b80f0707396824889789cf1971048a1c4b" + }, + { + "path": "artifacts/results/attempts/main/franka_M_repair_norm_shift_21/opus-5-5/B_T0_K0_cap10/a0/init_reference.npz", + "destination": "results/attempts/main/franka_M_repair_norm_shift_21/opus-5-5/B_T0_K0_cap10/a0/init_reference.npz", + "kind": "initial_policy", + "bytes": 194148, + "sha256": "5f165e2433e09f4894d3cbde8b4ff23f8447cd2d92ebf954d538538930ff1ea3" + }, + { + "path": "artifacts/results/attempts/main/franka_M_repair_norm_shift_21/opus-5-5/B_T0_K0_cap10/a0/submitted.npz", + "destination": "results/attempts/main/franka_M_repair_norm_shift_21/opus-5-5/B_T0_K0_cap10/a0/submitted.npz", + "kind": "submitted_policy", + "bytes": 194396, + "sha256": "63b7062e0f0a2487244e5c5a0b3b12b47c7e60ef97e1a01e020bd2372eb3f557" + }, + { + "path": "artifacts/results/attempts/main/franka_M_repair_norm_shift_21/opus-5-5/B_T0_K0_cap10/blind0/init_reference.npz", + "destination": "results/attempts/main/franka_M_repair_norm_shift_21/opus-5-5/B_T0_K0_cap10/blind0/init_reference.npz", + "kind": "initial_policy", + "bytes": 194148, + "sha256": "5f165e2433e09f4894d3cbde8b4ff23f8447cd2d92ebf954d538538930ff1ea3" + }, + { + "path": "artifacts/results/attempts/main/franka_M_repair_norm_shift_21/opus-5-5/B_T0_K0_cap10/blind0/submitted.npz", + "destination": "results/attempts/main/franka_M_repair_norm_shift_21/opus-5-5/B_T0_K0_cap10/blind0/submitted.npz", + "kind": "submitted_policy", + "bytes": 194148, + "sha256": "ebcca5e3dba94685ad72fb457db86672e533b3d92a39223a31fbd206d3137a20" + }, + { + "path": "artifacts/results/attempts/main/franka_M_repair_swap_act_1_3/opus-5-5/A_T0_K0_cap10/a0/init_reference.npz", + "destination": "results/attempts/main/franka_M_repair_swap_act_1_3/opus-5-5/A_T0_K0_cap10/a0/init_reference.npz", + "kind": "initial_policy", + "bytes": 194148, + "sha256": "f0fb11211277f7b4974756cd24dbec18e9dbd7f3025e9de93a823e0c7b05202a" + }, + { + "path": "artifacts/results/attempts/main/franka_M_repair_swap_act_1_3/opus-5-5/A_T0_K0_cap10/a0/submitted.npz", + "destination": "results/attempts/main/franka_M_repair_swap_act_1_3/opus-5-5/A_T0_K0_cap10/a0/submitted.npz", + "kind": "submitted_policy", + "bytes": 194396, + "sha256": "5522200d12881600bc2aa93dc9a6b8a761030d95198140190204ca83f6cd1efa" + }, + { + "path": "artifacts/results/attempts/main/franka_M_repair_swap_act_1_3/opus-5-5/A_T0_K0_cap10/blind0/init_reference.npz", + "destination": "results/attempts/main/franka_M_repair_swap_act_1_3/opus-5-5/A_T0_K0_cap10/blind0/init_reference.npz", + "kind": "initial_policy", + "bytes": 194148, + "sha256": "f0fb11211277f7b4974756cd24dbec18e9dbd7f3025e9de93a823e0c7b05202a" + }, + { + "path": "artifacts/results/attempts/main/franka_M_repair_swap_act_1_3/opus-5-5/A_T0_K0_cap10/blind0/submitted.npz", + "destination": "results/attempts/main/franka_M_repair_swap_act_1_3/opus-5-5/A_T0_K0_cap10/blind0/submitted.npz", + "kind": "submitted_policy", + "bytes": 194396, + "sha256": "e84847ee59e93cfeec4f68273f72ab677bdec4db76f1a3ff06117c6db5bdec4e" + }, + { + "path": "artifacts/results/attempts/main/franka_M_repair_swap_act_1_3/opus-5-5/A_T2_K5_cap10/a0/init_reference.npz", + "destination": "results/attempts/main/franka_M_repair_swap_act_1_3/opus-5-5/A_T2_K5_cap10/a0/init_reference.npz", + "kind": "initial_policy", + "bytes": 194148, + "sha256": "f0fb11211277f7b4974756cd24dbec18e9dbd7f3025e9de93a823e0c7b05202a" + }, + { + "path": "artifacts/results/attempts/main/franka_M_repair_swap_act_1_3/opus-5-5/A_T2_K5_cap10/a0/submitted.npz", + "destination": "results/attempts/main/franka_M_repair_swap_act_1_3/opus-5-5/A_T2_K5_cap10/a0/submitted.npz", + "kind": "submitted_policy", + "bytes": 194148, + "sha256": "353f84b62e0dbd1bc10d6fa97d39db20f195fe968f30aa30d1e13e8f32b445e8" + }, + { + "path": "artifacts/results/attempts/main/franka_M_repair_swap_act_1_3/opus-5-5/A_T2_K5_cap10/blind0/init_reference.npz", + "destination": "results/attempts/main/franka_M_repair_swap_act_1_3/opus-5-5/A_T2_K5_cap10/blind0/init_reference.npz", + "kind": "initial_policy", + "bytes": 194148, + "sha256": "f0fb11211277f7b4974756cd24dbec18e9dbd7f3025e9de93a823e0c7b05202a" + }, + { + "path": "artifacts/results/attempts/main/franka_M_repair_swap_act_1_3/opus-5-5/A_T2_K5_cap10/blind0/submitted.npz", + "destination": "results/attempts/main/franka_M_repair_swap_act_1_3/opus-5-5/A_T2_K5_cap10/blind0/submitted.npz", + "kind": "submitted_policy", + "bytes": 194148, + "sha256": "353f84b62e0dbd1bc10d6fa97d39db20f195fe968f30aa30d1e13e8f32b445e8" + }, + { + "path": "artifacts/results/attempts/main/franka_M_repair_swap_act_1_3/opus-5-5/B_T0_K0_cap10/a0/init_reference.npz", + "destination": "results/attempts/main/franka_M_repair_swap_act_1_3/opus-5-5/B_T0_K0_cap10/a0/init_reference.npz", + "kind": "initial_policy", + "bytes": 194148, + "sha256": "f0fb11211277f7b4974756cd24dbec18e9dbd7f3025e9de93a823e0c7b05202a" + }, + { + "path": "artifacts/results/attempts/main/franka_M_repair_swap_act_1_3/opus-5-5/B_T0_K0_cap10/a0/submitted.npz", + "destination": "results/attempts/main/franka_M_repair_swap_act_1_3/opus-5-5/B_T0_K0_cap10/a0/submitted.npz", + "kind": "submitted_policy", + "bytes": 194396, + "sha256": "5522200d12881600bc2aa93dc9a6b8a761030d95198140190204ca83f6cd1efa" + }, + { + "path": "artifacts/results/attempts/main/franka_M_repair_swap_act_1_3/opus-5-5/B_T0_K0_cap10/blind0/init_reference.npz", + "destination": "results/attempts/main/franka_M_repair_swap_act_1_3/opus-5-5/B_T0_K0_cap10/blind0/init_reference.npz", + "kind": "initial_policy", + "bytes": 194148, + "sha256": "f0fb11211277f7b4974756cd24dbec18e9dbd7f3025e9de93a823e0c7b05202a" + }, + { + "path": "artifacts/results/attempts/main/franka_M_repair_swap_act_1_3/opus-5-5/B_T0_K0_cap10/blind0/submitted.npz", + "destination": "results/attempts/main/franka_M_repair_swap_act_1_3/opus-5-5/B_T0_K0_cap10/blind0/submitted.npz", + "kind": "submitted_policy", + "bytes": 194396, + "sha256": "3775df16902276e289c071f1dbbfe1f04c57c978759ff67a273d9023a9e25734" + }, + { + "path": "artifacts/results/attempts/main/franka_XS_write/opus-5-5/A_T0_K0/a0/submitted.npz", + "destination": "results/attempts/main/franka_XS_write/opus-5-5/A_T0_K0/a0/submitted.npz", + "kind": "submitted_policy", + "bytes": 27268, + "sha256": "e89c89dd8fc66e0c92ab5d7222c594e22157b3a5f02c468ecd4299ea658acb17" + }, + { + "path": "artifacts/results/attempts/main/franka_XS_write/opus-5-5/A_T2_K5/a0/init_reference.npz", + "destination": "results/attempts/main/franka_XS_write/opus-5-5/A_T2_K5/a0/init_reference.npz", + "kind": "initial_policy", + "bytes": 27268, + "sha256": "4fe489675c7f488ab1abae8d14c4ddfd6933ea28ab66f9e79b3540f6df86c6eb" + }, + { + "path": "artifacts/results/attempts/main/franka_XS_write/opus-5-5/A_T2_K5/a0/submitted.npz", + "destination": "results/attempts/main/franka_XS_write/opus-5-5/A_T2_K5/a0/submitted.npz", + "kind": "submitted_policy", + "bytes": 27620, + "sha256": "11cbf2c271810b25397769d8d1e19455ef7e62b789a90234a15803ae51a83321" + } + ], + "scope": "checkpoints_and_figures_only" +} diff --git a/artifacts/openvla_study/results/attempts/fable-5-1_single/submitted.npz b/artifacts/openvla_study/results/attempts/fable-5-1_single/submitted.npz new file mode 100644 index 0000000000000000000000000000000000000000..97fb17f9dc66d784e4864c2d8d2602564e5337ea --- /dev/null +++ b/artifacts/openvla_study/results/attempts/fable-5-1_single/submitted.npz @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:928557ceaf67ab2635f230ee52b77e0372bd2abe2658a25fa52486880ee63ca5 +size 50206 diff --git a/artifacts/openvla_study/results/attempts/fable-5-1_three/submitted.npz b/artifacts/openvla_study/results/attempts/fable-5-1_three/submitted.npz new file mode 100644 index 0000000000000000000000000000000000000000..92b9a149303e27e573b9b63e8f5b7e2fb7dbdda7 --- /dev/null +++ b/artifacts/openvla_study/results/attempts/fable-5-1_three/submitted.npz @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:886cbc79f4cfa23b6b9bfceafaaf9ceef0fa351f3ac88b5e17de1bf13681c046 +size 50206 diff --git a/artifacts/openvla_study/results/attempts/opus-5-5_single/submitted.npz b/artifacts/openvla_study/results/attempts/opus-5-5_single/submitted.npz new file mode 100644 index 0000000000000000000000000000000000000000..16d9e92d0bdb573d29ec4bf23f7be7ee2815e46f --- /dev/null +++ b/artifacts/openvla_study/results/attempts/opus-5-5_single/submitted.npz @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:bf5bd8eee9e056055bf920c2391541bfd57c402631da9fc966ee81a4235889ad +size 50206 diff --git a/artifacts/openvla_study/results/attempts/opus-5-5_three/submitted.npz b/artifacts/openvla_study/results/attempts/opus-5-5_three/submitted.npz new file mode 100644 index 0000000000000000000000000000000000000000..09d5beeadc2c357d978a09e409690d8b707bbb09 --- /dev/null +++ b/artifacts/openvla_study/results/attempts/opus-5-5_three/submitted.npz @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4309dc07ccd8e2195f7bb2f32a1e572cf22886036fa8ee6c52385fc0fb5b9eed +size 278 diff --git a/artifacts/policies/allegro/nvidia/pretrained.npz b/artifacts/policies/allegro/nvidia/pretrained.npz new file mode 100644 index 0000000000000000000000000000000000000000..e90bb4a149ca37a2e6e48a9a90993fe790493930 --- /dev/null +++ b/artifacts/policies/allegro/nvidia/pretrained.npz @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:64b82670444794c53bb02867c8e483b0f70e98a5dbaaaff46cdcb5a7aa022964 +size 3281400 diff --git a/artifacts/policies/allegro/ppo/allegro_L_s0_early/best.npz b/artifacts/policies/allegro/ppo/allegro_L_s0_early/best.npz new file mode 100644 index 0000000000000000000000000000000000000000..30591af6ae0a24c2d0aeb572744f03a7689cf591 --- /dev/null +++ b/artifacts/policies/allegro/ppo/allegro_L_s0_early/best.npz @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:9ad19006d9d16e17b4d5b38f1a0fd3adc2e0af25610e73c127b763133b12dc60 +size 3281256 diff --git a/artifacts/policies/allegro/ppo/allegro_L_s0_early/ep00005.npz b/artifacts/policies/allegro/ppo/allegro_L_s0_early/ep00005.npz new file mode 100644 index 0000000000000000000000000000000000000000..78d9ce4890f45f4362ea409a860a40a5e4cb5e0b --- /dev/null +++ b/artifacts/policies/allegro/ppo/allegro_L_s0_early/ep00005.npz @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f8ad8ce9f689c5f6f8cfb31cfb850842306d4c8419e0f8338c10ea7cb562ef62 +size 3281340 diff --git a/artifacts/policies/allegro/ppo/allegro_L_s0_early/ep00010.npz b/artifacts/policies/allegro/ppo/allegro_L_s0_early/ep00010.npz new file mode 100644 index 0000000000000000000000000000000000000000..1a17c63f7f05a1f9cd684275dbf75d62f62b5ec1 --- /dev/null +++ b/artifacts/policies/allegro/ppo/allegro_L_s0_early/ep00010.npz @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a149f04cacf3f61d78dc3e8b9e75afc5b7c10537eb7027ef5697649397b3d4b1 +size 3281352 diff --git a/artifacts/policies/allegro/ppo/allegro_L_s0_early/ep00015.npz b/artifacts/policies/allegro/ppo/allegro_L_s0_early/ep00015.npz new file mode 100644 index 0000000000000000000000000000000000000000..ce27153c205cc979d03909fdf7873852d72f3843 --- /dev/null +++ b/artifacts/policies/allegro/ppo/allegro_L_s0_early/ep00015.npz @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4d4d08f7e6b80ee2e42b5d30ae5b000c351c1ea5290e519960d8da95ac99143a +size 3281348 diff --git a/artifacts/policies/allegro/ppo/allegro_L_s0_early/ep00020.npz b/artifacts/policies/allegro/ppo/allegro_L_s0_early/ep00020.npz new file mode 100644 index 0000000000000000000000000000000000000000..f9c0028088dae8fe57dc32818eccf738f1bb3379 --- /dev/null +++ b/artifacts/policies/allegro/ppo/allegro_L_s0_early/ep00020.npz @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c5cd88279145a82c0fc8337a2071894ff68ba38386fa339f4ce59cb17e9fdb6e +size 3281348 diff --git a/artifacts/policies/allegro/ppo/allegro_L_s0_early/ep00025.npz b/artifacts/policies/allegro/ppo/allegro_L_s0_early/ep00025.npz new file mode 100644 index 0000000000000000000000000000000000000000..cdaeb95bf53a588d09b579cece7a1a9e719e28f3 --- /dev/null +++ b/artifacts/policies/allegro/ppo/allegro_L_s0_early/ep00025.npz @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ded7b54c3bf7762b3527b80cce2a42e5858ad7e7082496e319e56e31c9e2f968 +size 3281348 diff --git a/artifacts/policies/allegro/ppo/allegro_L_s0_early/ep00030.npz b/artifacts/policies/allegro/ppo/allegro_L_s0_early/ep00030.npz new file mode 100644 index 0000000000000000000000000000000000000000..36bbc8ec5963459c6ca1440693971b3e50895671 --- /dev/null +++ b/artifacts/policies/allegro/ppo/allegro_L_s0_early/ep00030.npz @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f187ae23a3c4c0c8ed128b90dd9bc4820bbed8452953f33118fbd614f7f27276 +size 3281348 diff --git a/artifacts/policies/allegro/ppo/allegro_L_s0_early/ep00035.npz b/artifacts/policies/allegro/ppo/allegro_L_s0_early/ep00035.npz new file mode 100644 index 0000000000000000000000000000000000000000..5fd0fd5b49d2f21471044b887a5b17981d165179 --- /dev/null +++ b/artifacts/policies/allegro/ppo/allegro_L_s0_early/ep00035.npz @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:580d2214628a6fd306e207fb70dfff1b33f91754d8f1ca4d5b11219aae2dd522 +size 3281348 diff --git a/artifacts/policies/allegro/ppo/allegro_L_s0_early/ep00040.npz b/artifacts/policies/allegro/ppo/allegro_L_s0_early/ep00040.npz new file mode 100644 index 0000000000000000000000000000000000000000..fd61cbf5f950e50209c40f18d674a3ddfe6c0570 --- /dev/null +++ b/artifacts/policies/allegro/ppo/allegro_L_s0_early/ep00040.npz @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:41d65c4c5f7e85d4e8d22c049c434a42f21a39ed9ea8726cf83a6d13b330f6ab +size 3281344 diff --git a/artifacts/policies/allegro/ppo/allegro_L_s0_early/ep00045.npz b/artifacts/policies/allegro/ppo/allegro_L_s0_early/ep00045.npz new file mode 100644 index 0000000000000000000000000000000000000000..354b70f0dfe3d4cc7e60cc3e61c01d4b6e282615 --- /dev/null +++ b/artifacts/policies/allegro/ppo/allegro_L_s0_early/ep00045.npz @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:beed2ed568498740bb0c6aa72f5573f675c163842db041a1b1e1b9935c717cf0 +size 3281348 diff --git a/artifacts/policies/allegro/ppo/allegro_L_s0_early/ep00050.npz b/artifacts/policies/allegro/ppo/allegro_L_s0_early/ep00050.npz new file mode 100644 index 0000000000000000000000000000000000000000..14245530e23b058145352ca6004a04fda3bb5c91 --- /dev/null +++ b/artifacts/policies/allegro/ppo/allegro_L_s0_early/ep00050.npz @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:9864b0300d14c2d82918b0ee593efb4224f78c568ac7ca5da3d3b7442c17dac2 +size 3281348 diff --git a/artifacts/policies/allegro/ppo/allegro_L_s0_early/ep00055.npz b/artifacts/policies/allegro/ppo/allegro_L_s0_early/ep00055.npz new file mode 100644 index 0000000000000000000000000000000000000000..4d3e32db74c334a5b0ab8bfa4e1bc396c7bae5af --- /dev/null +++ b/artifacts/policies/allegro/ppo/allegro_L_s0_early/ep00055.npz @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d0448e98084231409f848e4836a57fd78512703f0c2528d04386ddd03d3fec8f +size 3281348 diff --git a/artifacts/policies/allegro/ppo/allegro_L_s0_early/ep00060.npz b/artifacts/policies/allegro/ppo/allegro_L_s0_early/ep00060.npz new file mode 100644 index 0000000000000000000000000000000000000000..0e8c8aea7af851a77b1921f4188844fa07aadfc9 --- /dev/null +++ b/artifacts/policies/allegro/ppo/allegro_L_s0_early/ep00060.npz @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:1111543b74bb463ceef767c2e34547dc7c8ebd9b679cfc967c5564d1e4ff443a +size 3281344 diff --git a/artifacts/policies/allegro/ppo/allegro_L_s0_early/ep00065.npz b/artifacts/policies/allegro/ppo/allegro_L_s0_early/ep00065.npz new file mode 100644 index 0000000000000000000000000000000000000000..4c9ec5f1f15b6d6a453a90c306192e74d85c70cd --- /dev/null +++ b/artifacts/policies/allegro/ppo/allegro_L_s0_early/ep00065.npz @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c1bb0a830681442da728fd0d784917472f1e08238b6d7ffca8c66bc6b316662e +size 3281344 diff --git a/artifacts/policies/allegro/ppo/allegro_L_s0_early/ep00070.npz b/artifacts/policies/allegro/ppo/allegro_L_s0_early/ep00070.npz new file mode 100644 index 0000000000000000000000000000000000000000..53638b8adedae7506c3451a25318e17ccd17704b --- /dev/null +++ b/artifacts/policies/allegro/ppo/allegro_L_s0_early/ep00070.npz @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:5a8e1ad5ef8d7a662be1c90b6b86f5bf12146ca9aadf5c0a5f39bc1ad53e1ffb +size 3281344 diff --git a/artifacts/policies/allegro/ppo/allegro_L_s0_early/ep00075.npz b/artifacts/policies/allegro/ppo/allegro_L_s0_early/ep00075.npz new file mode 100644 index 0000000000000000000000000000000000000000..0cf6612c95b32a31ad208c361c07963a3517d380 --- /dev/null +++ b/artifacts/policies/allegro/ppo/allegro_L_s0_early/ep00075.npz @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4af159a1caa95ee01567e958145e31ab9ab845f29ae33290e0ba740d3902b03f +size 3281344 diff --git a/artifacts/policies/allegro/ppo/allegro_L_s0_early/ep00080.npz b/artifacts/policies/allegro/ppo/allegro_L_s0_early/ep00080.npz new file mode 100644 index 0000000000000000000000000000000000000000..6e48f89729baf3243b3162bd966d5413709ed258 --- /dev/null +++ b/artifacts/policies/allegro/ppo/allegro_L_s0_early/ep00080.npz @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3dfae14a21d4df103dc937f06dd440fb629ba76fa09b48444599db1326badf31 +size 3281352 diff --git a/artifacts/policies/allegro/ppo/allegro_L_s0_early/ep00085.npz b/artifacts/policies/allegro/ppo/allegro_L_s0_early/ep00085.npz new file mode 100644 index 0000000000000000000000000000000000000000..d5428ca9fa5a216462aec3796f357c901082d65c --- /dev/null +++ b/artifacts/policies/allegro/ppo/allegro_L_s0_early/ep00085.npz @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b929091a8cded631a2ca69eb57049045cffdae0d70ba6698eb7eece929233a29 +size 3281348 diff --git a/artifacts/policies/allegro/ppo/allegro_L_s0_early/ep00090.npz b/artifacts/policies/allegro/ppo/allegro_L_s0_early/ep00090.npz new file mode 100644 index 0000000000000000000000000000000000000000..dd7ec0d4eb220c1daa96a6e21674f62fd26e0a31 --- /dev/null +++ b/artifacts/policies/allegro/ppo/allegro_L_s0_early/ep00090.npz @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:83cae2e05a28869157badc699838c8f18a1f0b5a2d3a36d7bdbed715e9a2b897 +size 3281348 diff --git a/artifacts/policies/allegro/ppo/allegro_L_s0_early/ep00095.npz b/artifacts/policies/allegro/ppo/allegro_L_s0_early/ep00095.npz new file mode 100644 index 0000000000000000000000000000000000000000..aa825f4ae267ba4d429cfdc6c4caf6c91919f6d7 --- /dev/null +++ b/artifacts/policies/allegro/ppo/allegro_L_s0_early/ep00095.npz @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c8525261d1144654748cb36abb9aef7ecfb13203d48cd2736fe43086da244191 +size 3281348 diff --git a/artifacts/policies/allegro/ppo/allegro_L_s0_early/ep00100.npz b/artifacts/policies/allegro/ppo/allegro_L_s0_early/ep00100.npz new file mode 100644 index 0000000000000000000000000000000000000000..935ae056b51db2eaa4a30564f322293057fe6fd8 --- /dev/null +++ b/artifacts/policies/allegro/ppo/allegro_L_s0_early/ep00100.npz @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a38877448b3f2811303b417822d0949902e8be682e4022df89c3b644f62cda08 +size 3281356 diff --git a/artifacts/policies/allegro/ppo/allegro_L_s0_early/ep00105.npz b/artifacts/policies/allegro/ppo/allegro_L_s0_early/ep00105.npz new file mode 100644 index 0000000000000000000000000000000000000000..64eefe5ec45a6630ee3eabf09eac3e1615e565ad --- /dev/null +++ b/artifacts/policies/allegro/ppo/allegro_L_s0_early/ep00105.npz @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:2c9c9b0c225c6dd76b152508b9ed1f78a62478cf7536dec003df18c51cbb8699 +size 3281356 diff --git a/artifacts/policies/allegro/ppo/allegro_L_s0_early/ep00110.npz b/artifacts/policies/allegro/ppo/allegro_L_s0_early/ep00110.npz new file mode 100644 index 0000000000000000000000000000000000000000..3fb42932b6e31936744a137ac87307877949cc2e --- /dev/null +++ b/artifacts/policies/allegro/ppo/allegro_L_s0_early/ep00110.npz @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d8f584685c5e8911c0b4c8fd6b64a1bab03344cb6f2c685a2ed8108ea90f97a0 +size 3281360 diff --git a/artifacts/policies/allegro/ppo/allegro_L_s0_early/ep00115.npz b/artifacts/policies/allegro/ppo/allegro_L_s0_early/ep00115.npz new file mode 100644 index 0000000000000000000000000000000000000000..7e2fec58cac981db552f481008d1d0f26051d5ee --- /dev/null +++ b/artifacts/policies/allegro/ppo/allegro_L_s0_early/ep00115.npz @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:70af62ad78b069d63c7600a1296e35e3e8591841fa161391c25fb0f7a3737a69 +size 3281360 diff --git a/artifacts/policies/allegro/ppo/allegro_L_s0_early/ep00120.npz b/artifacts/policies/allegro/ppo/allegro_L_s0_early/ep00120.npz new file mode 100644 index 0000000000000000000000000000000000000000..c8c2d5f4f0f0e6c68ccdb29c8ec3efe545889def --- /dev/null +++ b/artifacts/policies/allegro/ppo/allegro_L_s0_early/ep00120.npz @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e2817fa0e5f42b204134a32109d9cb163802c811a506cc4205000525605e2cac +size 3281364 diff --git a/artifacts/policies/allegro/ppo/allegro_L_s0_early/ep00125.npz b/artifacts/policies/allegro/ppo/allegro_L_s0_early/ep00125.npz new file mode 100644 index 0000000000000000000000000000000000000000..ddb5609d814364397413883f5860d7b9a74309ce --- /dev/null +++ b/artifacts/policies/allegro/ppo/allegro_L_s0_early/ep00125.npz @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c6a62bd4946f410eb269ed0376d4acd250e334d2f32e35895b740d5a8231f23d +size 3281356 diff --git a/artifacts/policies/allegro/ppo/allegro_L_s0_early/ep00130.npz b/artifacts/policies/allegro/ppo/allegro_L_s0_early/ep00130.npz new file mode 100644 index 0000000000000000000000000000000000000000..b7ef4c38f25cf68cf08f519b5d9a79804e758929 --- /dev/null +++ b/artifacts/policies/allegro/ppo/allegro_L_s0_early/ep00130.npz @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:5eaba8b893ed0602ce9ee59f44a5982804f2c32a1254d20f3f3c1db063cbe4cc +size 3281360 diff --git a/artifacts/policies/allegro/ppo/allegro_L_s0_early/ep00135.npz b/artifacts/policies/allegro/ppo/allegro_L_s0_early/ep00135.npz new file mode 100644 index 0000000000000000000000000000000000000000..201b01193886a0d634331dfceb9267db924ad991 --- /dev/null +++ b/artifacts/policies/allegro/ppo/allegro_L_s0_early/ep00135.npz @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c1bfc954ab3edb2a8982590c3b2e432af5241ee74cd5450d64288b7b01a6140f +size 3281364 diff --git a/artifacts/policies/allegro/ppo/allegro_L_s0_early/ep00140.npz b/artifacts/policies/allegro/ppo/allegro_L_s0_early/ep00140.npz new file mode 100644 index 0000000000000000000000000000000000000000..3f0b60202e871eafd1c4f4c0b59761925a1cbb86 --- /dev/null +++ b/artifacts/policies/allegro/ppo/allegro_L_s0_early/ep00140.npz @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:50e4647c753046e561ae972bbfc18d81b651b45e070c02576391635cf6569b8e +size 3281364 diff --git a/artifacts/policies/allegro/ppo/allegro_L_s0_early/ep00145.npz b/artifacts/policies/allegro/ppo/allegro_L_s0_early/ep00145.npz new file mode 100644 index 0000000000000000000000000000000000000000..a5427cd32dfb0a3e79c2dabbc4c0dcc45c12237b --- /dev/null +++ b/artifacts/policies/allegro/ppo/allegro_L_s0_early/ep00145.npz @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:beafac9e8991686f3e7ba5d802abe893eb1442ba13e42a5e231bc086f9c46508 +size 3281360 diff --git a/artifacts/policies/allegro/ppo/allegro_L_s0_early/ep00150.npz b/artifacts/policies/allegro/ppo/allegro_L_s0_early/ep00150.npz new file mode 100644 index 0000000000000000000000000000000000000000..7a4be2fc6c3fdef6f7bd8c1b5db3c929bf5286f3 --- /dev/null +++ b/artifacts/policies/allegro/ppo/allegro_L_s0_early/ep00150.npz @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:fdda749f4e5dc0755819348a8e70edce3c72e3b3116edf33e87f044d319d49f3 +size 3281364 diff --git a/artifacts/policies/allegro/ppo/allegro_L_s0_early/ep00155.npz b/artifacts/policies/allegro/ppo/allegro_L_s0_early/ep00155.npz new file mode 100644 index 0000000000000000000000000000000000000000..6dbb7a224799d2a3def67cb851183a43e8eb2ea0 --- /dev/null +++ b/artifacts/policies/allegro/ppo/allegro_L_s0_early/ep00155.npz @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ee86e5c9ea997718d27d7d35417b96f1ec2c6150dc022546be8a86160d508b4f +size 3281360 diff --git a/artifacts/policies/allegro/ppo/allegro_L_s0_early/ep00160.npz b/artifacts/policies/allegro/ppo/allegro_L_s0_early/ep00160.npz new file mode 100644 index 0000000000000000000000000000000000000000..700f7fcadf5c72bc060d6bddbb14abc9b0e458ae --- /dev/null +++ b/artifacts/policies/allegro/ppo/allegro_L_s0_early/ep00160.npz @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f2cb000dc6441e18dc2c3db9552aa69ffdb6b6c6000d255c61dfedc7de5e239a +size 3281360 diff --git a/artifacts/policies/allegro/ppo/allegro_L_s0_early/ep00165.npz b/artifacts/policies/allegro/ppo/allegro_L_s0_early/ep00165.npz new file mode 100644 index 0000000000000000000000000000000000000000..74b7b29d62a83956006a2dcd5df5d38d02db5fa0 --- /dev/null +++ b/artifacts/policies/allegro/ppo/allegro_L_s0_early/ep00165.npz @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:bd3305b4c63bdd156328ec2229485d30e4f3ab7e462ea3b02b5520b2e3f6626c +size 3281360 diff --git a/artifacts/policies/allegro/ppo/allegro_L_s0_early/ep00170.npz b/artifacts/policies/allegro/ppo/allegro_L_s0_early/ep00170.npz new file mode 100644 index 0000000000000000000000000000000000000000..b080818bb3c10903fe6e982fff213cc857c60c6a --- /dev/null +++ b/artifacts/policies/allegro/ppo/allegro_L_s0_early/ep00170.npz @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4515bc6614a29a44f05cabfe971b13e6863bd2381c8a990093e88d53483ebaa4 +size 3281360 diff --git a/artifacts/policies/allegro/ppo/allegro_L_s0_early/ep00175.npz b/artifacts/policies/allegro/ppo/allegro_L_s0_early/ep00175.npz new file mode 100644 index 0000000000000000000000000000000000000000..93f0719a510fa1e1d90bf10dc23986435511cfbb --- /dev/null +++ b/artifacts/policies/allegro/ppo/allegro_L_s0_early/ep00175.npz @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4633f84fc0ac0e5dadd08ad0eeb53023b7ed7512537142485ab0a845860fa324 +size 3281356 diff --git a/artifacts/policies/allegro/ppo/allegro_L_s0_early/ep00180.npz b/artifacts/policies/allegro/ppo/allegro_L_s0_early/ep00180.npz new file mode 100644 index 0000000000000000000000000000000000000000..d41d4d964c196b94b89028634aab347f12272449 --- /dev/null +++ b/artifacts/policies/allegro/ppo/allegro_L_s0_early/ep00180.npz @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:cbf0acb9fbcf70ff8388ff1039ef3d681cd14d398e060f8db0b635d33bcaf932 +size 3281360 diff --git a/artifacts/policies/allegro/ppo/allegro_L_s0_early/ep00185.npz b/artifacts/policies/allegro/ppo/allegro_L_s0_early/ep00185.npz new file mode 100644 index 0000000000000000000000000000000000000000..6690c1a1de10d4103e15a495f80f89d30cc4a084 --- /dev/null +++ b/artifacts/policies/allegro/ppo/allegro_L_s0_early/ep00185.npz @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ab95b2725e86eaee0e42521897ac68094062a1981da35f9670d5fbae5787d611 +size 3281360 diff --git a/artifacts/policies/allegro/ppo/allegro_L_s0_early/ep00190.npz b/artifacts/policies/allegro/ppo/allegro_L_s0_early/ep00190.npz new file mode 100644 index 0000000000000000000000000000000000000000..a787d566220f0ec5f41bcf91cfd6225eeb37b745 --- /dev/null +++ b/artifacts/policies/allegro/ppo/allegro_L_s0_early/ep00190.npz @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:98a9ab06300b7c294e5f602f514501acd6d37b8a9569db90cd8d09d1b9234b93 +size 3281360 diff --git a/artifacts/policies/allegro/ppo/allegro_L_s0_early/ep00195.npz b/artifacts/policies/allegro/ppo/allegro_L_s0_early/ep00195.npz new file mode 100644 index 0000000000000000000000000000000000000000..3f73bd9b733b65a5700a792c6b5fd97227bb480a --- /dev/null +++ b/artifacts/policies/allegro/ppo/allegro_L_s0_early/ep00195.npz @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:416415e2a289fb1fed350ac57b3f4445f1045b023ebe5931969ce4be41f76da9 +size 3281356 diff --git a/artifacts/policies/allegro/ppo/allegro_L_s0_early/ep00200.npz b/artifacts/policies/allegro/ppo/allegro_L_s0_early/ep00200.npz new file mode 100644 index 0000000000000000000000000000000000000000..6ce17c5386d413f425abc54f882a56a6da3e2915 --- /dev/null +++ b/artifacts/policies/allegro/ppo/allegro_L_s0_early/ep00200.npz @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:49e27a1f639c6a08bfaf79874daa1adf813bcc24a6c0013758c9570b28210e50 +size 3281368 diff --git a/artifacts/policies/allegro/ppo/allegro_L_s0_early/final.npz b/artifacts/policies/allegro/ppo/allegro_L_s0_early/final.npz new file mode 100644 index 0000000000000000000000000000000000000000..6ce17c5386d413f425abc54f882a56a6da3e2915 --- /dev/null +++ b/artifacts/policies/allegro/ppo/allegro_L_s0_early/final.npz @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:49e27a1f639c6a08bfaf79874daa1adf813bcc24a6c0013758c9570b28210e50 +size 3281368 diff --git a/artifacts/policies/allegro/repair/L/damaged/norm_shift_53.npz b/artifacts/policies/allegro/repair/L/damaged/norm_shift_53.npz new file mode 100644 index 0000000000000000000000000000000000000000..76a77e5551d2c78b3e1ae9b78d23b640ab8d7d6c --- /dev/null +++ b/artifacts/policies/allegro/repair/L/damaged/norm_shift_53.npz @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:0c0a717b5194e7344188ef6d70a7cbe4670c90f38b4da80c9dc77cea178cde47 +size 3281860 diff --git a/artifacts/policies/allegro/repair/L/damaged/swap_act_4_7.npz b/artifacts/policies/allegro/repair/L/damaged/swap_act_4_7.npz new file mode 100644 index 0000000000000000000000000000000000000000..6f8bdc16e326f8c92372eaad984bb6cb5a558d0c --- /dev/null +++ b/artifacts/policies/allegro/repair/L/damaged/swap_act_4_7.npz @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7d50544613eb33aa4360de409115fddeb7bca50dd42f9c5ce278301a5f12f14f +size 3281928 diff --git a/artifacts/policies/allegro/repair/L/natural/early_ep200.npz b/artifacts/policies/allegro/repair/L/natural/early_ep200.npz new file mode 100644 index 0000000000000000000000000000000000000000..6ce17c5386d413f425abc54f882a56a6da3e2915 --- /dev/null +++ b/artifacts/policies/allegro/repair/L/natural/early_ep200.npz @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:49e27a1f639c6a08bfaf79874daa1adf813bcc24a6c0013758c9570b28210e50 +size 3281368 diff --git a/artifacts/policies/allegro/repair/L/natural/early_ep50.npz b/artifacts/policies/allegro/repair/L/natural/early_ep50.npz new file mode 100644 index 0000000000000000000000000000000000000000..14245530e23b058145352ca6004a04fda3bb5c91 --- /dev/null +++ b/artifacts/policies/allegro/repair/L/natural/early_ep50.npz @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:9864b0300d14c2d82918b0ee593efb4224f78c568ac7ca5da3d3b7442c17dac2 +size 3281348 diff --git a/artifacts/policies/franka/broken/nvidia/dead_units_0_0p5.npz b/artifacts/policies/franka/broken/nvidia/dead_units_0_0p5.npz new file mode 100644 index 0000000000000000000000000000000000000000..0231235a29ba28d868771a3dbc70d51cba47da1f --- /dev/null +++ b/artifacts/policies/franka/broken/nvidia/dead_units_0_0p5.npz @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:04b8ebc89662881aff7b99412b7cef54e4a12ac93f94e426bc94ffdf6ffaa4d5 +size 198080 diff --git a/artifacts/policies/franka/broken/nvidia/dead_units_1_0p5.npz b/artifacts/policies/franka/broken/nvidia/dead_units_1_0p5.npz new file mode 100644 index 0000000000000000000000000000000000000000..7f116d386691b7453bf4781849f43d65537cc2a6 --- /dev/null +++ b/artifacts/policies/franka/broken/nvidia/dead_units_1_0p5.npz @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:579041b5e23ad1d48299e0d691501b572a933180073146f3a8aeb2921c5f2ffe +size 196792 diff --git a/artifacts/policies/franka/broken/nvidia/flip_obs_18.npz b/artifacts/policies/franka/broken/nvidia/flip_obs_18.npz new file mode 100644 index 0000000000000000000000000000000000000000..041ec938325c0d971f93d5f540a24518e88b093b --- /dev/null +++ b/artifacts/policies/franka/broken/nvidia/flip_obs_18.npz @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e05ec7d90eec94d80c54e5ef7a22d96f55966c4231738f2605d0a662cda5e653 +size 195676 diff --git a/artifacts/policies/franka/broken/nvidia/flip_obs_20.npz b/artifacts/policies/franka/broken/nvidia/flip_obs_20.npz new file mode 100644 index 0000000000000000000000000000000000000000..aeaf4551dd431f1f1d85d35dc0af12103e67dd51 --- /dev/null +++ b/artifacts/policies/franka/broken/nvidia/flip_obs_20.npz @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:41b6fe92e4e5c6d964f75e8ed6c51882e83fba69c994bf57e586c519bfe5b571 +size 195676 diff --git a/artifacts/policies/franka/broken/nvidia/noise_0p5.npz b/artifacts/policies/franka/broken/nvidia/noise_0p5.npz new file mode 100644 index 0000000000000000000000000000000000000000..0e7af31bd37990eedd65957219851ced5b24812c --- /dev/null +++ b/artifacts/policies/franka/broken/nvidia/noise_0p5.npz @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7b7edf6b15c7958c9b25644b56556f13ce6eec7d12d2fdbeeb75295ef4875cb3 +size 195688 diff --git a/artifacts/policies/franka/broken/nvidia/noise_1p0.npz b/artifacts/policies/franka/broken/nvidia/noise_1p0.npz new file mode 100644 index 0000000000000000000000000000000000000000..82ac3f3b9c338447a80cc130928e5e2161016fc6 --- /dev/null +++ b/artifacts/policies/franka/broken/nvidia/noise_1p0.npz @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ea47dcdedef23d6d9258d70e9a83a48de3796f533747a195f0cb72a2ca26c9cf +size 195688 diff --git a/artifacts/policies/franka/broken/nvidia/norm_shift_20.npz b/artifacts/policies/franka/broken/nvidia/norm_shift_20.npz new file mode 100644 index 0000000000000000000000000000000000000000..57b7343d6094d1dd3b765a3cdf86f76a974f7b32 --- /dev/null +++ b/artifacts/policies/franka/broken/nvidia/norm_shift_20.npz @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:526f0681f94a24179e306dee3962eb9d6c647f63ae0d667c591b08d3d0e5321e +size 195692 diff --git a/artifacts/policies/franka/broken/nvidia/norm_shift_21.npz b/artifacts/policies/franka/broken/nvidia/norm_shift_21.npz new file mode 100644 index 0000000000000000000000000000000000000000..a009d92f294179a7e930af0edc83bd57a8690d05 --- /dev/null +++ b/artifacts/policies/franka/broken/nvidia/norm_shift_21.npz @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ead4935799c5181033ffacda6fb8b62f820d3c88a8bc2c9716d0152f6372c0fa +size 195692 diff --git a/artifacts/policies/franka/broken/nvidia/swap_act_0_6.npz b/artifacts/policies/franka/broken/nvidia/swap_act_0_6.npz new file mode 100644 index 0000000000000000000000000000000000000000..8c7e0c68bf537f5b14fd5f9217738ce9e013529f --- /dev/null +++ b/artifacts/policies/franka/broken/nvidia/swap_act_0_6.npz @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:048ae7a5a105f045c83b5d1ec339570d7c00059f28cb3a06082d4f5d3feb5158 +size 195760 diff --git a/artifacts/policies/franka/broken/nvidia/swap_act_1_3.npz b/artifacts/policies/franka/broken/nvidia/swap_act_1_3.npz new file mode 100644 index 0000000000000000000000000000000000000000..9e5fed77024a6737cb81bc5d4cfb6a3399766722 --- /dev/null +++ b/artifacts/policies/franka/broken/nvidia/swap_act_1_3.npz @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f370f46ab6809dd97917ac4e3e569ffd0aa540bae4a1d6447bc82dec434b994e +size 195760 diff --git a/artifacts/policies/franka/nvidia/pretrained.npz b/artifacts/policies/franka/nvidia/pretrained.npz new file mode 100644 index 0000000000000000000000000000000000000000..41d94f3a95135202951dd5f8263fa9e4acc5cf9f --- /dev/null +++ b/artifacts/policies/franka/nvidia/pretrained.npz @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ba6647ce56efd520d509a0d42eed94d3763b15bc5d48adbffb37292d025f06bc +size 195236 diff --git a/artifacts/policies/franka/ppo/franka_M_s0/best.npz b/artifacts/policies/franka/ppo/franka_M_s0/best.npz new file mode 100644 index 0000000000000000000000000000000000000000..6bf84ec38984f9567efac5fe546db47458a8e386 --- /dev/null +++ b/artifacts/policies/franka/ppo/franka_M_s0/best.npz @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e9ec0d6ffc9fb8c50364752b686875bac3a651e7b9b63cc0beb84a0347caddea +size 195152 diff --git a/artifacts/policies/franka/ppo/franka_M_s0/ep00050.npz b/artifacts/policies/franka/ppo/franka_M_s0/ep00050.npz new file mode 100644 index 0000000000000000000000000000000000000000..8fd547c6cf2ee16e4371e966deacda3afafb36e9 --- /dev/null +++ b/artifacts/policies/franka/ppo/franka_M_s0/ep00050.npz @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e6114c5d6539857a117cffd47b2c0c2de51270c86326dd7a077fd02f036eb6d1 +size 195240 diff --git a/artifacts/policies/franka/ppo/franka_M_s0/ep00100.npz b/artifacts/policies/franka/ppo/franka_M_s0/ep00100.npz new file mode 100644 index 0000000000000000000000000000000000000000..2bfd461b206bdaab637b8735f4af6ed27b801f55 --- /dev/null +++ b/artifacts/policies/franka/ppo/franka_M_s0/ep00100.npz @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e65be9e303953f6051a044f738d2f95509ae3ca37b4b7df45b8c9128adedc01c +size 195252 diff --git a/artifacts/policies/franka/ppo/franka_M_s0/ep00150.npz b/artifacts/policies/franka/ppo/franka_M_s0/ep00150.npz new file mode 100644 index 0000000000000000000000000000000000000000..41a87c21c22b95f6c61b79cd4b092ed84144b6e9 --- /dev/null +++ b/artifacts/policies/franka/ppo/franka_M_s0/ep00150.npz @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b6d3eda2dd5ed127dc958d0f7d9ba84b8b5a6f16690d88711fd53b2226ceff17 +size 195252 diff --git a/artifacts/policies/franka/ppo/franka_M_s0/ep00200.npz b/artifacts/policies/franka/ppo/franka_M_s0/ep00200.npz new file mode 100644 index 0000000000000000000000000000000000000000..69a5cc040c7f8bc75adeca61a7fb4206f48f8af7 --- /dev/null +++ b/artifacts/policies/franka/ppo/franka_M_s0/ep00200.npz @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f1020ccc3254f58d301e2281a2cc312a434a9ecb4e72d0a143d388bd396395e5 +size 195256 diff --git a/artifacts/policies/franka/ppo/franka_M_s0/ep00250.npz b/artifacts/policies/franka/ppo/franka_M_s0/ep00250.npz new file mode 100644 index 0000000000000000000000000000000000000000..92ac0e5e8ad9655d47cf26c2a6ada6508234d07b --- /dev/null +++ b/artifacts/policies/franka/ppo/franka_M_s0/ep00250.npz @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:499717c9a688ec198f573b35388e0c6c37611721ceee6d55208fe6609464c47d +size 195256 diff --git a/artifacts/policies/franka/ppo/franka_M_s0/ep00300.npz b/artifacts/policies/franka/ppo/franka_M_s0/ep00300.npz new file mode 100644 index 0000000000000000000000000000000000000000..301e603175055df34b5059407c18fdda62c3ee7f --- /dev/null +++ b/artifacts/policies/franka/ppo/franka_M_s0/ep00300.npz @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:555a19029e2ddb4f0bba120c28abfc48b25d6fb6b10a9f73c8fa416c167f77f1 +size 195252 diff --git a/artifacts/policies/franka/ppo/franka_M_s0/ep00350.npz b/artifacts/policies/franka/ppo/franka_M_s0/ep00350.npz new file mode 100644 index 0000000000000000000000000000000000000000..6d6f4a6489777a2648cfd225b6319c7422335269 --- /dev/null +++ b/artifacts/policies/franka/ppo/franka_M_s0/ep00350.npz @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b29e94b86d62ad2628225b0e7f6459811e60cc880b5aa1b9c2fc16a314ebd929 +size 195256 diff --git a/artifacts/policies/franka/ppo/franka_M_s0/ep00400.npz b/artifacts/policies/franka/ppo/franka_M_s0/ep00400.npz new file mode 100644 index 0000000000000000000000000000000000000000..bf29ae1d8f00641332162993b82da2bcd35c47e4 --- /dev/null +++ b/artifacts/policies/franka/ppo/franka_M_s0/ep00400.npz @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:fb5b0b2292080c6ecb1e1dfc058685e50ff47b1b36c76cdc65426dcb1899599b +size 195256 diff --git a/artifacts/policies/franka/ppo/franka_M_s0/ep00450.npz b/artifacts/policies/franka/ppo/franka_M_s0/ep00450.npz new file mode 100644 index 0000000000000000000000000000000000000000..de82a429654397d16dcf20435a2ffc9313b979d9 --- /dev/null +++ b/artifacts/policies/franka/ppo/franka_M_s0/ep00450.npz @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b12cdecb04002635ab6ead5b0f1881dd2e979efe1dc402115f74e0af2be34ed4 +size 195256 diff --git a/artifacts/policies/franka/ppo/franka_M_s0/ep00500.npz b/artifacts/policies/franka/ppo/franka_M_s0/ep00500.npz new file mode 100644 index 0000000000000000000000000000000000000000..b5325d4882fcd5c0d6634672fb89bd6df9edfac8 --- /dev/null +++ b/artifacts/policies/franka/ppo/franka_M_s0/ep00500.npz @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:5e19af9cf2a1452f826c4d525f4cda60b1e5e344a460131d91f9dafcaabeab21 +size 195252 diff --git a/artifacts/policies/franka/ppo/franka_M_s0/ep00550.npz b/artifacts/policies/franka/ppo/franka_M_s0/ep00550.npz new file mode 100644 index 0000000000000000000000000000000000000000..338f973ff96e9799f34e099b8ee894d7da3458fc --- /dev/null +++ b/artifacts/policies/franka/ppo/franka_M_s0/ep00550.npz @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:07d025a059e0c13773940d6ffe942a6c6bbeb3a7a6a5bd80724fb898fc42c3da +size 195256 diff --git a/artifacts/policies/franka/ppo/franka_M_s0/ep00600.npz b/artifacts/policies/franka/ppo/franka_M_s0/ep00600.npz new file mode 100644 index 0000000000000000000000000000000000000000..eb2d45bdba928bf93eb162a227e2864417e825b0 --- /dev/null +++ b/artifacts/policies/franka/ppo/franka_M_s0/ep00600.npz @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:bdb6f6153a399fcf443c80d9fab965b567b221b2ee27211d14828d84fa2445c0 +size 195252 diff --git a/artifacts/policies/franka/ppo/franka_M_s0/ep00650.npz b/artifacts/policies/franka/ppo/franka_M_s0/ep00650.npz new file mode 100644 index 0000000000000000000000000000000000000000..ab78189ab19a180b27985e18400c1e91979e7dba --- /dev/null +++ b/artifacts/policies/franka/ppo/franka_M_s0/ep00650.npz @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:471d0906e755d9181aa7a4eac0bd1e2ab0f21ecf095a4f0895c1e847d6a24035 +size 195256 diff --git a/artifacts/policies/franka/ppo/franka_M_s0/ep00700.npz b/artifacts/policies/franka/ppo/franka_M_s0/ep00700.npz new file mode 100644 index 0000000000000000000000000000000000000000..a0e9d96b3f7b94b53e79f75c13939d0581bd4ad7 --- /dev/null +++ b/artifacts/policies/franka/ppo/franka_M_s0/ep00700.npz @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4dbe083b60ff9b19b720dd4893aa03355afe679aca8853fe3c93ebf6441f00eb +size 195256 diff --git a/artifacts/policies/franka/ppo/franka_M_s0/ep00750.npz b/artifacts/policies/franka/ppo/franka_M_s0/ep00750.npz new file mode 100644 index 0000000000000000000000000000000000000000..933b54a082aea067867db282727cde8ecc578484 --- /dev/null +++ b/artifacts/policies/franka/ppo/franka_M_s0/ep00750.npz @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:328c6615ae3e3027af99cc8090e9e7f58e6f6bca7376d526d767691cda198727 +size 195256 diff --git a/artifacts/policies/franka/ppo/franka_M_s0/ep00800.npz b/artifacts/policies/franka/ppo/franka_M_s0/ep00800.npz new file mode 100644 index 0000000000000000000000000000000000000000..b56943123b05eee58d3034e2b04e9f7006a7243b --- /dev/null +++ b/artifacts/policies/franka/ppo/franka_M_s0/ep00800.npz @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b843b149cdb97ae5ab47e874ac9462dd799a72e1657f8e8c71edd30bc1abb5c0 +size 195256 diff --git a/artifacts/policies/franka/ppo/franka_M_s0/ep00850.npz b/artifacts/policies/franka/ppo/franka_M_s0/ep00850.npz new file mode 100644 index 0000000000000000000000000000000000000000..1a0afea9117c97ca870ee96e3d965189fb8b9eeb --- /dev/null +++ b/artifacts/policies/franka/ppo/franka_M_s0/ep00850.npz @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:8cc264293108216e740ed99c513db4c580749a5f850fb604c0415bed6dc02d18 +size 195256 diff --git a/artifacts/policies/franka/ppo/franka_M_s0/ep00900.npz b/artifacts/policies/franka/ppo/franka_M_s0/ep00900.npz new file mode 100644 index 0000000000000000000000000000000000000000..42e86f77a74a302e2b02bd9e21c9ddfd398c41bc --- /dev/null +++ b/artifacts/policies/franka/ppo/franka_M_s0/ep00900.npz @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:8846903a32033a0305b32ecfddcb5a314c4e41bfcf3c3d15da98a7cd56fdeea7 +size 195256 diff --git a/artifacts/policies/franka/ppo/franka_M_s0/ep00950.npz b/artifacts/policies/franka/ppo/franka_M_s0/ep00950.npz new file mode 100644 index 0000000000000000000000000000000000000000..56a52409181f89b756e1c459a16c14cba415c3bf --- /dev/null +++ b/artifacts/policies/franka/ppo/franka_M_s0/ep00950.npz @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:40116ab6e19a882f0e581dcfca6e51d5537e27190a04a9905972e03a5146d567 +size 195256 diff --git a/artifacts/policies/franka/ppo/franka_M_s0/ep01000.npz b/artifacts/policies/franka/ppo/franka_M_s0/ep01000.npz new file mode 100644 index 0000000000000000000000000000000000000000..77a88002a329f6e6475a1c40eb17d6ffd6fb7a60 --- /dev/null +++ b/artifacts/policies/franka/ppo/franka_M_s0/ep01000.npz @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6ff864a87778d473236990ce129f36594306caf89d421437ffa36972c39c3dc2 +size 195264 diff --git a/artifacts/policies/franka/ppo/franka_M_s0/ep01050.npz b/artifacts/policies/franka/ppo/franka_M_s0/ep01050.npz new file mode 100644 index 0000000000000000000000000000000000000000..6135bc8c5894a937aca838ee3fdbdd3ff873664c --- /dev/null +++ b/artifacts/policies/franka/ppo/franka_M_s0/ep01050.npz @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:fe7871a81b1f12e5a2497520ec0b915dbe9d800055e9daa7888b9c9f0f21322b +size 195260 diff --git a/artifacts/policies/franka/ppo/franka_M_s0/ep01100.npz b/artifacts/policies/franka/ppo/franka_M_s0/ep01100.npz new file mode 100644 index 0000000000000000000000000000000000000000..1acaead8751ab72a7d6ef31b36c711805723ad28 --- /dev/null +++ b/artifacts/policies/franka/ppo/franka_M_s0/ep01100.npz @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b40c8bc4de3577ea6589eeec1547c7a26095a3dbc89cfa612ce2c89469ea9eb9 +size 195264 diff --git a/artifacts/policies/franka/ppo/franka_M_s0/ep01150.npz b/artifacts/policies/franka/ppo/franka_M_s0/ep01150.npz new file mode 100644 index 0000000000000000000000000000000000000000..3dca7c60849a1f2643a1f9d6b2c5c89b97a1f308 --- /dev/null +++ b/artifacts/policies/franka/ppo/franka_M_s0/ep01150.npz @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:bbe75080e851fb6743cd9abc359a698dd9376cb4d45332adce5facffa70de922 +size 195256 diff --git a/artifacts/policies/franka/ppo/franka_M_s0/ep01200.npz b/artifacts/policies/franka/ppo/franka_M_s0/ep01200.npz new file mode 100644 index 0000000000000000000000000000000000000000..95abe43ad89889f5af80806b62fff7700dfd3bb1 --- /dev/null +++ b/artifacts/policies/franka/ppo/franka_M_s0/ep01200.npz @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:73043d250a15220b6c16017d88af5bbf4ce75c660f69be07868429997f2f3ec1 +size 195264 diff --git a/artifacts/policies/franka/ppo/franka_M_s0/ep01250.npz b/artifacts/policies/franka/ppo/franka_M_s0/ep01250.npz new file mode 100644 index 0000000000000000000000000000000000000000..45f105686d06b6fc6fe8dc024974bc176dc9d946 --- /dev/null +++ b/artifacts/policies/franka/ppo/franka_M_s0/ep01250.npz @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:cf51bb09d70b2a4bd777d00ac585a24a9dc1317497166e1de0ffb87d8585f63e +size 195264 diff --git a/artifacts/policies/franka/ppo/franka_M_s0/ep01300.npz b/artifacts/policies/franka/ppo/franka_M_s0/ep01300.npz new file mode 100644 index 0000000000000000000000000000000000000000..afc91aaca8713e16338e083be4e196c5c007ba55 --- /dev/null +++ b/artifacts/policies/franka/ppo/franka_M_s0/ep01300.npz @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:89ef0496f1c14cba151e57b2a1c12d12186696e7796eb7ddb82b04b0d0ffaa1b +size 195264 diff --git a/artifacts/policies/franka/ppo/franka_M_s0/ep01350.npz b/artifacts/policies/franka/ppo/franka_M_s0/ep01350.npz new file mode 100644 index 0000000000000000000000000000000000000000..42b63a30684e0d668fce6bc6a5ce457339586434 --- /dev/null +++ b/artifacts/policies/franka/ppo/franka_M_s0/ep01350.npz @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:76a4a48340d02ce2d7b90920b8b6f8b9e9188620911680a06cd21fc0f06345ef +size 195260 diff --git a/artifacts/policies/franka/ppo/franka_M_s0/ep01400.npz b/artifacts/policies/franka/ppo/franka_M_s0/ep01400.npz new file mode 100644 index 0000000000000000000000000000000000000000..3c7fd10026f58e3b90c9faf496ecaefcbd5dd14d --- /dev/null +++ b/artifacts/policies/franka/ppo/franka_M_s0/ep01400.npz @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:56c29bd7d50f9df1595beff48e79d27d64ac1b4678cb722cdd0e735345aa179a +size 195264 diff --git a/artifacts/policies/franka/ppo/franka_M_s0/ep01450.npz b/artifacts/policies/franka/ppo/franka_M_s0/ep01450.npz new file mode 100644 index 0000000000000000000000000000000000000000..85745a87011f730430bdf64aa17dd7765f2d8cb4 --- /dev/null +++ b/artifacts/policies/franka/ppo/franka_M_s0/ep01450.npz @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c17499dfeb79f608645a9a613ad8dcaaec2b1284556ac2ec171a80073cbe5679 +size 195264 diff --git a/artifacts/policies/franka/ppo/franka_M_s0/ep01500.npz b/artifacts/policies/franka/ppo/franka_M_s0/ep01500.npz new file mode 100644 index 0000000000000000000000000000000000000000..a6faf4637bf304c0c2cee21f4502f6d464d1ed47 --- /dev/null +++ b/artifacts/policies/franka/ppo/franka_M_s0/ep01500.npz @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:478eb1eec676b8038ebeaea8b567f22904f56cc792dfb18b80142bdcff814e3d +size 195268 diff --git a/artifacts/policies/franka/ppo/franka_M_s0/final.npz b/artifacts/policies/franka/ppo/franka_M_s0/final.npz new file mode 100644 index 0000000000000000000000000000000000000000..a6faf4637bf304c0c2cee21f4502f6d464d1ed47 --- /dev/null +++ b/artifacts/policies/franka/ppo/franka_M_s0/final.npz @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:478eb1eec676b8038ebeaea8b567f22904f56cc792dfb18b80142bdcff814e3d +size 195268 diff --git a/artifacts/policies/franka/ppo/franka_M_s0_early/ep00015.npz b/artifacts/policies/franka/ppo/franka_M_s0_early/ep00015.npz new file mode 100644 index 0000000000000000000000000000000000000000..60432f8e33622fcfba2083f5ff1ff4d13890a5b3 --- /dev/null +++ b/artifacts/policies/franka/ppo/franka_M_s0_early/ep00015.npz @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:947d0248e901b216aa337bfffb81c609727eff2c8f8f7dee315b0ac658eba353 +size 195284 diff --git a/artifacts/policies/franka/ppo/franka_M_s0_early/ep00020.npz b/artifacts/policies/franka/ppo/franka_M_s0_early/ep00020.npz new file mode 100644 index 0000000000000000000000000000000000000000..f06469539c4b305448703f0c3c6c75b4b9a6cff4 --- /dev/null +++ b/artifacts/policies/franka/ppo/franka_M_s0_early/ep00020.npz @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7714181a1cfef781267cd42ae9df7cc6b6a3104d642e7b5d5629a35938ff898f +size 195292 diff --git a/artifacts/policies/franka/ppo/franka_M_s0_early/ep00025.npz b/artifacts/policies/franka/ppo/franka_M_s0_early/ep00025.npz new file mode 100644 index 0000000000000000000000000000000000000000..dd121a67ccf107a02752380f5809d6051c789d2a --- /dev/null +++ b/artifacts/policies/franka/ppo/franka_M_s0_early/ep00025.npz @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f30841d5f3d3f4636c5bedb684bc8682eb013b282a4f1274274f3851b50c95f4 +size 195288 diff --git a/artifacts/policies/franka/ppo/franka_M_s0_early/ep00030.npz b/artifacts/policies/franka/ppo/franka_M_s0_early/ep00030.npz new file mode 100644 index 0000000000000000000000000000000000000000..24b5a02dd806ee809ebeaae8bc5e565e93d45aa3 --- /dev/null +++ b/artifacts/policies/franka/ppo/franka_M_s0_early/ep00030.npz @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:da9bb6cad7e6e2d6d3e96edf23825c237ed4021b6549aa14111c54524a234512 +size 195292 diff --git a/artifacts/policies/franka/ppo/franka_M_s0_early/ep00035.npz b/artifacts/policies/franka/ppo/franka_M_s0_early/ep00035.npz new file mode 100644 index 0000000000000000000000000000000000000000..b6773e9eff1ff6cf83dafd1a6b5c8ec9d341fc66 --- /dev/null +++ b/artifacts/policies/franka/ppo/franka_M_s0_early/ep00035.npz @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c909f0d61cce86e35dfc79643136366f8563107b61b5cb164cc8ed08844a3939 +size 195292 diff --git a/artifacts/policies/franka/ppo/franka_M_s0_early/ep00040.npz b/artifacts/policies/franka/ppo/franka_M_s0_early/ep00040.npz new file mode 100644 index 0000000000000000000000000000000000000000..9e72a08b274826603e351dd84c580b26bb01f0b8 --- /dev/null +++ b/artifacts/policies/franka/ppo/franka_M_s0_early/ep00040.npz @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:71b11e761d8b3bd89caeeae97a714c8ad91fe4b6a8d8954d21f1d91650f53a1d +size 195292 diff --git a/artifacts/policies/franka/ppo/franka_M_s0_early/ep00045.npz b/artifacts/policies/franka/ppo/franka_M_s0_early/ep00045.npz new file mode 100644 index 0000000000000000000000000000000000000000..36aabe61adb88539f95d091d07542c35ceeb1736 --- /dev/null +++ b/artifacts/policies/franka/ppo/franka_M_s0_early/ep00045.npz @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:53e498ac22f68f45e656fd8ba0fec558c898dc6a7d642a424695e4413abe3e17 +size 195292 diff --git a/artifacts/policies/franka/ppo/franka_M_s0_early/ep00050.npz b/artifacts/policies/franka/ppo/franka_M_s0_early/ep00050.npz new file mode 100644 index 0000000000000000000000000000000000000000..278da43fd31e24c7db6754f0c865c342fb6e1da9 --- /dev/null +++ b/artifacts/policies/franka/ppo/franka_M_s0_early/ep00050.npz @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:59ae57e332b786478e45629f60d802ddd51a747058bf498d1af8752823f3dc13 +size 195288 diff --git a/artifacts/policies/franka/ppo/franka_M_s0_early/ep00055.npz b/artifacts/policies/franka/ppo/franka_M_s0_early/ep00055.npz new file mode 100644 index 0000000000000000000000000000000000000000..8e0731500f27b63485392e7f4047f8d0edb4bc08 --- /dev/null +++ b/artifacts/policies/franka/ppo/franka_M_s0_early/ep00055.npz @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:117e3cd962ba2b8bdf172dcb456ec5ff4053bbc50dc5b4da1756e418e7ea0396 +size 195288 diff --git a/artifacts/policies/franka/ppo/franka_M_s0_early/ep00060.npz b/artifacts/policies/franka/ppo/franka_M_s0_early/ep00060.npz new file mode 100644 index 0000000000000000000000000000000000000000..572bfba954856255183a110d713e7bd6a22904f3 --- /dev/null +++ b/artifacts/policies/franka/ppo/franka_M_s0_early/ep00060.npz @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:683a819d25974976a1ed2d61b7ddc8e430ebae5c6200e08c54f78bd1146499af +size 195300 diff --git a/artifacts/policies/franka/ppo/franka_M_s0_early/final.npz b/artifacts/policies/franka/ppo/franka_M_s0_early/final.npz new file mode 100644 index 0000000000000000000000000000000000000000..572bfba954856255183a110d713e7bd6a22904f3 --- /dev/null +++ b/artifacts/policies/franka/ppo/franka_M_s0_early/final.npz @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:683a819d25974976a1ed2d61b7ddc8e430ebae5c6200e08c54f78bd1146499af +size 195300 diff --git a/artifacts/policies/franka/ppo/franka_XS_s0/best.npz b/artifacts/policies/franka/ppo/franka_XS_s0/best.npz new file mode 100644 index 0000000000000000000000000000000000000000..a703ce2ed92a6245f9bed8bad129ab07b70c5dbb --- /dev/null +++ b/artifacts/policies/franka/ppo/franka_XS_s0/best.npz @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:72f8d29aaa20b8edef7e5baaccd247b5018524d7202bcaedb2106863d633a172 +size 28284 diff --git a/artifacts/policies/franka/ppo/franka_XS_s0/ep00050.npz b/artifacts/policies/franka/ppo/franka_XS_s0/ep00050.npz new file mode 100644 index 0000000000000000000000000000000000000000..d66541f9bb92a6e9a57edc56e2a9adb0093b9f00 --- /dev/null +++ b/artifacts/policies/franka/ppo/franka_XS_s0/ep00050.npz @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:0bac0b60b803ff32b116c1a9ed3368403dc00ee877680b95241a6f3b104a5d15 +size 28376 diff --git a/artifacts/policies/franka/ppo/franka_XS_s0/ep00100.npz b/artifacts/policies/franka/ppo/franka_XS_s0/ep00100.npz new file mode 100644 index 0000000000000000000000000000000000000000..125add8890113644e27bf408124db90d55984aed --- /dev/null +++ b/artifacts/policies/franka/ppo/franka_XS_s0/ep00100.npz @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:037936e786a985dd868c9150250caa76c2cd8cfad53883ec083cdab127e7ad88 +size 28384 diff --git a/artifacts/policies/franka/ppo/franka_XS_s0/ep00150.npz b/artifacts/policies/franka/ppo/franka_XS_s0/ep00150.npz new file mode 100644 index 0000000000000000000000000000000000000000..f6153fa2aa13410216b1a5d295bd88e29aa22f2d --- /dev/null +++ b/artifacts/policies/franka/ppo/franka_XS_s0/ep00150.npz @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3ee88a008b06a1f505a280e7092362b51a2c1ec60568dab1a65b697208458059 +size 28384 diff --git a/artifacts/policies/franka/ppo/franka_XS_s0/ep00200.npz b/artifacts/policies/franka/ppo/franka_XS_s0/ep00200.npz new file mode 100644 index 0000000000000000000000000000000000000000..d553ae1f26c9761667c18480a5a18c9026e6a2ec --- /dev/null +++ b/artifacts/policies/franka/ppo/franka_XS_s0/ep00200.npz @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7de00fe6e49241c5c2ac42eb3037324b7fbefa9d3205fdee16316738c2a9cd57 +size 28388 diff --git a/artifacts/policies/franka/ppo/franka_XS_s0/ep00250.npz b/artifacts/policies/franka/ppo/franka_XS_s0/ep00250.npz new file mode 100644 index 0000000000000000000000000000000000000000..1c7bc028b8ea81b9701287598a41d710f4b08e0f --- /dev/null +++ b/artifacts/policies/franka/ppo/franka_XS_s0/ep00250.npz @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:31b0373d3889ea6d59b81933f3f019f068e7995d9053b1c545ac26a9728a35ea +size 28388 diff --git a/artifacts/policies/franka/ppo/franka_XS_s0/ep00300.npz b/artifacts/policies/franka/ppo/franka_XS_s0/ep00300.npz new file mode 100644 index 0000000000000000000000000000000000000000..66dbe30936ed177361b25e9b6922e43a496a3c03 --- /dev/null +++ b/artifacts/policies/franka/ppo/franka_XS_s0/ep00300.npz @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a4c1c47c4d39725b857091907e5b6c23236af4e036b9011f352788a4f26843c2 +size 28388 diff --git a/artifacts/policies/franka/ppo/franka_XS_s0/ep00350.npz b/artifacts/policies/franka/ppo/franka_XS_s0/ep00350.npz new file mode 100644 index 0000000000000000000000000000000000000000..1e34870325fa994976311ef52b02080d676ad1c5 --- /dev/null +++ b/artifacts/policies/franka/ppo/franka_XS_s0/ep00350.npz @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:40d025d53e54c52b795dc5c553d90fdb7332d95e62fefa287584260cf8b752a8 +size 28388 diff --git a/artifacts/policies/franka/ppo/franka_XS_s0/ep00400.npz b/artifacts/policies/franka/ppo/franka_XS_s0/ep00400.npz new file mode 100644 index 0000000000000000000000000000000000000000..166e673de7267920122b258090367c261f785005 --- /dev/null +++ b/artifacts/policies/franka/ppo/franka_XS_s0/ep00400.npz @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:1097caf4b62f19847dbeeda605f54c344a142f92e48bb9a09e9899ca48aaeeab +size 28384 diff --git a/artifacts/policies/franka/ppo/franka_XS_s0/ep00450.npz b/artifacts/policies/franka/ppo/franka_XS_s0/ep00450.npz new file mode 100644 index 0000000000000000000000000000000000000000..ce94854978bb6650ddc200411b4563556079bc9c --- /dev/null +++ b/artifacts/policies/franka/ppo/franka_XS_s0/ep00450.npz @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a3ebf99d282f24970a50be6905a053ce287aa3121ebcaac49f760e469f6caccf +size 28388 diff --git a/artifacts/policies/franka/ppo/franka_XS_s0/ep00500.npz b/artifacts/policies/franka/ppo/franka_XS_s0/ep00500.npz new file mode 100644 index 0000000000000000000000000000000000000000..dd8e21c113e9df7e35b0f9cedee89290dc2337fb --- /dev/null +++ b/artifacts/policies/franka/ppo/franka_XS_s0/ep00500.npz @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:77a87855aa4cd662a749db1e2c4656a0795bac2a0ab7d3f6adac5b19bcd23d2e +size 28388 diff --git a/artifacts/policies/franka/ppo/franka_XS_s0/ep00550.npz b/artifacts/policies/franka/ppo/franka_XS_s0/ep00550.npz new file mode 100644 index 0000000000000000000000000000000000000000..5c62b520b4d0f79beaabeacfebc69ebd86dda776 --- /dev/null +++ b/artifacts/policies/franka/ppo/franka_XS_s0/ep00550.npz @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:5cda281633d3dee552ba06689b5d620e73dfd3506dce4fc72b6d819a69ed84b7 +size 28388 diff --git a/artifacts/policies/franka/ppo/franka_XS_s0/ep00600.npz b/artifacts/policies/franka/ppo/franka_XS_s0/ep00600.npz new file mode 100644 index 0000000000000000000000000000000000000000..7b90d5fae3a7d5fb67ecfa05b73dd8669e0c2060 --- /dev/null +++ b/artifacts/policies/franka/ppo/franka_XS_s0/ep00600.npz @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e4051fb42949c72b236ca712d774409e9eb71aa64b3190b843dc4746a1969e10 +size 28388 diff --git a/artifacts/policies/franka/ppo/franka_XS_s0/ep00650.npz b/artifacts/policies/franka/ppo/franka_XS_s0/ep00650.npz new file mode 100644 index 0000000000000000000000000000000000000000..f38dbee37838980e63ffbe9e75746070d335f8c9 --- /dev/null +++ b/artifacts/policies/franka/ppo/franka_XS_s0/ep00650.npz @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b446f163d4a0a502d22c84c50e98848d0b6972f74053c11499aaa594f234a819 +size 28388 diff --git a/artifacts/policies/franka/ppo/franka_XS_s0/ep00700.npz b/artifacts/policies/franka/ppo/franka_XS_s0/ep00700.npz new file mode 100644 index 0000000000000000000000000000000000000000..48eecf33e6bb72143757a1a3c645808a7a8d8e43 --- /dev/null +++ b/artifacts/policies/franka/ppo/franka_XS_s0/ep00700.npz @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:57185f5da23c7390f580dada67d68219c52acbb1983e6e345ba82eed90d04512 +size 28384 diff --git a/artifacts/policies/franka/ppo/franka_XS_s0/ep00750.npz b/artifacts/policies/franka/ppo/franka_XS_s0/ep00750.npz new file mode 100644 index 0000000000000000000000000000000000000000..324f421413bad820b0dff4a689374f0794863cbe --- /dev/null +++ b/artifacts/policies/franka/ppo/franka_XS_s0/ep00750.npz @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:59ee38d1b7ac9a1db94a2b323d4fcdcfa5fc9aba95ee13c99a40807811f7077b +size 28388 diff --git a/artifacts/policies/franka/ppo/franka_XS_s0/ep00800.npz b/artifacts/policies/franka/ppo/franka_XS_s0/ep00800.npz new file mode 100644 index 0000000000000000000000000000000000000000..53e83ebe8c73427f28bdd69fd3214cf901fd198d --- /dev/null +++ b/artifacts/policies/franka/ppo/franka_XS_s0/ep00800.npz @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3ad6899bb4168dbc43e0b115b73382a8deb5877cae807afdcd8b1990371165d5 +size 28388 diff --git a/artifacts/policies/franka/ppo/franka_XS_s0/ep00850.npz b/artifacts/policies/franka/ppo/franka_XS_s0/ep00850.npz new file mode 100644 index 0000000000000000000000000000000000000000..11eccb7a9117aff9c54db029fee9c6938154530d --- /dev/null +++ b/artifacts/policies/franka/ppo/franka_XS_s0/ep00850.npz @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:5645ed6ea470f01e02ff12d12298a0237058af393d3925dd1389e4aa45775591 +size 28388 diff --git a/artifacts/policies/franka/ppo/franka_XS_s0/ep00900.npz b/artifacts/policies/franka/ppo/franka_XS_s0/ep00900.npz new file mode 100644 index 0000000000000000000000000000000000000000..e0bee1999d4af0256e3687cad2ba50408fa815e4 --- /dev/null +++ b/artifacts/policies/franka/ppo/franka_XS_s0/ep00900.npz @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f288aab3baacd95ed8b33af71f93e9c71731f4aa16561484d9749ce4e639aaf4 +size 28388 diff --git a/artifacts/policies/franka/ppo/franka_XS_s0/ep00950.npz b/artifacts/policies/franka/ppo/franka_XS_s0/ep00950.npz new file mode 100644 index 0000000000000000000000000000000000000000..b30fcd222e694f15992b1999c3d456869f508a18 --- /dev/null +++ b/artifacts/policies/franka/ppo/franka_XS_s0/ep00950.npz @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4eb583cc7d2be763e2b70924d486aaaa070a950e5707c113a9b0ce8b6acea808 +size 28384 diff --git a/artifacts/policies/franka/ppo/franka_XS_s0/ep01000.npz b/artifacts/policies/franka/ppo/franka_XS_s0/ep01000.npz new file mode 100644 index 0000000000000000000000000000000000000000..07ff8d9fb0b36f3ecb1c86f50d4dedf01c93f879 --- /dev/null +++ b/artifacts/policies/franka/ppo/franka_XS_s0/ep01000.npz @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:50043ae736d8e83c96d121b9f4848f2df527da1fc4d33b2a14e7bd201c43ce16 +size 28396 diff --git a/artifacts/policies/franka/ppo/franka_XS_s0/ep01050.npz b/artifacts/policies/franka/ppo/franka_XS_s0/ep01050.npz new file mode 100644 index 0000000000000000000000000000000000000000..4219f4009c64025aa7cfde5659e5c04900a4ce45 --- /dev/null +++ b/artifacts/policies/franka/ppo/franka_XS_s0/ep01050.npz @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:cefc35f790fd33223ecd77e5a7697f4107d88b1a2e93aab14a0762b471e02193 +size 28396 diff --git a/artifacts/policies/franka/ppo/franka_XS_s0/ep01100.npz b/artifacts/policies/franka/ppo/franka_XS_s0/ep01100.npz new file mode 100644 index 0000000000000000000000000000000000000000..8c5d597985b75c83aaf690fdd999e8e1d667c535 --- /dev/null +++ b/artifacts/policies/franka/ppo/franka_XS_s0/ep01100.npz @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:89113f69b1d132974465e464596af47349c8ab75fa0d8df203ff54ab0c828230 +size 28396 diff --git a/artifacts/policies/franka/ppo/franka_XS_s0/ep01150.npz b/artifacts/policies/franka/ppo/franka_XS_s0/ep01150.npz new file mode 100644 index 0000000000000000000000000000000000000000..47809efcae47c0bab12fe7831ca4582730eb41f8 --- /dev/null +++ b/artifacts/policies/franka/ppo/franka_XS_s0/ep01150.npz @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:2886fc31f55384a51df17c83175663aa307c10b38520f166848fa4e83fd22474 +size 28388 diff --git a/artifacts/policies/franka/ppo/franka_XS_s0/ep01200.npz b/artifacts/policies/franka/ppo/franka_XS_s0/ep01200.npz new file mode 100644 index 0000000000000000000000000000000000000000..66b44572237d3fea8fc0a581b939774e0a5bb66d --- /dev/null +++ b/artifacts/policies/franka/ppo/franka_XS_s0/ep01200.npz @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:754cb938a4eb3c2275c99756f217e72211c53d1375711c2da138c5af26ad7a1c +size 28396 diff --git a/artifacts/policies/franka/ppo/franka_XS_s0/ep01250.npz b/artifacts/policies/franka/ppo/franka_XS_s0/ep01250.npz new file mode 100644 index 0000000000000000000000000000000000000000..eb28f0518d5f1f121ddd3ef2bcbc6c76fae84341 --- /dev/null +++ b/artifacts/policies/franka/ppo/franka_XS_s0/ep01250.npz @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:78e18bb37342ff5593800188e33e3c7831462517bcd6af56a5a6779a7e2ce48c +size 28396 diff --git a/artifacts/policies/franka/ppo/franka_XS_s0/ep01300.npz b/artifacts/policies/franka/ppo/franka_XS_s0/ep01300.npz new file mode 100644 index 0000000000000000000000000000000000000000..58f01121bc5fb1f08ea0e03ae897afdd5a9efac2 --- /dev/null +++ b/artifacts/policies/franka/ppo/franka_XS_s0/ep01300.npz @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:252e5e81c39d6843a226b9e82709478f19ced5b4068202562641504e799be7ec +size 28392 diff --git a/artifacts/policies/franka/ppo/franka_XS_s0/ep01350.npz b/artifacts/policies/franka/ppo/franka_XS_s0/ep01350.npz new file mode 100644 index 0000000000000000000000000000000000000000..dfb152c8e5466c84abcb4a4cd66dd188f3bbb0ba --- /dev/null +++ b/artifacts/policies/franka/ppo/franka_XS_s0/ep01350.npz @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a88e0ffb90b4e68918e676c679b17c37a8f9933a327bf66813d747c6a6a78f16 +size 28396 diff --git a/artifacts/policies/franka/ppo/franka_XS_s0/ep01400.npz b/artifacts/policies/franka/ppo/franka_XS_s0/ep01400.npz new file mode 100644 index 0000000000000000000000000000000000000000..c80c5d3c88157ff357153db16d7200e7d03f29c6 --- /dev/null +++ b/artifacts/policies/franka/ppo/franka_XS_s0/ep01400.npz @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e967e600f3595f09886aac7dbed20935d48745a420cffeee6833d499f8364d13 +size 28396 diff --git a/artifacts/policies/franka/ppo/franka_XS_s0/ep01450.npz b/artifacts/policies/franka/ppo/franka_XS_s0/ep01450.npz new file mode 100644 index 0000000000000000000000000000000000000000..e090cb2a92ce59a1cebe83c9980d50bd54fabcab --- /dev/null +++ b/artifacts/policies/franka/ppo/franka_XS_s0/ep01450.npz @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:5b7c6d892a86a85abbba73597a9af3fb37e9d65e41a13197fa25ff3e6f0743a5 +size 28392 diff --git a/artifacts/policies/franka/ppo/franka_XS_s0/ep01500.npz b/artifacts/policies/franka/ppo/franka_XS_s0/ep01500.npz new file mode 100644 index 0000000000000000000000000000000000000000..49d982cfa9ef1e38938983588f4445f51cf96771 --- /dev/null +++ b/artifacts/policies/franka/ppo/franka_XS_s0/ep01500.npz @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ecd3d1a5a6facf8eba802e42ba1c8bcfd37683e963c1babd22362fdaa0d31956 +size 28404 diff --git a/artifacts/policies/franka/ppo/franka_XS_s0/final.npz b/artifacts/policies/franka/ppo/franka_XS_s0/final.npz new file mode 100644 index 0000000000000000000000000000000000000000..49d982cfa9ef1e38938983588f4445f51cf96771 --- /dev/null +++ b/artifacts/policies/franka/ppo/franka_XS_s0/final.npz @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ecd3d1a5a6facf8eba802e42ba1c8bcfd37683e963c1babd22362fdaa0d31956 +size 28404 diff --git a/artifacts/policies/franka/ppo_v2_targetbug_ARCHIVE/franka_M_s0/best.npz b/artifacts/policies/franka/ppo_v2_targetbug_ARCHIVE/franka_M_s0/best.npz new file mode 100644 index 0000000000000000000000000000000000000000..b1578ec411813bec79a6224824bb5f61c5773131 --- /dev/null +++ b/artifacts/policies/franka/ppo_v2_targetbug_ARCHIVE/franka_M_s0/best.npz @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:50c5df0e6b0159f2395bbfffe5555320d9233f64c25d2ca255ec06cc6a5e427f +size 195152 diff --git a/artifacts/policies/franka/ppo_v2_targetbug_ARCHIVE/franka_M_s0/ep00050.npz b/artifacts/policies/franka/ppo_v2_targetbug_ARCHIVE/franka_M_s0/ep00050.npz new file mode 100644 index 0000000000000000000000000000000000000000..43c498928c6421ccefd0e0a0097235efd183ad62 --- /dev/null +++ b/artifacts/policies/franka/ppo_v2_targetbug_ARCHIVE/franka_M_s0/ep00050.npz @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:283af74d00df44d970be18593810958c429763438663728b9e831a2dd2f988b8 +size 195240 diff --git a/artifacts/policies/franka/ppo_v2_targetbug_ARCHIVE/franka_M_s0/ep00100.npz b/artifacts/policies/franka/ppo_v2_targetbug_ARCHIVE/franka_M_s0/ep00100.npz new file mode 100644 index 0000000000000000000000000000000000000000..088c035afa9ac4d23fdca78750b7ccdcfe3ab8f0 --- /dev/null +++ b/artifacts/policies/franka/ppo_v2_targetbug_ARCHIVE/franka_M_s0/ep00100.npz @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:52d39101f3579c3c39d5c1fe1d3cff9b800b08c7c3b80020ecb26dad281a0693 +size 195252 diff --git a/artifacts/policies/franka/ppo_v2_targetbug_ARCHIVE/franka_M_s0/ep00150.npz b/artifacts/policies/franka/ppo_v2_targetbug_ARCHIVE/franka_M_s0/ep00150.npz new file mode 100644 index 0000000000000000000000000000000000000000..05a29425e404e947d495f9c9bca339d74b258af1 --- /dev/null +++ b/artifacts/policies/franka/ppo_v2_targetbug_ARCHIVE/franka_M_s0/ep00150.npz @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:5377b77f92ddf031ef978fd677f7c4ee1e7cc2e34ba6baf5a1c9f9b20944651c +size 195252 diff --git a/artifacts/policies/franka/ppo_v2_targetbug_ARCHIVE/franka_M_s0/ep00200.npz b/artifacts/policies/franka/ppo_v2_targetbug_ARCHIVE/franka_M_s0/ep00200.npz new file mode 100644 index 0000000000000000000000000000000000000000..1397a36d070fb8130dae41e0e9287d8bf22c183c --- /dev/null +++ b/artifacts/policies/franka/ppo_v2_targetbug_ARCHIVE/franka_M_s0/ep00200.npz @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:2a80dfa8f2c4e3776a4cb9d48f94c84b828e46a7b26fed3ceb89ebb6cf9396a0 +size 195252 diff --git a/artifacts/policies/franka/ppo_v2_targetbug_ARCHIVE/franka_M_s0/ep00250.npz b/artifacts/policies/franka/ppo_v2_targetbug_ARCHIVE/franka_M_s0/ep00250.npz new file mode 100644 index 0000000000000000000000000000000000000000..3292ed6e57556f96db7eb6ae0e6120a46917ded9 --- /dev/null +++ b/artifacts/policies/franka/ppo_v2_targetbug_ARCHIVE/franka_M_s0/ep00250.npz @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:69c8d8ed9372c5db2c49c1d4b77eb7047634297834a2028330ea67a4963042fe +size 195256 diff --git a/artifacts/policies/franka/ppo_v2_targetbug_ARCHIVE/franka_M_s0/ep00300.npz b/artifacts/policies/franka/ppo_v2_targetbug_ARCHIVE/franka_M_s0/ep00300.npz new file mode 100644 index 0000000000000000000000000000000000000000..f4a0dd244567c428fd7e56aa62383dbbc4c3ea0a --- /dev/null +++ b/artifacts/policies/franka/ppo_v2_targetbug_ARCHIVE/franka_M_s0/ep00300.npz @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:8bf3b5dc31d3084f5aef5b6c121d6e2d6089608df95e198f79644d87caaece22 +size 195256 diff --git a/artifacts/policies/franka/ppo_v2_targetbug_ARCHIVE/franka_M_s0/ep00350.npz b/artifacts/policies/franka/ppo_v2_targetbug_ARCHIVE/franka_M_s0/ep00350.npz new file mode 100644 index 0000000000000000000000000000000000000000..23a64f6345effa994934c9cf6d830ee316e9f904 --- /dev/null +++ b/artifacts/policies/franka/ppo_v2_targetbug_ARCHIVE/franka_M_s0/ep00350.npz @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6b679b2366c71767c017380c1e95b87584dedfd429d26019618c08eafc7937cc +size 195248 diff --git a/artifacts/policies/franka/ppo_v2_targetbug_ARCHIVE/franka_M_s0/ep00400.npz b/artifacts/policies/franka/ppo_v2_targetbug_ARCHIVE/franka_M_s0/ep00400.npz new file mode 100644 index 0000000000000000000000000000000000000000..8710eef58eba807b3d531551c534d5d3a93c2cdc --- /dev/null +++ b/artifacts/policies/franka/ppo_v2_targetbug_ARCHIVE/franka_M_s0/ep00400.npz @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:78eb7e08015b1c9b598616aa6575bb18a19d444e0d7d58afc1f0a3d2769d3efd +size 195256 diff --git a/artifacts/policies/franka/ppo_v2_targetbug_ARCHIVE/franka_M_s0/ep00450.npz b/artifacts/policies/franka/ppo_v2_targetbug_ARCHIVE/franka_M_s0/ep00450.npz new file mode 100644 index 0000000000000000000000000000000000000000..28403c1e9a8786ee976e5f3d69c9072594cf6605 --- /dev/null +++ b/artifacts/policies/franka/ppo_v2_targetbug_ARCHIVE/franka_M_s0/ep00450.npz @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:195ea47d34501e937a33f63e4859b950dd52aa5ffff175ef9f41de3d5782d2ae +size 195252 diff --git a/artifacts/policies/franka/ppo_v2_targetbug_ARCHIVE/franka_M_s0/ep00500.npz b/artifacts/policies/franka/ppo_v2_targetbug_ARCHIVE/franka_M_s0/ep00500.npz new file mode 100644 index 0000000000000000000000000000000000000000..4fc60ae161fa13050936deec0bc398db546ef022 --- /dev/null +++ b/artifacts/policies/franka/ppo_v2_targetbug_ARCHIVE/franka_M_s0/ep00500.npz @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:40c74685cbf4030a0327ec71ca15a5e1cd5a91d86045d035143aec56a990ff49 +size 195256 diff --git a/artifacts/policies/franka/ppo_v2_targetbug_ARCHIVE/franka_M_s0/final.npz b/artifacts/policies/franka/ppo_v2_targetbug_ARCHIVE/franka_M_s0/final.npz new file mode 100644 index 0000000000000000000000000000000000000000..4fc60ae161fa13050936deec0bc398db546ef022 --- /dev/null +++ b/artifacts/policies/franka/ppo_v2_targetbug_ARCHIVE/franka_M_s0/final.npz @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:40c74685cbf4030a0327ec71ca15a5e1cd5a91d86045d035143aec56a990ff49 +size 195256 diff --git a/artifacts/policies/franka/ppo_v2_targetbug_ARCHIVE/franka_XS_s0/best.npz b/artifacts/policies/franka/ppo_v2_targetbug_ARCHIVE/franka_XS_s0/best.npz new file mode 100644 index 0000000000000000000000000000000000000000..b92cba406bdf07ea03e284fc1411407f81149638 --- /dev/null +++ b/artifacts/policies/franka/ppo_v2_targetbug_ARCHIVE/franka_XS_s0/best.npz @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:2279b71d78e65cbe98489e24ae00c395f6e0f35d9174532353c78ce6be71c6d5 +size 28284 diff --git a/artifacts/policies/franka/ppo_v2_targetbug_ARCHIVE/franka_XS_s0/ep00050.npz b/artifacts/policies/franka/ppo_v2_targetbug_ARCHIVE/franka_XS_s0/ep00050.npz new file mode 100644 index 0000000000000000000000000000000000000000..45ee035f715b9071b6dc98a8ac8bdc10917976ef --- /dev/null +++ b/artifacts/policies/franka/ppo_v2_targetbug_ARCHIVE/franka_XS_s0/ep00050.npz @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:0f8c924994d9a9b3b7ec1b100ce16cdc015e3ada35591bfb9394266216c58d21 +size 28376 diff --git a/artifacts/policies/franka/ppo_v2_targetbug_ARCHIVE/franka_XS_s0/ep00100.npz b/artifacts/policies/franka/ppo_v2_targetbug_ARCHIVE/franka_XS_s0/ep00100.npz new file mode 100644 index 0000000000000000000000000000000000000000..2c541a3804d22740c65aa8b93ff8fe6c52dfbc18 --- /dev/null +++ b/artifacts/policies/franka/ppo_v2_targetbug_ARCHIVE/franka_XS_s0/ep00100.npz @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:cd199b60d281d058880843d7ccb199c681a842c80d07737cc47415af9cfeb217 +size 28384 diff --git a/artifacts/policies/franka/ppo_v2_targetbug_ARCHIVE/franka_XS_s0/ep00150.npz b/artifacts/policies/franka/ppo_v2_targetbug_ARCHIVE/franka_XS_s0/ep00150.npz new file mode 100644 index 0000000000000000000000000000000000000000..eac77fc308d4c607795327daa1b20b68a01cbb82 --- /dev/null +++ b/artifacts/policies/franka/ppo_v2_targetbug_ARCHIVE/franka_XS_s0/ep00150.npz @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:946c89acf589d6618d675c87ed6d89d32565ad8997841086849dad94eb241b49 +size 28384 diff --git a/artifacts/policies/franka/ppo_v2_targetbug_ARCHIVE/franka_XS_s0/ep00200.npz b/artifacts/policies/franka/ppo_v2_targetbug_ARCHIVE/franka_XS_s0/ep00200.npz new file mode 100644 index 0000000000000000000000000000000000000000..9f352e9977ce9e59b89a0fd3919e0866dfb1581c --- /dev/null +++ b/artifacts/policies/franka/ppo_v2_targetbug_ARCHIVE/franka_XS_s0/ep00200.npz @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b65567dfbd5db6cf07ecc5d041f67d35aa46d972761bd8ccce3543e28b8286fe +size 28384 diff --git a/artifacts/policies/franka/ppo_v2_targetbug_ARCHIVE/franka_XS_s0/ep00250.npz b/artifacts/policies/franka/ppo_v2_targetbug_ARCHIVE/franka_XS_s0/ep00250.npz new file mode 100644 index 0000000000000000000000000000000000000000..1d9e300594eacf3f4897dd9929e7af62d48b70d5 --- /dev/null +++ b/artifacts/policies/franka/ppo_v2_targetbug_ARCHIVE/franka_XS_s0/ep00250.npz @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:bde038e64a7e854b2e041be42cbdcc06b0773c2663157aa341c3e30d636e0e0c +size 28388 diff --git a/artifacts/policies/franka/ppo_v2_targetbug_ARCHIVE/franka_XS_s0/ep00300.npz b/artifacts/policies/franka/ppo_v2_targetbug_ARCHIVE/franka_XS_s0/ep00300.npz new file mode 100644 index 0000000000000000000000000000000000000000..7b1e64e125edc2d921d41b4e508e40fd8004cd7d --- /dev/null +++ b/artifacts/policies/franka/ppo_v2_targetbug_ARCHIVE/franka_XS_s0/ep00300.npz @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:5fb5a17861da792127db6636fd11d2636280c9a1a501f179e0bfadb5ae39c7f4 +size 28388 diff --git a/artifacts/policies/franka/ppo_v2_targetbug_ARCHIVE/franka_XS_s0/ep00350.npz b/artifacts/policies/franka/ppo_v2_targetbug_ARCHIVE/franka_XS_s0/ep00350.npz new file mode 100644 index 0000000000000000000000000000000000000000..3ad6bed2fb1bf5aad674cd28c924b944f7993eb6 --- /dev/null +++ b/artifacts/policies/franka/ppo_v2_targetbug_ARCHIVE/franka_XS_s0/ep00350.npz @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a64d50f98fc61b22729060d7c182582d7eebe5a1e0f63b12ff7d2641d71ff6da +size 28384 diff --git a/artifacts/policies/franka/ppo_v2_targetbug_ARCHIVE/franka_XS_s0/ep00400.npz b/artifacts/policies/franka/ppo_v2_targetbug_ARCHIVE/franka_XS_s0/ep00400.npz new file mode 100644 index 0000000000000000000000000000000000000000..89c216e9c3f06c9df2f7fc93365f5c2596b858d7 --- /dev/null +++ b/artifacts/policies/franka/ppo_v2_targetbug_ARCHIVE/franka_XS_s0/ep00400.npz @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d1adff6f890e68fa017804998635830fb00ece375630305942eacad6a300aed9 +size 28388 diff --git a/artifacts/policies/franka/ppo_v2_targetbug_ARCHIVE/franka_XS_s0/ep00450.npz b/artifacts/policies/franka/ppo_v2_targetbug_ARCHIVE/franka_XS_s0/ep00450.npz new file mode 100644 index 0000000000000000000000000000000000000000..6469ca784912d66fde456cc30c3805366e5b83f5 --- /dev/null +++ b/artifacts/policies/franka/ppo_v2_targetbug_ARCHIVE/franka_XS_s0/ep00450.npz @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:88fea983697c32488e5481bdcd5f8e50aad801591a3a393c500015abc1cc4741 +size 28384 diff --git a/artifacts/policies/franka/ppo_v2_targetbug_ARCHIVE/franka_XS_s0/ep00500.npz b/artifacts/policies/franka/ppo_v2_targetbug_ARCHIVE/franka_XS_s0/ep00500.npz new file mode 100644 index 0000000000000000000000000000000000000000..4949cdb35f154db152e35e13d9ed069ad4b333d6 --- /dev/null +++ b/artifacts/policies/franka/ppo_v2_targetbug_ARCHIVE/franka_XS_s0/ep00500.npz @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6ec50033204675878e1dc901f8f0b844123b2f8978a4d81a535fac0ef362ff2d +size 28388 diff --git a/artifacts/policies/franka/ppo_v2_targetbug_ARCHIVE/franka_XS_s0/ep00550.npz b/artifacts/policies/franka/ppo_v2_targetbug_ARCHIVE/franka_XS_s0/ep00550.npz new file mode 100644 index 0000000000000000000000000000000000000000..5eadc76e4b943aaeb249f2ad3a527b8d816ac962 --- /dev/null +++ b/artifacts/policies/franka/ppo_v2_targetbug_ARCHIVE/franka_XS_s0/ep00550.npz @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:cb8a61937dee641c396aca7863ac6bf14c0992b0c85c0e50c21da1e1e521f202 +size 28384 diff --git a/artifacts/policies/franka/ppo_v2_targetbug_ARCHIVE/franka_XS_s0/final.npz b/artifacts/policies/franka/ppo_v2_targetbug_ARCHIVE/franka_XS_s0/final.npz new file mode 100644 index 0000000000000000000000000000000000000000..5eadc76e4b943aaeb249f2ad3a527b8d816ac962 --- /dev/null +++ b/artifacts/policies/franka/ppo_v2_targetbug_ARCHIVE/franka_XS_s0/final.npz @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:cb8a61937dee641c396aca7863ac6bf14c0992b0c85c0e50c21da1e1e521f202 +size 28384 diff --git a/artifacts/policies/franka/repair/M_s0/damaged/dead_units_0_0p5.npz b/artifacts/policies/franka/repair/M_s0/damaged/dead_units_0_0p5.npz new file mode 100644 index 0000000000000000000000000000000000000000..23d9c37956d1deb629020a0790ba52a4bb542bbc --- /dev/null +++ b/artifacts/policies/franka/repair/M_s0/damaged/dead_units_0_0p5.npz @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f6ba89fc05b6b3b3600841b8a7ff6b238d952b32088265bd3e78809b1bb75f5f +size 198192 diff --git a/artifacts/policies/franka/repair/M_s0/damaged/dead_units_1_0p5.npz b/artifacts/policies/franka/repair/M_s0/damaged/dead_units_1_0p5.npz new file mode 100644 index 0000000000000000000000000000000000000000..016333d42be87b9ea77b68abec9606bcf34dedd4 --- /dev/null +++ b/artifacts/policies/franka/repair/M_s0/damaged/dead_units_1_0p5.npz @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d578e0fb9e194a23a62b04842ba85b5cf7c09fdbd4ad6cf0f0e005d2f3124253 +size 196904 diff --git a/artifacts/policies/franka/repair/M_s0/damaged/flip_obs_18.npz b/artifacts/policies/franka/repair/M_s0/damaged/flip_obs_18.npz new file mode 100644 index 0000000000000000000000000000000000000000..caa1c71f19906dba1f5bfbbcf131ad8488c87bf2 --- /dev/null +++ b/artifacts/policies/franka/repair/M_s0/damaged/flip_obs_18.npz @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:1313b921cc00dcd2c965f7f0831a0576e959f2a5e9c2dc54ec0cdc337b2107f2 +size 195788 diff --git a/artifacts/policies/franka/repair/M_s0/damaged/flip_obs_20.npz b/artifacts/policies/franka/repair/M_s0/damaged/flip_obs_20.npz new file mode 100644 index 0000000000000000000000000000000000000000..22b08c8b41428cea84a9ae5ef8783f45a48e7346 --- /dev/null +++ b/artifacts/policies/franka/repair/M_s0/damaged/flip_obs_20.npz @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:aa34f3f3b8e9188c6de943ab7496048fc773dd2c06fd959efc478f4b29f516d9 +size 195788 diff --git a/artifacts/policies/franka/repair/M_s0/damaged/noise_0p5.npz b/artifacts/policies/franka/repair/M_s0/damaged/noise_0p5.npz new file mode 100644 index 0000000000000000000000000000000000000000..391c0d98b748da70c0a9f8f6c50a0fe279e6b94f --- /dev/null +++ b/artifacts/policies/franka/repair/M_s0/damaged/noise_0p5.npz @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c8b68164075b6d68bdcaf745f6d9c5ba10e5af52d7abf2e0306ce469f15c2060 +size 195800 diff --git a/artifacts/policies/franka/repair/M_s0/damaged/noise_1p0.npz b/artifacts/policies/franka/repair/M_s0/damaged/noise_1p0.npz new file mode 100644 index 0000000000000000000000000000000000000000..824e2b8df3b943a45c6cd5bc39431e26dfd09656 --- /dev/null +++ b/artifacts/policies/franka/repair/M_s0/damaged/noise_1p0.npz @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:29c0e415867e0350ddcfefd0e45f06e404f1a2f5435472faaf7675868eb6a74b +size 195800 diff --git a/artifacts/policies/franka/repair/M_s0/damaged/norm_shift_20.npz b/artifacts/policies/franka/repair/M_s0/damaged/norm_shift_20.npz new file mode 100644 index 0000000000000000000000000000000000000000..26316de2a32327268939cceb7af90456b399a74e --- /dev/null +++ b/artifacts/policies/franka/repair/M_s0/damaged/norm_shift_20.npz @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:49c1e6e23a5a244bb94f577780df25025c4bec191d5663f5fd8fc2f54be8a366 +size 195804 diff --git a/artifacts/policies/franka/repair/M_s0/damaged/norm_shift_21.npz b/artifacts/policies/franka/repair/M_s0/damaged/norm_shift_21.npz new file mode 100644 index 0000000000000000000000000000000000000000..2d5dd6c6ce21afb0c21ad6fbbcd475e063f6f244 --- /dev/null +++ b/artifacts/policies/franka/repair/M_s0/damaged/norm_shift_21.npz @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:bf18be659506e3c5bba72be2f4f7abbe64fd24c7bc557010388d55278723317a +size 195804 diff --git a/artifacts/policies/franka/repair/M_s0/damaged/swap_act_0_6.npz b/artifacts/policies/franka/repair/M_s0/damaged/swap_act_0_6.npz new file mode 100644 index 0000000000000000000000000000000000000000..6e9040e5a40555ab933cf0306ca28f9ff47816e5 --- /dev/null +++ b/artifacts/policies/franka/repair/M_s0/damaged/swap_act_0_6.npz @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:005a6fc046b36a5413ecd6d01c36d2c17466ccad7f404842197dc191e01328fc +size 195872 diff --git a/artifacts/policies/franka/repair/M_s0/damaged/swap_act_1_3.npz b/artifacts/policies/franka/repair/M_s0/damaged/swap_act_1_3.npz new file mode 100644 index 0000000000000000000000000000000000000000..f1916851cd01d9789c234e42319e6c5987e8aadb --- /dev/null +++ b/artifacts/policies/franka/repair/M_s0/damaged/swap_act_1_3.npz @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:647d82de6330edcc91cf601942d2532fa27791bc928561faf87794c63d90d75a +size 195872 diff --git a/artifacts/policies/franka/repair/M_s0/expert.npz b/artifacts/policies/franka/repair/M_s0/expert.npz new file mode 100644 index 0000000000000000000000000000000000000000..a6faf4637bf304c0c2cee21f4502f6d464d1ed47 --- /dev/null +++ b/artifacts/policies/franka/repair/M_s0/expert.npz @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:478eb1eec676b8038ebeaea8b567f22904f56cc792dfb18b80142bdcff814e3d +size 195268 diff --git a/artifacts/policies/franka/repair/M_s0/natural/collapse_ep500.npz b/artifacts/policies/franka/repair/M_s0/natural/collapse_ep500.npz new file mode 100644 index 0000000000000000000000000000000000000000..b5325d4882fcd5c0d6634672fb89bd6df9edfac8 --- /dev/null +++ b/artifacts/policies/franka/repair/M_s0/natural/collapse_ep500.npz @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:5e19af9cf2a1452f826c4d525f4cda60b1e5e344a460131d91f9dafcaabeab21 +size 195252 diff --git a/artifacts/policies/franka/repair/M_s0/natural/early_ep15.npz b/artifacts/policies/franka/repair/M_s0/natural/early_ep15.npz new file mode 100644 index 0000000000000000000000000000000000000000..60432f8e33622fcfba2083f5ff1ff4d13890a5b3 --- /dev/null +++ b/artifacts/policies/franka/repair/M_s0/natural/early_ep15.npz @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:947d0248e901b216aa337bfffb81c609727eff2c8f8f7dee315b0ac658eba353 +size 195284 diff --git a/artifacts/policies/franka/repair/M_s0/natural/early_ep20.npz b/artifacts/policies/franka/repair/M_s0/natural/early_ep20.npz new file mode 100644 index 0000000000000000000000000000000000000000..f06469539c4b305448703f0c3c6c75b4b9a6cff4 --- /dev/null +++ b/artifacts/policies/franka/repair/M_s0/natural/early_ep20.npz @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7714181a1cfef781267cd42ae9df7cc6b6a3104d642e7b5d5629a35938ff898f +size 195292 diff --git a/artifacts/policies/franka/repair/M_s0/natural/early_ep40.npz b/artifacts/policies/franka/repair/M_s0/natural/early_ep40.npz new file mode 100644 index 0000000000000000000000000000000000000000..9e72a08b274826603e351dd84c580b26bb01f0b8 --- /dev/null +++ b/artifacts/policies/franka/repair/M_s0/natural/early_ep40.npz @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:71b11e761d8b3bd89caeeae97a714c8ad91fe4b6a8d8954d21f1d91650f53a1d +size 195292 diff --git a/artifacts/policies/shadow/nvidia/pretrained.npz b/artifacts/policies/shadow/nvidia/pretrained.npz new file mode 100644 index 0000000000000000000000000000000000000000..499c97327d872674c5c280ae3835f61cb37ee2d5 --- /dev/null +++ b/artifacts/policies/shadow/nvidia/pretrained.npz @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:37805e1bea85477edacb64f73e0e1ab6f8bb0fb2374353b1ac95a5a01e999dd2 +size 2046724 diff --git a/artifacts/policies/shadow/ppo/shadow_L_s0_early/best.npz b/artifacts/policies/shadow/ppo/shadow_L_s0_early/best.npz new file mode 100644 index 0000000000000000000000000000000000000000..83b060ed04fc67a7dadf16afc71ddccc5846bce1 --- /dev/null +++ b/artifacts/policies/shadow/ppo/shadow_L_s0_early/best.npz @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:8c6864ec6990f6b904b2961067054229526cbb5f90f3644632a7ea5c392d53c5 +size 2046572 diff --git a/artifacts/policies/shadow/ppo/shadow_L_s0_early/ep00005.npz b/artifacts/policies/shadow/ppo/shadow_L_s0_early/ep00005.npz new file mode 100644 index 0000000000000000000000000000000000000000..5357e5f3445b362adc1a06a92eb5d58285fe9bf0 --- /dev/null +++ b/artifacts/policies/shadow/ppo/shadow_L_s0_early/ep00005.npz @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:1a1a66f496c4459437441661c47aca2fffb5db81a448042274bf6fc993bfc43c +size 2046652 diff --git a/artifacts/policies/shadow/ppo/shadow_L_s0_early/ep00010.npz b/artifacts/policies/shadow/ppo/shadow_L_s0_early/ep00010.npz new file mode 100644 index 0000000000000000000000000000000000000000..9b75282abd15bfe4ab143035d3a34b945cd73175 --- /dev/null +++ b/artifacts/policies/shadow/ppo/shadow_L_s0_early/ep00010.npz @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:39e725f26f0d4048070e4370372a3f636f8b0bd867f7573776ec684138a1dc9c +size 2046668 diff --git a/artifacts/policies/shadow/ppo/shadow_L_s0_early/ep00015.npz b/artifacts/policies/shadow/ppo/shadow_L_s0_early/ep00015.npz new file mode 100644 index 0000000000000000000000000000000000000000..9711b8f13eadf9646c18152981ce12c4032354b3 --- /dev/null +++ b/artifacts/policies/shadow/ppo/shadow_L_s0_early/ep00015.npz @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:af11e9a5bc080388ff743069ba3d67ed588c8a1904c8aded3f391d976ed51015 +size 2046668 diff --git a/artifacts/policies/shadow/ppo/shadow_L_s0_early/ep00020.npz b/artifacts/policies/shadow/ppo/shadow_L_s0_early/ep00020.npz new file mode 100644 index 0000000000000000000000000000000000000000..af4b68cb4a522902f68b28285de372962abe9954 --- /dev/null +++ b/artifacts/policies/shadow/ppo/shadow_L_s0_early/ep00020.npz @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d089a8267b1e51b6e9adff697da966860d3b46ac57cf8c63ebeb2b81b880326d +size 2046664 diff --git a/artifacts/policies/shadow/ppo/shadow_L_s0_early/ep00025.npz b/artifacts/policies/shadow/ppo/shadow_L_s0_early/ep00025.npz new file mode 100644 index 0000000000000000000000000000000000000000..928df2b333db53cf5e480daf1d53ce37be484332 --- /dev/null +++ b/artifacts/policies/shadow/ppo/shadow_L_s0_early/ep00025.npz @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e5b8eee8680cb11857f929e756b1f5a6fbf15babaaaad0930f80bd7183785e26 +size 2046668 diff --git a/artifacts/policies/shadow/ppo/shadow_L_s0_early/ep00030.npz b/artifacts/policies/shadow/ppo/shadow_L_s0_early/ep00030.npz new file mode 100644 index 0000000000000000000000000000000000000000..936de78d554eb1b4af9cd5bde62f6c206806aa9e --- /dev/null +++ b/artifacts/policies/shadow/ppo/shadow_L_s0_early/ep00030.npz @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:546638919411dbdb070317e6499ea11106a66198df74a50563a8f167c827b72a +size 2046668 diff --git a/artifacts/policies/shadow/ppo/shadow_L_s0_early/ep00035.npz b/artifacts/policies/shadow/ppo/shadow_L_s0_early/ep00035.npz new file mode 100644 index 0000000000000000000000000000000000000000..a6ae6467f9af0f3b62fbe900f478d1e50785c59e --- /dev/null +++ b/artifacts/policies/shadow/ppo/shadow_L_s0_early/ep00035.npz @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:cd8b78599e4ed8da5d7ff401f4eac9e33317d8a6d2b8e9cc29317a7f6bd08855 +size 2046668 diff --git a/artifacts/policies/shadow/ppo/shadow_L_s0_early/ep00040.npz b/artifacts/policies/shadow/ppo/shadow_L_s0_early/ep00040.npz new file mode 100644 index 0000000000000000000000000000000000000000..0b09a7868848ead94bc790f537eed1a4d26fa758 --- /dev/null +++ b/artifacts/policies/shadow/ppo/shadow_L_s0_early/ep00040.npz @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:022c98c12a6be87bce0c0f9a79234456e82ef7d134df586bd21bc07ab0f4dbbf +size 2046660 diff --git a/artifacts/policies/shadow/ppo/shadow_L_s0_early/ep00045.npz b/artifacts/policies/shadow/ppo/shadow_L_s0_early/ep00045.npz new file mode 100644 index 0000000000000000000000000000000000000000..7f42f525a0e542b8357d2f4d21c572915b06358b --- /dev/null +++ b/artifacts/policies/shadow/ppo/shadow_L_s0_early/ep00045.npz @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:78b8630b1786c40a3dd0d1235f759b6077c94569a30d9edcb1bc615ebff8116d +size 2046664 diff --git a/artifacts/policies/shadow/ppo/shadow_L_s0_early/ep00050.npz b/artifacts/policies/shadow/ppo/shadow_L_s0_early/ep00050.npz new file mode 100644 index 0000000000000000000000000000000000000000..b46ca865c75e96b20e242489be53e034a74fc7de --- /dev/null +++ b/artifacts/policies/shadow/ppo/shadow_L_s0_early/ep00050.npz @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d7748ce5cf91c36510b0f5ffc0510c623aa2f0d64a660f77998d7a222a8d7a07 +size 2046664 diff --git a/artifacts/policies/shadow/ppo/shadow_L_s0_early/ep00055.npz b/artifacts/policies/shadow/ppo/shadow_L_s0_early/ep00055.npz new file mode 100644 index 0000000000000000000000000000000000000000..0e5754f2504ce8d1313b65090e67f8b127cb5d56 --- /dev/null +++ b/artifacts/policies/shadow/ppo/shadow_L_s0_early/ep00055.npz @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b3ee97f4977d46d007b395015377572a713432843c79374af3074d6f2dfc7361 +size 2046656 diff --git a/artifacts/policies/shadow/ppo/shadow_L_s0_early/ep00060.npz b/artifacts/policies/shadow/ppo/shadow_L_s0_early/ep00060.npz new file mode 100644 index 0000000000000000000000000000000000000000..739e7752dee9385a4d034cd31135752b7a2401c4 --- /dev/null +++ b/artifacts/policies/shadow/ppo/shadow_L_s0_early/ep00060.npz @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6165d0788dc00fc0b4cfbdbe9887665545c776dd6c933902a9c95d01bd2412f6 +size 2046668 diff --git a/artifacts/policies/shadow/ppo/shadow_L_s0_early/ep00065.npz b/artifacts/policies/shadow/ppo/shadow_L_s0_early/ep00065.npz new file mode 100644 index 0000000000000000000000000000000000000000..f659e081cb6a2c242a399a98bf89c322902e1af7 --- /dev/null +++ b/artifacts/policies/shadow/ppo/shadow_L_s0_early/ep00065.npz @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d4a3d5b809008e5ed449a11586914b59885aa55a51cdc3f07cc0fa40eff615a4 +size 2046664 diff --git a/artifacts/policies/shadow/ppo/shadow_L_s0_early/ep00070.npz b/artifacts/policies/shadow/ppo/shadow_L_s0_early/ep00070.npz new file mode 100644 index 0000000000000000000000000000000000000000..bd62b03be1ba7fc92b4734dc5a4a3df084a65059 --- /dev/null +++ b/artifacts/policies/shadow/ppo/shadow_L_s0_early/ep00070.npz @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:35167dd5e9b332bd973a71dba20bb96f21455deeb0e7da44c6a21679e47a1872 +size 2046664 diff --git a/artifacts/policies/shadow/ppo/shadow_L_s0_early/ep00075.npz b/artifacts/policies/shadow/ppo/shadow_L_s0_early/ep00075.npz new file mode 100644 index 0000000000000000000000000000000000000000..bbe12f368da4068e86ffc99039b2c2b7158ebbe9 --- /dev/null +++ b/artifacts/policies/shadow/ppo/shadow_L_s0_early/ep00075.npz @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a7eef4a92c6dc72d8942726635c449c465e3574ed50689fc63b1216ae5596159 +size 2046656 diff --git a/artifacts/policies/shadow/ppo/shadow_L_s0_early/ep00080.npz b/artifacts/policies/shadow/ppo/shadow_L_s0_early/ep00080.npz new file mode 100644 index 0000000000000000000000000000000000000000..84936c375f5091bd93beb07c299e84e7c71f0bcd --- /dev/null +++ b/artifacts/policies/shadow/ppo/shadow_L_s0_early/ep00080.npz @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ad0f923de554f482d13ab57c8da949d2f8f558c90d2f52d0f208f549dc5a4d40 +size 2046668 diff --git a/artifacts/policies/shadow/ppo/shadow_L_s0_early/ep00085.npz b/artifacts/policies/shadow/ppo/shadow_L_s0_early/ep00085.npz new file mode 100644 index 0000000000000000000000000000000000000000..81fc475382aed18331b57e00775918770c9c5342 --- /dev/null +++ b/artifacts/policies/shadow/ppo/shadow_L_s0_early/ep00085.npz @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c41febe1c4be2c36995f27a75fd69ccf5fc4cbaa764e21a07105e5029a3aea68 +size 2046668 diff --git a/artifacts/policies/shadow/ppo/shadow_L_s0_early/ep00090.npz b/artifacts/policies/shadow/ppo/shadow_L_s0_early/ep00090.npz new file mode 100644 index 0000000000000000000000000000000000000000..aaf7152e8a114e398952aeee5fc24d492a87d9a4 --- /dev/null +++ b/artifacts/policies/shadow/ppo/shadow_L_s0_early/ep00090.npz @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e12397699c76a2d09736a288662bc93af4509cd8156828a96b765164d7b75ff4 +size 2046664 diff --git a/artifacts/policies/shadow/ppo/shadow_L_s0_early/ep00095.npz b/artifacts/policies/shadow/ppo/shadow_L_s0_early/ep00095.npz new file mode 100644 index 0000000000000000000000000000000000000000..661bf7750217015ae754ae7b107b478a3a9d8b61 --- /dev/null +++ b/artifacts/policies/shadow/ppo/shadow_L_s0_early/ep00095.npz @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:27111e0e67cf5f83fc3a6da5cda9852cb6a2327058bb5d765e48e7ce89389669 +size 2046668 diff --git a/artifacts/policies/shadow/ppo/shadow_L_s0_early/ep00100.npz b/artifacts/policies/shadow/ppo/shadow_L_s0_early/ep00100.npz new file mode 100644 index 0000000000000000000000000000000000000000..628bf00618394a015820f18ad018df907a9f7505 --- /dev/null +++ b/artifacts/policies/shadow/ppo/shadow_L_s0_early/ep00100.npz @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:fbc800488d39a927d40c525640b04ff50caee8c66753e4333c0a203239427536 +size 2046676 diff --git a/artifacts/policies/shadow/ppo/shadow_L_s0_early/ep00105.npz b/artifacts/policies/shadow/ppo/shadow_L_s0_early/ep00105.npz new file mode 100644 index 0000000000000000000000000000000000000000..78b60683f8b45b19f8bc7e48b0ee1d1490fab6d9 --- /dev/null +++ b/artifacts/policies/shadow/ppo/shadow_L_s0_early/ep00105.npz @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:31bba83b83672b78bec5a13c8c0952bf391c6132355d28fd16065fafbae7d7f8 +size 2046672 diff --git a/artifacts/policies/shadow/ppo/shadow_L_s0_early/ep00110.npz b/artifacts/policies/shadow/ppo/shadow_L_s0_early/ep00110.npz new file mode 100644 index 0000000000000000000000000000000000000000..d69c73d2528c24a8df0fd42c401fad0b26d0ca5d --- /dev/null +++ b/artifacts/policies/shadow/ppo/shadow_L_s0_early/ep00110.npz @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:9dd50a10dd96fd67d8dd727ba809dca79dc38ca3d61a20bf973586018306c6c1 +size 2046676 diff --git a/artifacts/policies/shadow/ppo/shadow_L_s0_early/ep00115.npz b/artifacts/policies/shadow/ppo/shadow_L_s0_early/ep00115.npz new file mode 100644 index 0000000000000000000000000000000000000000..9a6b7f43e44348081ae3aa2b54b2d15077bb76b4 --- /dev/null +++ b/artifacts/policies/shadow/ppo/shadow_L_s0_early/ep00115.npz @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:1f9b11f3d21249fb45f153bebbcc1a02f0c694c7024383edb8be1d6fd58c94a7 +size 2046672 diff --git a/artifacts/policies/shadow/ppo/shadow_L_s0_early/ep00120.npz b/artifacts/policies/shadow/ppo/shadow_L_s0_early/ep00120.npz new file mode 100644 index 0000000000000000000000000000000000000000..ef70fb57f4b8740a5a38619e4717a70042c1044c --- /dev/null +++ b/artifacts/policies/shadow/ppo/shadow_L_s0_early/ep00120.npz @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:5b03f735124f5bdeee63e4fb1ebdd672252ded24e39c16319a79dec4ddb6a807 +size 2046676 diff --git a/artifacts/policies/shadow/ppo/shadow_L_s0_early/ep00125.npz b/artifacts/policies/shadow/ppo/shadow_L_s0_early/ep00125.npz new file mode 100644 index 0000000000000000000000000000000000000000..934fd036620cceff3c284010ce39e4b48899e2dd --- /dev/null +++ b/artifacts/policies/shadow/ppo/shadow_L_s0_early/ep00125.npz @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:db56523f06ec462420bf85ef5451b6975297bbc9a4c912e8f3027a06df292d91 +size 2046664 diff --git a/artifacts/policies/shadow/ppo/shadow_L_s0_early/ep00130.npz b/artifacts/policies/shadow/ppo/shadow_L_s0_early/ep00130.npz new file mode 100644 index 0000000000000000000000000000000000000000..dcbd3f480e7f318095cba3594ac24bb8c0412211 --- /dev/null +++ b/artifacts/policies/shadow/ppo/shadow_L_s0_early/ep00130.npz @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ab4455b7f9878aa032f1c96ded5a0495c03463b326d66cc41bfacfef037c2ccf +size 2046676 diff --git a/artifacts/policies/shadow/ppo/shadow_L_s0_early/ep00135.npz b/artifacts/policies/shadow/ppo/shadow_L_s0_early/ep00135.npz new file mode 100644 index 0000000000000000000000000000000000000000..55ae3f0ce618e7ff5011919b7476c23542063476 --- /dev/null +++ b/artifacts/policies/shadow/ppo/shadow_L_s0_early/ep00135.npz @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:dab37e2814e17e6be9cd7e5c11aa9cd620885e2d06f43ea82e2bf3940089750f +size 2046676 diff --git a/artifacts/policies/shadow/ppo/shadow_L_s0_early/ep00140.npz b/artifacts/policies/shadow/ppo/shadow_L_s0_early/ep00140.npz new file mode 100644 index 0000000000000000000000000000000000000000..5e0d212ae9f9d54ccc86f5df7815bdf74bbd0160 --- /dev/null +++ b/artifacts/policies/shadow/ppo/shadow_L_s0_early/ep00140.npz @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d53ca805e1de68ae5712a3eeca40d922b68789ce9421114f9f861ad7e3fa3c24 +size 2046676 diff --git a/artifacts/policies/shadow/ppo/shadow_L_s0_early/ep00145.npz b/artifacts/policies/shadow/ppo/shadow_L_s0_early/ep00145.npz new file mode 100644 index 0000000000000000000000000000000000000000..483295f22764c996c552d2ad23e4112a0353b3f1 --- /dev/null +++ b/artifacts/policies/shadow/ppo/shadow_L_s0_early/ep00145.npz @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6e9309fdb47e363ec6b0aee534c26b7667928bcf908c0d23a53975e6631d55ee +size 2046672 diff --git a/artifacts/policies/shadow/ppo/shadow_L_s0_early/ep00150.npz b/artifacts/policies/shadow/ppo/shadow_L_s0_early/ep00150.npz new file mode 100644 index 0000000000000000000000000000000000000000..ea555adaeef7cc49751f35d1222ec041913ccf08 --- /dev/null +++ b/artifacts/policies/shadow/ppo/shadow_L_s0_early/ep00150.npz @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3e925e3e45f07c0d493e29911e45a9928ac25fd0c736709a932fc4a6d8658d07 +size 2046676 diff --git a/artifacts/policies/shadow/ppo/shadow_L_s0_early/ep00155.npz b/artifacts/policies/shadow/ppo/shadow_L_s0_early/ep00155.npz new file mode 100644 index 0000000000000000000000000000000000000000..cc511adddfde24e2f0a1f51e55a038aaba955543 --- /dev/null +++ b/artifacts/policies/shadow/ppo/shadow_L_s0_early/ep00155.npz @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4f35fb495f4b2317b23a4a4152868fee8be1fae5ba07886cf7cbfd4998cfd710 +size 2046672 diff --git a/artifacts/policies/shadow/ppo/shadow_L_s0_early/ep00160.npz b/artifacts/policies/shadow/ppo/shadow_L_s0_early/ep00160.npz new file mode 100644 index 0000000000000000000000000000000000000000..1951824e9398df95868ea1b781a4e32e8d17a3a6 --- /dev/null +++ b/artifacts/policies/shadow/ppo/shadow_L_s0_early/ep00160.npz @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:fee4e9fa16e821934f5a69ee41393110f6e50313281df7f046a3b726ea1fbb4c +size 2046676 diff --git a/artifacts/policies/shadow/ppo/shadow_L_s0_early/ep00165.npz b/artifacts/policies/shadow/ppo/shadow_L_s0_early/ep00165.npz new file mode 100644 index 0000000000000000000000000000000000000000..7fe0c54edc1132a688c76fa57e8d7210c116d0f3 --- /dev/null +++ b/artifacts/policies/shadow/ppo/shadow_L_s0_early/ep00165.npz @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:76ce6f74c3cfbc97157a5b94488cc70eac3bc5b6df53f30eae8d743a012589c0 +size 2046672 diff --git a/artifacts/policies/shadow/ppo/shadow_L_s0_early/ep00170.npz b/artifacts/policies/shadow/ppo/shadow_L_s0_early/ep00170.npz new file mode 100644 index 0000000000000000000000000000000000000000..f1ac445ef55c26c126d49e563b92b22c5c6890ed --- /dev/null +++ b/artifacts/policies/shadow/ppo/shadow_L_s0_early/ep00170.npz @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:1b2d81358781e29b16ede3ddb3716716036df123ee664398f8a67d5995ad9c49 +size 2046676 diff --git a/artifacts/policies/shadow/ppo/shadow_L_s0_early/ep00175.npz b/artifacts/policies/shadow/ppo/shadow_L_s0_early/ep00175.npz new file mode 100644 index 0000000000000000000000000000000000000000..6dc14b852a8757ac57238a622c9b16daf4cdc4c5 --- /dev/null +++ b/artifacts/policies/shadow/ppo/shadow_L_s0_early/ep00175.npz @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:30fd9fcd6e75b3ec7022aa96fb4a17a9e382a6b5bfbbb5ced0c5cd2efd2e9b82 +size 2046672 diff --git a/artifacts/policies/shadow/ppo/shadow_L_s0_early/ep00180.npz b/artifacts/policies/shadow/ppo/shadow_L_s0_early/ep00180.npz new file mode 100644 index 0000000000000000000000000000000000000000..a9fec654de6e2c35c366904b023b3f559b6955ed --- /dev/null +++ b/artifacts/policies/shadow/ppo/shadow_L_s0_early/ep00180.npz @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e704b08bed4e39ed1c92491e10a4c835f1fcdb47affc39f50a99d6abcad9ca9d +size 2046676 diff --git a/artifacts/policies/shadow/ppo/shadow_L_s0_early/ep00185.npz b/artifacts/policies/shadow/ppo/shadow_L_s0_early/ep00185.npz new file mode 100644 index 0000000000000000000000000000000000000000..ac020a705c87552fb626585ae051ca1d8e0f2dc9 --- /dev/null +++ b/artifacts/policies/shadow/ppo/shadow_L_s0_early/ep00185.npz @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c55849d56e120c9dfca932416c77bc449642948684e1ea4bb0040710a9135b6c +size 2046672 diff --git a/artifacts/policies/shadow/ppo/shadow_L_s0_early/ep00190.npz b/artifacts/policies/shadow/ppo/shadow_L_s0_early/ep00190.npz new file mode 100644 index 0000000000000000000000000000000000000000..63b762b9849698c7b158d97ae062e71ccf0e1460 --- /dev/null +++ b/artifacts/policies/shadow/ppo/shadow_L_s0_early/ep00190.npz @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:0d2a130a7e4ec49ad3a66dfca398551b85e986123b83067a6ede8e141aa18cba +size 2046676 diff --git a/artifacts/policies/shadow/ppo/shadow_L_s0_early/ep00195.npz b/artifacts/policies/shadow/ppo/shadow_L_s0_early/ep00195.npz new file mode 100644 index 0000000000000000000000000000000000000000..23915c55527ea18287c9c979c1dcff52915408a4 --- /dev/null +++ b/artifacts/policies/shadow/ppo/shadow_L_s0_early/ep00195.npz @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e29816378164c81a669428c8fc65e9007c39b4dcb9dabd350e93b5304f8c0493 +size 2046668 diff --git a/artifacts/policies/shadow/ppo/shadow_L_s0_early/ep00200.npz b/artifacts/policies/shadow/ppo/shadow_L_s0_early/ep00200.npz new file mode 100644 index 0000000000000000000000000000000000000000..e95a09a75c7279f9e17c9dcd47a2184128fdae7b --- /dev/null +++ b/artifacts/policies/shadow/ppo/shadow_L_s0_early/ep00200.npz @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:bc3fcc10d5dfeb49a5c476316ca9c59b199427dca99da479c18b36d4a4c9636c +size 2046688 diff --git a/artifacts/policies/shadow/ppo/shadow_L_s0_early/final.npz b/artifacts/policies/shadow/ppo/shadow_L_s0_early/final.npz new file mode 100644 index 0000000000000000000000000000000000000000..e95a09a75c7279f9e17c9dcd47a2184128fdae7b --- /dev/null +++ b/artifacts/policies/shadow/ppo/shadow_L_s0_early/final.npz @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:bc3fcc10d5dfeb49a5c476316ca9c59b199427dca99da479c18b36d4a4c9636c +size 2046688 diff --git a/artifacts/policies/shadow/repair/L/damaged/norm_shift_69.npz b/artifacts/policies/shadow/repair/L/damaged/norm_shift_69.npz new file mode 100644 index 0000000000000000000000000000000000000000..2efbfaa002e2d239871229ba08e6f00c93bb1a4d --- /dev/null +++ b/artifacts/policies/shadow/repair/L/damaged/norm_shift_69.npz @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:74f84dc98ea89893fe05bbe396e4409b01d29d5af1f6b6eb5767c33703596789 +size 2047180 diff --git a/artifacts/policies/shadow/repair/L/damaged/swap_act_7_11.npz b/artifacts/policies/shadow/repair/L/damaged/swap_act_7_11.npz new file mode 100644 index 0000000000000000000000000000000000000000..412ace4a294f7474826a965ae1992ad04a8d24cf --- /dev/null +++ b/artifacts/policies/shadow/repair/L/damaged/swap_act_7_11.npz @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:35394fc29fce6b150bd33bf18d3940a305bbf83a4dd42fede1622dd20eec349e +size 2047260 diff --git a/artifacts/policies/shadow/repair/L/natural/early_ep100.npz b/artifacts/policies/shadow/repair/L/natural/early_ep100.npz new file mode 100644 index 0000000000000000000000000000000000000000..628bf00618394a015820f18ad018df907a9f7505 --- /dev/null +++ b/artifacts/policies/shadow/repair/L/natural/early_ep100.npz @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:fbc800488d39a927d40c525640b04ff50caee8c66753e4333c0a203239427536 +size 2046676 diff --git a/artifacts/policies/shadow/repair/L/natural/early_ep200.npz b/artifacts/policies/shadow/repair/L/natural/early_ep200.npz new file mode 100644 index 0000000000000000000000000000000000000000..e95a09a75c7279f9e17c9dcd47a2184128fdae7b --- /dev/null +++ b/artifacts/policies/shadow/repair/L/natural/early_ep200.npz @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:bc3fcc10d5dfeb49a5c476316ca9c59b199427dca99da479c18b36d4a4c9636c +size 2046688 diff --git a/artifacts/reports/figures/capability_map.pdf b/artifacts/reports/figures/capability_map.pdf new file mode 100644 index 0000000000000000000000000000000000000000..5a7165cdfc66cb70ba80c4079afb01d4d49c4970 Binary files /dev/null and b/artifacts/reports/figures/capability_map.pdf differ diff --git a/artifacts/reports/figures/capability_map.png b/artifacts/reports/figures/capability_map.png new file mode 100644 index 0000000000000000000000000000000000000000..e229524f6061f3d50f938a6096fc9c31470f29c5 --- /dev/null +++ b/artifacts/reports/figures/capability_map.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:63bfa02ad6540c885de39073cd42e8bd3b85a995adb652345cf146f92a5273df +size 281914 diff --git a/artifacts/reports/figures/capability_map.svg b/artifacts/reports/figures/capability_map.svg new file mode 100644 index 0000000000000000000000000000000000000000..6080262eecc6e4941dcbe0d88b074ad25c999634 --- /dev/null +++ b/artifacts/reports/figures/capability_map.svg @@ -0,0 +1,1057 @@ + + + + + + + + 2026-11-04T19:21:06.625931 + image/svg+xml + + + Matplotlib v3.9.4, https://matplotlib.org/ + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + −.25 + + + + + + + + + + + + + 0 + + + + + + + + + + + + + .25 + + + + + + + + + + + + + .50 + + + + + + + + + + + + + .75 + + + + + + + + + + + + + 1 + + + + Progress from starting score to expert + + + + + + + + + + + + + + Franka + drawer success + + + + + + + + + + Allegro + goals / episode + + + + + + + + + + Shadow + goals / episode + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + n=10 + + + + + + + + + + + + + + + + + + + + + + + + n=3 + + + + + + + + + + + + + + + + + + n=3 + + + Original damaged runs are faded: fault names were visible. + Only completed blinded reruns are diamonds. + + + A Repair deliberate damage + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + −.25 + + + + + + + + + + + + + 0 + + + + + + + + + + + + + .25 + + + + + + + + + + + + + .50 + + + + + + + + + + + + + .75 + + + + + + + + + + + + + 1 + + + + Progress from starting score to expert + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + n=7 + + + + + + + + + + + + + + + + + + + + + + + + n=4 + + + + + + + + + + + + + + + + + + + + + n=4 + + + Hollow = controller replacement within the edit budget. + An entry-change cap does not guarantee preserving behaviour. + + + B Repair natural failure + + + + + + + + + + + + + + + + + + + + + −.25 + + + + + + + + + + + + + 0 + + + + + + + + + + + + + .25 + + + + + + + + + + + + + .50 + + + + + + + + + + + + + .75 + + + + + + + + + + + + + 1 + + + + Progress from starting score to expert + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + n=2 + + + + + + + + + + + + + + + + + + + + + + + n=4 + + + + + + + + + + + + + + + + + + + + + + + n=4 + + + Franka: Opus reaches 99.85%; Fable was not tested here. + Hand scratch policies remain below 0.5% of expert score. + + + C Write from scratch + + + + Direct weight programming works in selected settings; general repair is unproven + + + 41 scored attempts · 6/10 blinded reruns scored · 0 = no improvement, 1 = expert score · one attempt per cell + + + Progress = (after − before) / (expert − before). Scratch writing uses a zero-score reference; no zero-policy baseline was measured. + Scores: 4,096 held-out environments per policy. Repeated episodes do not establish reliability across independent LLM attempts. + + + + + + + + + + + + Opus 5.5 + + + + + + + + + + + Fable 5.1 + + + + + + + + + + + Blinded + + + + + + + + + + + Replacement + + + + + + + + + + + Audit unavailable + + + + + + + + + + + + + + + diff --git a/artifacts/reports/figures/openvla_experiment.pdf b/artifacts/reports/figures/openvla_experiment.pdf new file mode 100644 index 0000000000000000000000000000000000000000..637c76b7edfca1a85c5aacbc6f8de51ab30f4444 Binary files /dev/null and b/artifacts/reports/figures/openvla_experiment.pdf differ diff --git a/artifacts/reports/figures/openvla_experiment.png b/artifacts/reports/figures/openvla_experiment.png new file mode 100644 index 0000000000000000000000000000000000000000..3378f589dd530e462c8da897896d5151ac94a44e --- /dev/null +++ b/artifacts/reports/figures/openvla_experiment.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:229da8b1bb057acc6598f9d73b5b2319f167885aa3ef668457728e4daf831d01 +size 590654 diff --git a/artifacts/reports/figures/openvla_experiment.svg b/artifacts/reports/figures/openvla_experiment.svg new file mode 100644 index 0000000000000000000000000000000000000000..eb53e37972ed27652d7c3f73b703cc933d28d758 --- /dev/null +++ b/artifacts/reports/figures/openvla_experiment.svg @@ -0,0 +1,705 @@ + + + + + + + + 2026-11-04T19:05:44.754329 + image/svg+xml + + + Matplotlib v3.9.4, https://matplotlib.org/ + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + 0% + + + + + + + + + 50% + + + + + + + + + 100% + + + + + + + + Baseline + + + + + + Opus + + + + + + Fable + + + + + + + + + + + + + + 27/30 + + + 27/30 + + + 24/30 + + + + + + + + + + + + + + 0% + + + + + + + + + 50% + + + + + + + + + 100% + + + + + + + + Baseline + + + + + + Opus + + + + + + Fable + + + + + + + + + + + + + + 27/30 + + + 27/30 + + + 26/30 + + + + Can Claude improve OpenVLA by editing its weights? + + + LIBERO-Spatial / Six tasks / Five final rollouts per task / Four independent editing attempts + + + ONE FINE-TUNED OPENVLA CHECKPOINT + + + Unmodified baseline: 27/30 successes (90%) • Every attempt starts from these exact weights + + + A Single-task editing + + + Move the bowl from the drawer to the plate + + + Original checkpoint + + + Opus 5.5 + + + 4,023 weights changed + + + Fable 5.1 + + + 4,096 weights changed + + + Selected-task results + + + Successes in five final rollouts per task + + + Baseline + + + Opus + + + Fable + + + Inside top drawer + + + 3/5 + + + 3/5 + + + 3/5 + + + Target success: 60% for all three policies + + + Final evaluation on all six tasks + + + B Three-task editing + + + Move the bowl to the plate from three starting locations + + + Original checkpoint + + + Opus 5.5 + + + 0 weights changed + + + Fable 5.1 + + + 4,089 weights changed + + + Selected-task results + + + Successes in five final rollouts per task + + + Baseline + + + Opus + + + Fable + + + Inside top drawer + + + 3/5 + + + 3/5 + + + 2/5 + + + On ramekin + + + 4/5 + + + 4/5 + + + 5/5 + + + Between plate & ramekin + + + 5/5 + + + 5/5 + + + 5/5 + + + Combined target success: 12/15 (80%) for all three + + + Final evaluation on all six tasks + + + Design: editors see only their selected tasks; frozen policies are evaluated on all six tasks. No final feedback returns to editors. + + + Result: aggregate target success did not improve. Fable’s edits reduced overall success; Opus’s three-task submission made no changes. + + + Pilot limits: one attempt per condition and five episodes per task. Three attempts resumed after sandbox memory failures. + + + Fable single-task had no valid development feedback because of a renderer failure. The results do not establish a model ranking. + + + + + + + + + + + diff --git a/artifacts/reports/figures/scratch_vs_ppo.pdf b/artifacts/reports/figures/scratch_vs_ppo.pdf new file mode 100644 index 0000000000000000000000000000000000000000..fdf6638d961b28d5ebe0b8d9075a10dfb3b86421 Binary files /dev/null and b/artifacts/reports/figures/scratch_vs_ppo.pdf differ diff --git a/artifacts/reports/figures/scratch_vs_ppo.png b/artifacts/reports/figures/scratch_vs_ppo.png new file mode 100644 index 0000000000000000000000000000000000000000..e554bc2bd951c23f97cdb590506dd4edd06699c4 --- /dev/null +++ b/artifacts/reports/figures/scratch_vs_ppo.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:5da764bffad4f95ca052c88d603a488be42510fa742c8967c5405feb71f53d6a +size 229268 diff --git a/artifacts/reports/figures/scratch_vs_ppo.svg b/artifacts/reports/figures/scratch_vs_ppo.svg new file mode 100644 index 0000000000000000000000000000000000000000..0dfbd108df6334e9dded58cd8b9533c0b6b1ed38 --- /dev/null +++ b/artifacts/reports/figures/scratch_vs_ppo.svg @@ -0,0 +1,1490 @@ + + + + + + + + 2026-11-04T19:21:08.048270 + image/svg+xml + + + Matplotlib v3.9.4, https://matplotlib.org/ + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + No sim + + + + + + + + + + + + + 100k + + + + + + + + + + + + + 1M + + + + + + + + + + + + + 10M + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + Practice transition upper bound / PPO frames + + + + + + + + + + + + + + + + + 0% + + + + + + + + + + + + + 20% + + + + + + + + + + + + + 40% + + + + + + + + + + + + + 60% + + + + + + + + + + + + + 80% + + + + + + + + + + + + + 100% + + + + Held-out drawer success + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + Franka + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + No sim + + + + + + + + + + + + + 100k + + + + + + + + + + + + + 1M + + + + + + + + + + + + + 10M + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + Practice transition upper bound / PPO frames + + + + + + + + + + + + + + + + + 0 + + + + + + + + + + + + + + + + + + 10 + −3 + + + + + + + + + + + + + + + + + + 10 + −2 + + + + + + + + + + + + + + + + + + 10 + −1 + + + + + + + + + + + + + + + + + + 10 + 0 + + + + + + + + + + + + + + + + + + 10 + 1 + + + + + + Held-out goals / episode (symlog) + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + Allegro + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + No sim + + + + + + + + + + + + + 100k + + + + + + + + + + + + + 1M + + + + + + + + + + + + + 10M + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + Practice transition upper bound / PPO frames + + + + + + + + + + + + + + + + + 0 + + + + + + + + + + + + + + + + + + 10 + −3 + + + + + + + + + + + + + + + + + + 10 + −2 + + + + + + + + + + + + + + + + + + 10 + −1 + + + + + + + + + + + + + + + + + + 10 + 0 + + + + + + + + + + + + + + + + + + 10 + 1 + + + + + + Held-out goals / episode (symlog) + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + Shadow + + + + + + + + + + + + + + + + + + + + + + + + Scratch policies: strong drawer control, weak dexterous manipulation + + + PPO checkpoints provide context; this is not a controlled sample-efficiency comparison. + + + LLM budget = rollouts × 256 environments × episode limit; excludes final evaluation. No-simulator attempts occupy a separate display lane. + Architectures, training/evaluation noise, compute, and prior knowledge differ. All plotted scores and budgets are available in attempts.csv. + + + + + + + + + + + + Opus 5.5 + + + + + + + + + + + Fable 5.1 + + + + + + Early PPO + + + + + + Expert + + + + + + + + + + + Rule B (A = circle) + + + + + + + + + + + + + + + diff --git a/artifacts/reports/figures/summary.pdf b/artifacts/reports/figures/summary.pdf new file mode 100644 index 0000000000000000000000000000000000000000..c4bae06f0b83451ff2a0c7950a68829de91ce983 Binary files /dev/null and b/artifacts/reports/figures/summary.pdf differ diff --git a/artifacts/reports/figures/summary.png b/artifacts/reports/figures/summary.png new file mode 100644 index 0000000000000000000000000000000000000000..95c2d761fb365aae68110f3e0cd7fafe40b1c0b8 --- /dev/null +++ b/artifacts/reports/figures/summary.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:9f46ca82747450f3a9a44e2dac323c33c7c5dff67efded0ddb3597b31c25fffc +size 546876 diff --git a/artifacts/reports/figures/summary.svg b/artifacts/reports/figures/summary.svg new file mode 100644 index 0000000000000000000000000000000000000000..4be9da2ce11a58cab73c99e504da2ae747c78405 --- /dev/null +++ b/artifacts/reports/figures/summary.svg @@ -0,0 +1,218 @@ + + + + + + + + 2026-11-04T19:21:05.612773 + image/svg+xml + + + Matplotlib v3.9.4, https://matplotlib.org/ + + + + + + + + + + + + + + + + + + + + + + + Can AI fix a robot by changing the numbers in its neural network? + + + We asked Opus and Fable to repair or build the software that controls simulated robots. + + + They edited network parameters directly, without the usual automated training process. + + + The robots had two jobs: open a drawer, or turn a cube in a robotic hand. + + + 1 Fix something we broke + + + YES, IN SOME TESTS + + + Start with a robot that works. + Deliberately damage its controller. + Ask the AI to find and fix the fault. + + + 0% → 99.98% + + + drawer-opening success after an Opus repair + + + Fable also restored strong cube-turning + performance by correcting one number. + But another drawer repair attempt failed. + + + 2 Improve a poor controller + + + SOMETIMES + + + Start with a robot that learned badly. + Ask the AI to improve the controller + it already has. + + + 1% → 100% + + + drawer-opening success in one Opus test + + + Some cube-turning controllers improved, + but stayed far below the trained reference. + Other edits made performance worse. + + + 3 Build a new controller + + + DEPENDS ON THE TASK + + + Give the AI an empty network. + Ask it to write the numbers needed + to control the robot. + + + 99.85% + + + drawer-opening success with an Opus design + + + Full controller rewrite using IK (joint-position + calculations), encoded in network weights. + Required five rounds of testing and revision. + + + Both hand tasks failed to achieve strong control + + + Robot hand + + + New: Opus + + + New: Fable + + + Best repair of poor controller + + + Trained reference + + + Allegro + + + 0.052 + + + 0.027 + + + 0.275 + + + 12.552 + + + Shadow + + + 0.023 + + + 0.013 + + + 2.673 + + + 30.945 + + + Hand scores = average target cube orientations reached per trial; higher is better. Values shown are the best observed in each category. + + + Fixing a simple fault in an already capable hand controller did succeed in some tests; creating or recovering strong skills did not. + + + How to interpret the drawer success + + + The new drawer controller used inverse kinematics (IK), not recovery of a learned skill. A separate 99.85% “repair” also replaced the old controller. + + + Other drawer repairs preserved existing behaviour. A full controller rewrite does not require changing every numeric entry. + + + Limits: mostly one AI attempt per setting. Early folder names revealed faults; 6/10 follow-ups that hid those clues have scores. + + + Drawer percentages measure openings in 4,096 simulated trials of one final controller, not the probability that an AI repair will work. + + + diff --git a/artifacts/results/attempts/hands/allegro_L_repair_early_ep200/fable-5-1/A_T2_K5_cap10/a0/init_reference.npz b/artifacts/results/attempts/hands/allegro_L_repair_early_ep200/fable-5-1/A_T2_K5_cap10/a0/init_reference.npz new file mode 100644 index 0000000000000000000000000000000000000000..504e9da1cf989a6161ed112d32a2ce80c9e9b5ac --- /dev/null +++ b/artifacts/results/attempts/hands/allegro_L_repair_early_ep200/fable-5-1/A_T2_K5_cap10/a0/init_reference.npz @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:83dff15ac3fd4f36048634d78e2d2783d3cdd4efa64ebbc28b0e3941189b0e41 +size 3280264 diff --git a/artifacts/results/attempts/hands/allegro_L_repair_early_ep200/fable-5-1/A_T2_K5_cap10/a0/submitted.npz b/artifacts/results/attempts/hands/allegro_L_repair_early_ep200/fable-5-1/A_T2_K5_cap10/a0/submitted.npz new file mode 100644 index 0000000000000000000000000000000000000000..ecf108d5f0caf6375624c91900fc8e728d6cbc53 --- /dev/null +++ b/artifacts/results/attempts/hands/allegro_L_repair_early_ep200/fable-5-1/A_T2_K5_cap10/a0/submitted.npz @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:dee08c8b30a6aa858e2e99cef5fd19e69bcf77699099b8c8eb6f3eae89b0f80c +size 3280728 diff --git a/artifacts/results/attempts/hands/allegro_L_repair_early_ep200/opus-5-5/A_T2_K5_cap10/a0/init_reference.npz b/artifacts/results/attempts/hands/allegro_L_repair_early_ep200/opus-5-5/A_T2_K5_cap10/a0/init_reference.npz new file mode 100644 index 0000000000000000000000000000000000000000..504e9da1cf989a6161ed112d32a2ce80c9e9b5ac --- /dev/null +++ b/artifacts/results/attempts/hands/allegro_L_repair_early_ep200/opus-5-5/A_T2_K5_cap10/a0/init_reference.npz @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:83dff15ac3fd4f36048634d78e2d2783d3cdd4efa64ebbc28b0e3941189b0e41 +size 3280264 diff --git a/artifacts/results/attempts/hands/allegro_L_repair_early_ep200/opus-5-5/A_T2_K5_cap10/a0/submitted.npz b/artifacts/results/attempts/hands/allegro_L_repair_early_ep200/opus-5-5/A_T2_K5_cap10/a0/submitted.npz new file mode 100644 index 0000000000000000000000000000000000000000..a0d800e31d9f2e595f2e69279ff8e7d82a1e83ce --- /dev/null +++ b/artifacts/results/attempts/hands/allegro_L_repair_early_ep200/opus-5-5/A_T2_K5_cap10/a0/submitted.npz @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:091bdad7e094ecaeeaf3a44f64184f53391ccf3ed67cd8c213f084d3312458e1 +size 3280512 diff --git a/artifacts/results/attempts/hands/allegro_L_repair_early_ep50/fable-5-1/A_T2_K5_cap10/a0/init_reference.npz b/artifacts/results/attempts/hands/allegro_L_repair_early_ep50/fable-5-1/A_T2_K5_cap10/a0/init_reference.npz new file mode 100644 index 0000000000000000000000000000000000000000..deabfd55a13d27d3f493330744ff81dcfed12521 --- /dev/null +++ b/artifacts/results/attempts/hands/allegro_L_repair_early_ep50/fable-5-1/A_T2_K5_cap10/a0/init_reference.npz @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:962e5d38f977c80dc648d5765743ca242ab635e0b151064e9f205ffb81033a77 +size 3280264 diff --git a/artifacts/results/attempts/hands/allegro_L_repair_early_ep50/fable-5-1/A_T2_K5_cap10/a0/submitted.npz b/artifacts/results/attempts/hands/allegro_L_repair_early_ep50/fable-5-1/A_T2_K5_cap10/a0/submitted.npz new file mode 100644 index 0000000000000000000000000000000000000000..22e0f9c7f0252e8a3177e91a6f5d5316e13e1692 --- /dev/null +++ b/artifacts/results/attempts/hands/allegro_L_repair_early_ep50/fable-5-1/A_T2_K5_cap10/a0/submitted.npz @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4052904907fec150a9125bd0099f91f232667150579cb457fae39fa9618a9180 +size 3281208 diff --git a/artifacts/results/attempts/hands/allegro_L_repair_early_ep50/opus-5-5/A_T2_K5_cap10/a0/init_reference.npz b/artifacts/results/attempts/hands/allegro_L_repair_early_ep50/opus-5-5/A_T2_K5_cap10/a0/init_reference.npz new file mode 100644 index 0000000000000000000000000000000000000000..deabfd55a13d27d3f493330744ff81dcfed12521 --- /dev/null +++ b/artifacts/results/attempts/hands/allegro_L_repair_early_ep50/opus-5-5/A_T2_K5_cap10/a0/init_reference.npz @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:962e5d38f977c80dc648d5765743ca242ab635e0b151064e9f205ffb81033a77 +size 3280264 diff --git a/artifacts/results/attempts/hands/allegro_L_repair_early_ep50/opus-5-5/A_T2_K5_cap10/a0/submitted.npz b/artifacts/results/attempts/hands/allegro_L_repair_early_ep50/opus-5-5/A_T2_K5_cap10/a0/submitted.npz new file mode 100644 index 0000000000000000000000000000000000000000..e721d9d35db7ab3e9f577c8c9ca06735d9a4bf23 --- /dev/null +++ b/artifacts/results/attempts/hands/allegro_L_repair_early_ep50/opus-5-5/A_T2_K5_cap10/a0/submitted.npz @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ca1129bc3fee14240b00fa47136dce52875f7bd056ead94b4cc1655c786addaf +size 3280264 diff --git a/artifacts/results/attempts/hands/allegro_L_repair_norm_shift_53/fable-5-1/A_T2_K5_cap10/a0/init_reference.npz b/artifacts/results/attempts/hands/allegro_L_repair_norm_shift_53/fable-5-1/A_T2_K5_cap10/a0/init_reference.npz new file mode 100644 index 0000000000000000000000000000000000000000..ce15ea4b7652fde3699d14bb2532255cd7f0c1f1 --- /dev/null +++ b/artifacts/results/attempts/hands/allegro_L_repair_norm_shift_53/fable-5-1/A_T2_K5_cap10/a0/init_reference.npz @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:16bbf725d37b183b0cd7278cae9956034ad42fa0cfa4ccdf5069c49e68d26fce +size 3280264 diff --git a/artifacts/results/attempts/hands/allegro_L_repair_norm_shift_53/fable-5-1/A_T2_K5_cap10/a0/submitted.npz b/artifacts/results/attempts/hands/allegro_L_repair_norm_shift_53/fable-5-1/A_T2_K5_cap10/a0/submitted.npz new file mode 100644 index 0000000000000000000000000000000000000000..342f8d95fe5f3e0ddec4aaa2cf01a4018a03ad1f --- /dev/null +++ b/artifacts/results/attempts/hands/allegro_L_repair_norm_shift_53/fable-5-1/A_T2_K5_cap10/a0/submitted.npz @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:46bbae9a6f9470dd87182cfa6e03860f18294b116af8b56b39e269059d990205 +size 3280512 diff --git a/artifacts/results/attempts/hands/allegro_L_repair_norm_shift_53/fable-5-1/A_T2_K5_cap10/blind0/init_reference.npz b/artifacts/results/attempts/hands/allegro_L_repair_norm_shift_53/fable-5-1/A_T2_K5_cap10/blind0/init_reference.npz new file mode 100644 index 0000000000000000000000000000000000000000..ce15ea4b7652fde3699d14bb2532255cd7f0c1f1 --- /dev/null +++ b/artifacts/results/attempts/hands/allegro_L_repair_norm_shift_53/fable-5-1/A_T2_K5_cap10/blind0/init_reference.npz @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:16bbf725d37b183b0cd7278cae9956034ad42fa0cfa4ccdf5069c49e68d26fce +size 3280264 diff --git a/artifacts/results/attempts/hands/allegro_L_repair_norm_shift_53/fable-5-1/A_T2_K5_cap10/blind0/submitted.npz b/artifacts/results/attempts/hands/allegro_L_repair_norm_shift_53/fable-5-1/A_T2_K5_cap10/blind0/submitted.npz new file mode 100644 index 0000000000000000000000000000000000000000..342f8d95fe5f3e0ddec4aaa2cf01a4018a03ad1f --- /dev/null +++ b/artifacts/results/attempts/hands/allegro_L_repair_norm_shift_53/fable-5-1/A_T2_K5_cap10/blind0/submitted.npz @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:46bbae9a6f9470dd87182cfa6e03860f18294b116af8b56b39e269059d990205 +size 3280512 diff --git a/artifacts/results/attempts/hands/allegro_L_repair_swap_act_4_7/opus-5-5/A_T2_K5_cap10/a0/init_reference.npz b/artifacts/results/attempts/hands/allegro_L_repair_swap_act_4_7/opus-5-5/A_T2_K5_cap10/a0/init_reference.npz new file mode 100644 index 0000000000000000000000000000000000000000..693be250c205d7e51b1a88dfe0faea6759346141 --- /dev/null +++ b/artifacts/results/attempts/hands/allegro_L_repair_swap_act_4_7/opus-5-5/A_T2_K5_cap10/a0/init_reference.npz @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a912eb2398e920bb9129b08186b7ed8603c1325d842331bc38e4becb2ccf8078 +size 3280264 diff --git a/artifacts/results/attempts/hands/allegro_L_repair_swap_act_4_7/opus-5-5/A_T2_K5_cap10/a0/submitted.npz b/artifacts/results/attempts/hands/allegro_L_repair_swap_act_4_7/opus-5-5/A_T2_K5_cap10/a0/submitted.npz new file mode 100644 index 0000000000000000000000000000000000000000..19b5ff57beeb65a0079f74cc2a7c58eace4878ed --- /dev/null +++ b/artifacts/results/attempts/hands/allegro_L_repair_swap_act_4_7/opus-5-5/A_T2_K5_cap10/a0/submitted.npz @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b58b54ef6e6be2b2dd23f03175e142dccefeec2120d80933e83d281d6d690467 +size 3280264 diff --git a/artifacts/results/attempts/hands/allegro_XS_write/fable-5-1/A_T2_K5/a0/init_reference.npz b/artifacts/results/attempts/hands/allegro_XS_write/fable-5-1/A_T2_K5/a0/init_reference.npz new file mode 100644 index 0000000000000000000000000000000000000000..4595e38522721f4ccbba1f5f03662dfc42dce53c --- /dev/null +++ b/artifacts/results/attempts/hands/allegro_XS_write/fable-5-1/A_T2_K5/a0/init_reference.npz @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:734a84976e9e4596104bcce0e6053900d668396ca3dc17c5cf3fd4aa14183c24 +size 55752 diff --git a/artifacts/results/attempts/hands/allegro_XS_write/fable-5-1/A_T2_K5/a0/submitted.npz b/artifacts/results/attempts/hands/allegro_XS_write/fable-5-1/A_T2_K5/a0/submitted.npz new file mode 100644 index 0000000000000000000000000000000000000000..5eda9a2e8f6e69c8024926e7bcd0fefd22b251b5 --- /dev/null +++ b/artifacts/results/attempts/hands/allegro_XS_write/fable-5-1/A_T2_K5/a0/submitted.npz @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:5114a7071069acd95858fe577f02818079266049f7c2e75454ef34c8ea6f589d +size 56000 diff --git a/artifacts/results/attempts/hands/allegro_XS_write/fable-5-1/B_T2_K5/a0/init_reference.npz b/artifacts/results/attempts/hands/allegro_XS_write/fable-5-1/B_T2_K5/a0/init_reference.npz new file mode 100644 index 0000000000000000000000000000000000000000..4595e38522721f4ccbba1f5f03662dfc42dce53c --- /dev/null +++ b/artifacts/results/attempts/hands/allegro_XS_write/fable-5-1/B_T2_K5/a0/init_reference.npz @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:734a84976e9e4596104bcce0e6053900d668396ca3dc17c5cf3fd4aa14183c24 +size 55752 diff --git a/artifacts/results/attempts/hands/allegro_XS_write/fable-5-1/B_T2_K5/a0/submitted.npz b/artifacts/results/attempts/hands/allegro_XS_write/fable-5-1/B_T2_K5/a0/submitted.npz new file mode 100644 index 0000000000000000000000000000000000000000..6a1f1675b9f882756f04cd8ffac672e0c320ba6f --- /dev/null +++ b/artifacts/results/attempts/hands/allegro_XS_write/fable-5-1/B_T2_K5/a0/submitted.npz @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f3dd6e43164726fb1470c00460dcf541f273ec204e37439e54e5692e35cb9e19 +size 56000 diff --git a/artifacts/results/attempts/hands/allegro_XS_write/opus-5-5/A_T2_K5/a0/init_reference.npz b/artifacts/results/attempts/hands/allegro_XS_write/opus-5-5/A_T2_K5/a0/init_reference.npz new file mode 100644 index 0000000000000000000000000000000000000000..4595e38522721f4ccbba1f5f03662dfc42dce53c --- /dev/null +++ b/artifacts/results/attempts/hands/allegro_XS_write/opus-5-5/A_T2_K5/a0/init_reference.npz @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:734a84976e9e4596104bcce0e6053900d668396ca3dc17c5cf3fd4aa14183c24 +size 55752 diff --git a/artifacts/results/attempts/hands/allegro_XS_write/opus-5-5/A_T2_K5/a0/submitted.npz b/artifacts/results/attempts/hands/allegro_XS_write/opus-5-5/A_T2_K5/a0/submitted.npz new file mode 100644 index 0000000000000000000000000000000000000000..e5eaaf2bd2c46f47f80c9340f2fde95affc6562a --- /dev/null +++ b/artifacts/results/attempts/hands/allegro_XS_write/opus-5-5/A_T2_K5/a0/submitted.npz @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:1ca1ce3f969ce8bf81109757b8b4549cb900de4ccfc40d3863cf94e7a24682d9 +size 56000 diff --git a/artifacts/results/attempts/hands/allegro_XS_write/opus-5-5/B_T2_K5/a0/init_reference.npz b/artifacts/results/attempts/hands/allegro_XS_write/opus-5-5/B_T2_K5/a0/init_reference.npz new file mode 100644 index 0000000000000000000000000000000000000000..4595e38522721f4ccbba1f5f03662dfc42dce53c --- /dev/null +++ b/artifacts/results/attempts/hands/allegro_XS_write/opus-5-5/B_T2_K5/a0/init_reference.npz @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:734a84976e9e4596104bcce0e6053900d668396ca3dc17c5cf3fd4aa14183c24 +size 55752 diff --git a/artifacts/results/attempts/hands/allegro_XS_write/opus-5-5/B_T2_K5/a0/submitted.npz b/artifacts/results/attempts/hands/allegro_XS_write/opus-5-5/B_T2_K5/a0/submitted.npz new file mode 100644 index 0000000000000000000000000000000000000000..f7b414d506571c06b1af188f05acb25941eadb0e --- /dev/null +++ b/artifacts/results/attempts/hands/allegro_XS_write/opus-5-5/B_T2_K5/a0/submitted.npz @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3e087c510c31e8b9105f78e7749c99d667dbaaada2444f6d0756f02e557fc1d6 +size 57076 diff --git a/artifacts/results/attempts/hands/shadow_L_repair_early_ep100/fable-5-1/A_T2_K5_cap10/a0/init_reference.npz b/artifacts/results/attempts/hands/shadow_L_repair_early_ep100/fable-5-1/A_T2_K5_cap10/a0/init_reference.npz new file mode 100644 index 0000000000000000000000000000000000000000..ff24fb047e6a91773d37cdce8c8ff4b524b7b481 --- /dev/null +++ b/artifacts/results/attempts/hands/shadow_L_repair_early_ep100/fable-5-1/A_T2_K5_cap10/a0/init_reference.npz @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:9e4d9685886178e4258327a6b8f5a64ca9cf320be26454e1ba9dab29d345e13f +size 2045600 diff --git a/artifacts/results/attempts/hands/shadow_L_repair_early_ep100/fable-5-1/A_T2_K5_cap10/a0/submitted.npz b/artifacts/results/attempts/hands/shadow_L_repair_early_ep100/fable-5-1/A_T2_K5_cap10/a0/submitted.npz new file mode 100644 index 0000000000000000000000000000000000000000..cf915e2aa41c459b425c3653b1d60fa11dc30a5f --- /dev/null +++ b/artifacts/results/attempts/hands/shadow_L_repair_early_ep100/fable-5-1/A_T2_K5_cap10/a0/submitted.npz @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:2363589ada764b30c20f1cf5647b5001a93e1e350f69b34d941fbb93af3d351d +size 2045848 diff --git a/artifacts/results/attempts/hands/shadow_L_repair_early_ep100/opus-5-5/A_T2_K5_cap10/a0/init_reference.npz b/artifacts/results/attempts/hands/shadow_L_repair_early_ep100/opus-5-5/A_T2_K5_cap10/a0/init_reference.npz new file mode 100644 index 0000000000000000000000000000000000000000..ff24fb047e6a91773d37cdce8c8ff4b524b7b481 --- /dev/null +++ b/artifacts/results/attempts/hands/shadow_L_repair_early_ep100/opus-5-5/A_T2_K5_cap10/a0/init_reference.npz @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:9e4d9685886178e4258327a6b8f5a64ca9cf320be26454e1ba9dab29d345e13f +size 2045600 diff --git a/artifacts/results/attempts/hands/shadow_L_repair_early_ep100/opus-5-5/A_T2_K5_cap10/a0/submitted.npz b/artifacts/results/attempts/hands/shadow_L_repair_early_ep100/opus-5-5/A_T2_K5_cap10/a0/submitted.npz new file mode 100644 index 0000000000000000000000000000000000000000..8d7c4c36346e98cfe831b43db1267b5d73d11879 --- /dev/null +++ b/artifacts/results/attempts/hands/shadow_L_repair_early_ep100/opus-5-5/A_T2_K5_cap10/a0/submitted.npz @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ec3f825b99cde6db9850c8b0976e335fbfecf3c9994112dec621e473d90b7059 +size 2046100 diff --git a/artifacts/results/attempts/hands/shadow_L_repair_early_ep200/fable-5-1/A_T2_K5_cap10/a0/init_reference.npz b/artifacts/results/attempts/hands/shadow_L_repair_early_ep200/fable-5-1/A_T2_K5_cap10/a0/init_reference.npz new file mode 100644 index 0000000000000000000000000000000000000000..6f7d3a1cd974da5dceeaeda89f36b27eafa6bceb --- /dev/null +++ b/artifacts/results/attempts/hands/shadow_L_repair_early_ep200/fable-5-1/A_T2_K5_cap10/a0/init_reference.npz @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:553a9eb82d9680976f727df64ece1d43dae1a2f7cca25194b06fcb590f98b661 +size 2045600 diff --git a/artifacts/results/attempts/hands/shadow_L_repair_early_ep200/fable-5-1/A_T2_K5_cap10/a0/submitted.npz b/artifacts/results/attempts/hands/shadow_L_repair_early_ep200/fable-5-1/A_T2_K5_cap10/a0/submitted.npz new file mode 100644 index 0000000000000000000000000000000000000000..444fea75f4056a64a3b4dadeb08ce1521f208f87 --- /dev/null +++ b/artifacts/results/attempts/hands/shadow_L_repair_early_ep200/fable-5-1/A_T2_K5_cap10/a0/submitted.npz @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:8677efe5de181dbc3742e9d713e532d60ff881d26939da8fbe7ef906f533bf44 +size 2045848 diff --git a/artifacts/results/attempts/hands/shadow_L_repair_early_ep200/opus-5-5/A_T2_K5_cap10/a0/init_reference.npz b/artifacts/results/attempts/hands/shadow_L_repair_early_ep200/opus-5-5/A_T2_K5_cap10/a0/init_reference.npz new file mode 100644 index 0000000000000000000000000000000000000000..6f7d3a1cd974da5dceeaeda89f36b27eafa6bceb --- /dev/null +++ b/artifacts/results/attempts/hands/shadow_L_repair_early_ep200/opus-5-5/A_T2_K5_cap10/a0/init_reference.npz @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:553a9eb82d9680976f727df64ece1d43dae1a2f7cca25194b06fcb590f98b661 +size 2045600 diff --git a/artifacts/results/attempts/hands/shadow_L_repair_early_ep200/opus-5-5/A_T2_K5_cap10/a0/submitted.npz b/artifacts/results/attempts/hands/shadow_L_repair_early_ep200/opus-5-5/A_T2_K5_cap10/a0/submitted.npz new file mode 100644 index 0000000000000000000000000000000000000000..c1c63c4eb1cde0c11e291545647c471cf3ad1374 --- /dev/null +++ b/artifacts/results/attempts/hands/shadow_L_repair_early_ep200/opus-5-5/A_T2_K5_cap10/a0/submitted.npz @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:48da7a80a7c2bdeaecab9961e4a09109b29ab2dca79f48db596539cd73c0c1a2 +size 2045848 diff --git a/artifacts/results/attempts/hands/shadow_L_repair_norm_shift_69/fable-5-1/A_T2_K5_cap10/a0/init_reference.npz b/artifacts/results/attempts/hands/shadow_L_repair_norm_shift_69/fable-5-1/A_T2_K5_cap10/a0/init_reference.npz new file mode 100644 index 0000000000000000000000000000000000000000..8f1acd7ae07ebe10179a9e471ce95e5d88af44cf --- /dev/null +++ b/artifacts/results/attempts/hands/shadow_L_repair_norm_shift_69/fable-5-1/A_T2_K5_cap10/a0/init_reference.npz @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:90e53d7b453d254827a81c9bcbf2605ca2e2791c2c65aa3661b5eff7d91565a4 +size 2045600 diff --git a/artifacts/results/attempts/hands/shadow_L_repair_norm_shift_69/fable-5-1/A_T2_K5_cap10/a0/submitted.npz b/artifacts/results/attempts/hands/shadow_L_repair_norm_shift_69/fable-5-1/A_T2_K5_cap10/a0/submitted.npz new file mode 100644 index 0000000000000000000000000000000000000000..31de73cc38ce27b99881d7707f3f31f730dc4885 --- /dev/null +++ b/artifacts/results/attempts/hands/shadow_L_repair_norm_shift_69/fable-5-1/A_T2_K5_cap10/a0/submitted.npz @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3cf2562eef969b3c5b3d3f4fc05163c3bc3c121a33820d95cdeb60dd41de41d7 +size 2045848 diff --git a/artifacts/results/attempts/hands/shadow_L_repair_norm_shift_69/fable-5-1/A_T2_K5_cap10/blind0/init_reference.npz b/artifacts/results/attempts/hands/shadow_L_repair_norm_shift_69/fable-5-1/A_T2_K5_cap10/blind0/init_reference.npz new file mode 100644 index 0000000000000000000000000000000000000000..8f1acd7ae07ebe10179a9e471ce95e5d88af44cf --- /dev/null +++ b/artifacts/results/attempts/hands/shadow_L_repair_norm_shift_69/fable-5-1/A_T2_K5_cap10/blind0/init_reference.npz @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:90e53d7b453d254827a81c9bcbf2605ca2e2791c2c65aa3661b5eff7d91565a4 +size 2045600 diff --git a/artifacts/results/attempts/hands/shadow_L_repair_norm_shift_69/fable-5-1/A_T2_K5_cap10/blind0/submitted.npz b/artifacts/results/attempts/hands/shadow_L_repair_norm_shift_69/fable-5-1/A_T2_K5_cap10/blind0/submitted.npz new file mode 100644 index 0000000000000000000000000000000000000000..31de73cc38ce27b99881d7707f3f31f730dc4885 --- /dev/null +++ b/artifacts/results/attempts/hands/shadow_L_repair_norm_shift_69/fable-5-1/A_T2_K5_cap10/blind0/submitted.npz @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3cf2562eef969b3c5b3d3f4fc05163c3bc3c121a33820d95cdeb60dd41de41d7 +size 2045848 diff --git a/artifacts/results/attempts/hands/shadow_L_repair_swap_act_7_11/opus-5-5/A_T2_K5_cap10/a0/init_reference.npz b/artifacts/results/attempts/hands/shadow_L_repair_swap_act_7_11/opus-5-5/A_T2_K5_cap10/a0/init_reference.npz new file mode 100644 index 0000000000000000000000000000000000000000..86234cf55abbfc5f0920e56791b38741c9c608f3 --- /dev/null +++ b/artifacts/results/attempts/hands/shadow_L_repair_swap_act_7_11/opus-5-5/A_T2_K5_cap10/a0/init_reference.npz @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:469ff9e9a23f5b7b91fd61ee72466d1c070b906b7bf917f82b6a23dd4832579c +size 2045600 diff --git a/artifacts/results/attempts/hands/shadow_L_repair_swap_act_7_11/opus-5-5/A_T2_K5_cap10/a0/submitted.npz b/artifacts/results/attempts/hands/shadow_L_repair_swap_act_7_11/opus-5-5/A_T2_K5_cap10/a0/submitted.npz new file mode 100644 index 0000000000000000000000000000000000000000..b9546e545f3209d1441ebb05050c0eeeba50cd3c --- /dev/null +++ b/artifacts/results/attempts/hands/shadow_L_repair_swap_act_7_11/opus-5-5/A_T2_K5_cap10/a0/submitted.npz @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4fa39ebb60303fe1563aab7eb3d96bea2d7f6ac1f1ff11e368089aec981659a1 +size 2045600 diff --git a/artifacts/results/attempts/hands/shadow_XS_write/fable-5-1/A_T2_K5/a0/init_reference.npz b/artifacts/results/attempts/hands/shadow_XS_write/fable-5-1/A_T2_K5/a0/init_reference.npz new file mode 100644 index 0000000000000000000000000000000000000000..c35085af61481ca7dd5b8ca88f177e54a93e7c0b --- /dev/null +++ b/artifacts/results/attempts/hands/shadow_XS_write/fable-5-1/A_T2_K5/a0/init_reference.npz @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a2c82623b46c5121650a575c5a0b6525fd86b2a6ed1ace507d7829f902379ee0 +size 65504 diff --git a/artifacts/results/attempts/hands/shadow_XS_write/fable-5-1/A_T2_K5/a0/submitted.npz b/artifacts/results/attempts/hands/shadow_XS_write/fable-5-1/A_T2_K5/a0/submitted.npz new file mode 100644 index 0000000000000000000000000000000000000000..704f5593bfc00bb6481a10e05d04e3f851cee673 --- /dev/null +++ b/artifacts/results/attempts/hands/shadow_XS_write/fable-5-1/A_T2_K5/a0/submitted.npz @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:93bf86d50b379b33848ab89c99e25a74a31490669ad9745c78dad71c757be7dd +size 65752 diff --git a/artifacts/results/attempts/hands/shadow_XS_write/fable-5-1/B_T2_K5/a0/init_reference.npz b/artifacts/results/attempts/hands/shadow_XS_write/fable-5-1/B_T2_K5/a0/init_reference.npz new file mode 100644 index 0000000000000000000000000000000000000000..c35085af61481ca7dd5b8ca88f177e54a93e7c0b --- /dev/null +++ b/artifacts/results/attempts/hands/shadow_XS_write/fable-5-1/B_T2_K5/a0/init_reference.npz @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a2c82623b46c5121650a575c5a0b6525fd86b2a6ed1ace507d7829f902379ee0 +size 65504 diff --git a/artifacts/results/attempts/hands/shadow_XS_write/fable-5-1/B_T2_K5/a0/submitted.npz b/artifacts/results/attempts/hands/shadow_XS_write/fable-5-1/B_T2_K5/a0/submitted.npz new file mode 100644 index 0000000000000000000000000000000000000000..0a08abf10529de8aa0222a499018b2bb718a11a1 --- /dev/null +++ b/artifacts/results/attempts/hands/shadow_XS_write/fable-5-1/B_T2_K5/a0/submitted.npz @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ac4940efa41f77ca964d64e10d8471fa6ccb74064f7a025ad203fd8cd37b675e +size 65752 diff --git a/artifacts/results/attempts/hands/shadow_XS_write/opus-5-5/A_T2_K5/a0/init_reference.npz b/artifacts/results/attempts/hands/shadow_XS_write/opus-5-5/A_T2_K5/a0/init_reference.npz new file mode 100644 index 0000000000000000000000000000000000000000..c35085af61481ca7dd5b8ca88f177e54a93e7c0b --- /dev/null +++ b/artifacts/results/attempts/hands/shadow_XS_write/opus-5-5/A_T2_K5/a0/init_reference.npz @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a2c82623b46c5121650a575c5a0b6525fd86b2a6ed1ace507d7829f902379ee0 +size 65504 diff --git a/artifacts/results/attempts/hands/shadow_XS_write/opus-5-5/A_T2_K5/a0/submitted.npz b/artifacts/results/attempts/hands/shadow_XS_write/opus-5-5/A_T2_K5/a0/submitted.npz new file mode 100644 index 0000000000000000000000000000000000000000..ec3aa58bc70091ee23a7a6fcff5da959631dfa61 --- /dev/null +++ b/artifacts/results/attempts/hands/shadow_XS_write/opus-5-5/A_T2_K5/a0/submitted.npz @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e705e7871dc90da44eb7a5fc965017aeca214349233afda14a79e032d6213ba3 +size 66616 diff --git a/artifacts/results/attempts/hands/shadow_XS_write/opus-5-5/B_T2_K5/a0/init_reference.npz b/artifacts/results/attempts/hands/shadow_XS_write/opus-5-5/B_T2_K5/a0/init_reference.npz new file mode 100644 index 0000000000000000000000000000000000000000..c35085af61481ca7dd5b8ca88f177e54a93e7c0b --- /dev/null +++ b/artifacts/results/attempts/hands/shadow_XS_write/opus-5-5/B_T2_K5/a0/init_reference.npz @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a2c82623b46c5121650a575c5a0b6525fd86b2a6ed1ace507d7829f902379ee0 +size 65504 diff --git a/artifacts/results/attempts/hands/shadow_XS_write/opus-5-5/B_T2_K5/a0/submitted.npz b/artifacts/results/attempts/hands/shadow_XS_write/opus-5-5/B_T2_K5/a0/submitted.npz new file mode 100644 index 0000000000000000000000000000000000000000..e2f6e9c8b0c52b67d730f45728cba07575c86ef3 --- /dev/null +++ b/artifacts/results/attempts/hands/shadow_XS_write/opus-5-5/B_T2_K5/a0/submitted.npz @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:358a05853e5482d0b059b5725d5a80a8fb538aa6c479aba1a04d230af2eff77c +size 65752 diff --git a/artifacts/results/attempts/main/franka_M_repair_collapse_ep500/opus-5-5/A_T0_K0_cap10/a0/init_reference.npz b/artifacts/results/attempts/main/franka_M_repair_collapse_ep500/opus-5-5/A_T0_K0_cap10/a0/init_reference.npz new file mode 100644 index 0000000000000000000000000000000000000000..52bdc236f2dceb4b001121b21a078adbb85b8d08 --- /dev/null +++ b/artifacts/results/attempts/main/franka_M_repair_collapse_ep500/opus-5-5/A_T0_K0_cap10/a0/init_reference.npz @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:1e7762a844130a1f1ea096a79797557475aee2265f37ef6615a03ee90ee69ed8 +size 194148 diff --git a/artifacts/results/attempts/main/franka_M_repair_collapse_ep500/opus-5-5/A_T0_K0_cap10/a0/submitted.npz b/artifacts/results/attempts/main/franka_M_repair_collapse_ep500/opus-5-5/A_T0_K0_cap10/a0/submitted.npz new file mode 100644 index 0000000000000000000000000000000000000000..8275d235f217bb3d5eebf769cf57aae45e440424 --- /dev/null +++ b/artifacts/results/attempts/main/franka_M_repair_collapse_ep500/opus-5-5/A_T0_K0_cap10/a0/submitted.npz @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6d6c6bd70b5d9bd3d9b5be621bf2c6f0c016b17ca3dc94aa2b14db793bb06f2b +size 194148 diff --git a/artifacts/results/attempts/main/franka_M_repair_collapse_ep500/opus-5-5/A_T2_K5_cap10/a0/init_reference.npz b/artifacts/results/attempts/main/franka_M_repair_collapse_ep500/opus-5-5/A_T2_K5_cap10/a0/init_reference.npz new file mode 100644 index 0000000000000000000000000000000000000000..52bdc236f2dceb4b001121b21a078adbb85b8d08 --- /dev/null +++ b/artifacts/results/attempts/main/franka_M_repair_collapse_ep500/opus-5-5/A_T2_K5_cap10/a0/init_reference.npz @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:1e7762a844130a1f1ea096a79797557475aee2265f37ef6615a03ee90ee69ed8 +size 194148 diff --git a/artifacts/results/attempts/main/franka_M_repair_collapse_ep500/opus-5-5/A_T2_K5_cap10/a0/submitted.npz b/artifacts/results/attempts/main/franka_M_repair_collapse_ep500/opus-5-5/A_T2_K5_cap10/a0/submitted.npz new file mode 100644 index 0000000000000000000000000000000000000000..7d6ce0a3d21c635e3693a6bc9aec61d1ffbbbca7 --- /dev/null +++ b/artifacts/results/attempts/main/franka_M_repair_collapse_ep500/opus-5-5/A_T2_K5_cap10/a0/submitted.npz @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:68c10234c9d96fbde4f63ff8045219544bd963697a8f83ffc6e850c80868a36d +size 194396 diff --git a/artifacts/results/attempts/main/franka_M_repair_early_ep15/opus-5-5/A_T0_K0_cap10/a0/init_reference.npz b/artifacts/results/attempts/main/franka_M_repair_early_ep15/opus-5-5/A_T0_K0_cap10/a0/init_reference.npz new file mode 100644 index 0000000000000000000000000000000000000000..0345c2945b20baabf4c1a14b39b946715e39393a --- /dev/null +++ b/artifacts/results/attempts/main/franka_M_repair_early_ep15/opus-5-5/A_T0_K0_cap10/a0/init_reference.npz @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:feeb4d3f2ddfa9d0b6b585c7205c1d110a3db236213206d1f5558e6bc8507047 +size 194148 diff --git a/artifacts/results/attempts/main/franka_M_repair_early_ep15/opus-5-5/A_T0_K0_cap10/a0/submitted.npz b/artifacts/results/attempts/main/franka_M_repair_early_ep15/opus-5-5/A_T0_K0_cap10/a0/submitted.npz new file mode 100644 index 0000000000000000000000000000000000000000..42c0b736f1a45a1e92ce25d0b7395d453638638c --- /dev/null +++ b/artifacts/results/attempts/main/franka_M_repair_early_ep15/opus-5-5/A_T0_K0_cap10/a0/submitted.npz @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:9ad5cdd13c12b9af908d396c927e1318a4879bd2dfdc40b9dccd4fecda630fc0 +size 194396 diff --git a/artifacts/results/attempts/main/franka_M_repair_early_ep15/opus-5-5/A_T2_K5_cap10/a0/init_reference.npz b/artifacts/results/attempts/main/franka_M_repair_early_ep15/opus-5-5/A_T2_K5_cap10/a0/init_reference.npz new file mode 100644 index 0000000000000000000000000000000000000000..0345c2945b20baabf4c1a14b39b946715e39393a --- /dev/null +++ b/artifacts/results/attempts/main/franka_M_repair_early_ep15/opus-5-5/A_T2_K5_cap10/a0/init_reference.npz @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:feeb4d3f2ddfa9d0b6b585c7205c1d110a3db236213206d1f5558e6bc8507047 +size 194148 diff --git a/artifacts/results/attempts/main/franka_M_repair_early_ep15/opus-5-5/A_T2_K5_cap10/a0/submitted.npz b/artifacts/results/attempts/main/franka_M_repair_early_ep15/opus-5-5/A_T2_K5_cap10/a0/submitted.npz new file mode 100644 index 0000000000000000000000000000000000000000..46070c14a10213cad7d27334b47fea0cdca4f0b3 --- /dev/null +++ b/artifacts/results/attempts/main/franka_M_repair_early_ep15/opus-5-5/A_T2_K5_cap10/a0/submitted.npz @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:2aa13d70587bcacd3fd270156047e474b3aadfb41f2c1c4b1532cf521bae04a1 +size 194396 diff --git a/artifacts/results/attempts/main/franka_M_repair_early_ep20/opus-5-5/A_T0_K0_cap10/a0/init_reference.npz b/artifacts/results/attempts/main/franka_M_repair_early_ep20/opus-5-5/A_T0_K0_cap10/a0/init_reference.npz new file mode 100644 index 0000000000000000000000000000000000000000..ae3216b1750b1bd54e18d6d477bd5acb57be17e9 --- /dev/null +++ b/artifacts/results/attempts/main/franka_M_repair_early_ep20/opus-5-5/A_T0_K0_cap10/a0/init_reference.npz @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:fb12def9ecfbd7531312f7d0690d326bc550ac94d6ae3197c0c770fb8c1af0f8 +size 194148 diff --git a/artifacts/results/attempts/main/franka_M_repair_early_ep20/opus-5-5/A_T0_K0_cap10/a0/submitted.npz b/artifacts/results/attempts/main/franka_M_repair_early_ep20/opus-5-5/A_T0_K0_cap10/a0/submitted.npz new file mode 100644 index 0000000000000000000000000000000000000000..c3b997af46609f56833711d8eeba54509ffd3da0 --- /dev/null +++ b/artifacts/results/attempts/main/franka_M_repair_early_ep20/opus-5-5/A_T0_K0_cap10/a0/submitted.npz @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4a6ffa580c733429a97fbb429bdb5cb70a7712dbbfa3b211ca85468291b94387 +size 194148 diff --git a/artifacts/results/attempts/main/franka_M_repair_early_ep20/opus-5-5/A_T2_K5_cap10/a0/init_reference.npz b/artifacts/results/attempts/main/franka_M_repair_early_ep20/opus-5-5/A_T2_K5_cap10/a0/init_reference.npz new file mode 100644 index 0000000000000000000000000000000000000000..ae3216b1750b1bd54e18d6d477bd5acb57be17e9 --- /dev/null +++ b/artifacts/results/attempts/main/franka_M_repair_early_ep20/opus-5-5/A_T2_K5_cap10/a0/init_reference.npz @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:fb12def9ecfbd7531312f7d0690d326bc550ac94d6ae3197c0c770fb8c1af0f8 +size 194148 diff --git a/artifacts/results/attempts/main/franka_M_repair_early_ep20/opus-5-5/A_T2_K5_cap10/a0/submitted.npz b/artifacts/results/attempts/main/franka_M_repair_early_ep20/opus-5-5/A_T2_K5_cap10/a0/submitted.npz new file mode 100644 index 0000000000000000000000000000000000000000..a3f505ca6534fa3eb91f8b567db5ab3b1ad74f2d --- /dev/null +++ b/artifacts/results/attempts/main/franka_M_repair_early_ep20/opus-5-5/A_T2_K5_cap10/a0/submitted.npz @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:80882383fd905abcc536a4d5e47f4066c116d943a2149a88fdaf54e42005a12a +size 194396 diff --git a/artifacts/results/attempts/main/franka_M_repair_early_ep40/opus-5-5/A_T0_K0_cap10/a0/init_reference.npz b/artifacts/results/attempts/main/franka_M_repair_early_ep40/opus-5-5/A_T0_K0_cap10/a0/init_reference.npz new file mode 100644 index 0000000000000000000000000000000000000000..3a7ebe31bbd2ad0fbc01dc5f6638c18e2decf707 --- /dev/null +++ b/artifacts/results/attempts/main/franka_M_repair_early_ep40/opus-5-5/A_T0_K0_cap10/a0/init_reference.npz @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:29b96da17627852282b904f6b98e0d7fa16720fcb18be8a1005b97c6ebe1bf50 +size 194148 diff --git a/artifacts/results/attempts/main/franka_M_repair_early_ep40/opus-5-5/A_T0_K0_cap10/a0/submitted.npz b/artifacts/results/attempts/main/franka_M_repair_early_ep40/opus-5-5/A_T0_K0_cap10/a0/submitted.npz new file mode 100644 index 0000000000000000000000000000000000000000..a904715dac1a2e452f403653d65f2a45abd5bc6a --- /dev/null +++ b/artifacts/results/attempts/main/franka_M_repair_early_ep40/opus-5-5/A_T0_K0_cap10/a0/submitted.npz @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6e856294d2d666197fd7245f766487f89642dfe6826956c821fe7dc4b0d74ad8 +size 194148 diff --git a/artifacts/results/attempts/main/franka_M_repair_norm_shift_21/opus-5-5/A_T0_K0_cap10/a0/init_reference.npz b/artifacts/results/attempts/main/franka_M_repair_norm_shift_21/opus-5-5/A_T0_K0_cap10/a0/init_reference.npz new file mode 100644 index 0000000000000000000000000000000000000000..2e9a5f1caa5982a9a340df831140a7202e4aeca1 --- /dev/null +++ b/artifacts/results/attempts/main/franka_M_repair_norm_shift_21/opus-5-5/A_T0_K0_cap10/a0/init_reference.npz @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:5f165e2433e09f4894d3cbde8b4ff23f8447cd2d92ebf954d538538930ff1ea3 +size 194148 diff --git a/artifacts/results/attempts/main/franka_M_repair_norm_shift_21/opus-5-5/A_T0_K0_cap10/a0/submitted.npz b/artifacts/results/attempts/main/franka_M_repair_norm_shift_21/opus-5-5/A_T0_K0_cap10/a0/submitted.npz new file mode 100644 index 0000000000000000000000000000000000000000..c8156bf30df7e7e4b50cd26717493e8a720fbdb2 --- /dev/null +++ b/artifacts/results/attempts/main/franka_M_repair_norm_shift_21/opus-5-5/A_T0_K0_cap10/a0/submitted.npz @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:88de103d6c963a22904a46793f85ee7811665f6e955583eb4c355164a8bd474a +size 194148 diff --git a/artifacts/results/attempts/main/franka_M_repair_norm_shift_21/opus-5-5/A_T2_K5_cap10/a0/init_reference.npz b/artifacts/results/attempts/main/franka_M_repair_norm_shift_21/opus-5-5/A_T2_K5_cap10/a0/init_reference.npz new file mode 100644 index 0000000000000000000000000000000000000000..2e9a5f1caa5982a9a340df831140a7202e4aeca1 --- /dev/null +++ b/artifacts/results/attempts/main/franka_M_repair_norm_shift_21/opus-5-5/A_T2_K5_cap10/a0/init_reference.npz @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:5f165e2433e09f4894d3cbde8b4ff23f8447cd2d92ebf954d538538930ff1ea3 +size 194148 diff --git a/artifacts/results/attempts/main/franka_M_repair_norm_shift_21/opus-5-5/A_T2_K5_cap10/a0/submitted.npz b/artifacts/results/attempts/main/franka_M_repair_norm_shift_21/opus-5-5/A_T2_K5_cap10/a0/submitted.npz new file mode 100644 index 0000000000000000000000000000000000000000..eb67c62aa9eb4e6e7392732212ba37f84e64dd72 --- /dev/null +++ b/artifacts/results/attempts/main/franka_M_repair_norm_shift_21/opus-5-5/A_T2_K5_cap10/a0/submitted.npz @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ebfda4cf4cff5daf24025797841fa8b80f0707396824889789cf1971048a1c4b +size 194396 diff --git a/artifacts/results/attempts/main/franka_M_repair_norm_shift_21/opus-5-5/B_T0_K0_cap10/a0/init_reference.npz b/artifacts/results/attempts/main/franka_M_repair_norm_shift_21/opus-5-5/B_T0_K0_cap10/a0/init_reference.npz new file mode 100644 index 0000000000000000000000000000000000000000..2e9a5f1caa5982a9a340df831140a7202e4aeca1 --- /dev/null +++ b/artifacts/results/attempts/main/franka_M_repair_norm_shift_21/opus-5-5/B_T0_K0_cap10/a0/init_reference.npz @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:5f165e2433e09f4894d3cbde8b4ff23f8447cd2d92ebf954d538538930ff1ea3 +size 194148 diff --git a/artifacts/results/attempts/main/franka_M_repair_norm_shift_21/opus-5-5/B_T0_K0_cap10/a0/submitted.npz b/artifacts/results/attempts/main/franka_M_repair_norm_shift_21/opus-5-5/B_T0_K0_cap10/a0/submitted.npz new file mode 100644 index 0000000000000000000000000000000000000000..f542cf66fe56033adaa45eeec8beda5976ee44bb --- /dev/null +++ b/artifacts/results/attempts/main/franka_M_repair_norm_shift_21/opus-5-5/B_T0_K0_cap10/a0/submitted.npz @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:63b7062e0f0a2487244e5c5a0b3b12b47c7e60ef97e1a01e020bd2372eb3f557 +size 194396 diff --git a/artifacts/results/attempts/main/franka_M_repair_norm_shift_21/opus-5-5/B_T0_K0_cap10/blind0/init_reference.npz b/artifacts/results/attempts/main/franka_M_repair_norm_shift_21/opus-5-5/B_T0_K0_cap10/blind0/init_reference.npz new file mode 100644 index 0000000000000000000000000000000000000000..2e9a5f1caa5982a9a340df831140a7202e4aeca1 --- /dev/null +++ b/artifacts/results/attempts/main/franka_M_repair_norm_shift_21/opus-5-5/B_T0_K0_cap10/blind0/init_reference.npz @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:5f165e2433e09f4894d3cbde8b4ff23f8447cd2d92ebf954d538538930ff1ea3 +size 194148 diff --git a/artifacts/results/attempts/main/franka_M_repair_norm_shift_21/opus-5-5/B_T0_K0_cap10/blind0/submitted.npz b/artifacts/results/attempts/main/franka_M_repair_norm_shift_21/opus-5-5/B_T0_K0_cap10/blind0/submitted.npz new file mode 100644 index 0000000000000000000000000000000000000000..7fb6295b0b3093115e207fb48591cb6918a85bd8 --- /dev/null +++ b/artifacts/results/attempts/main/franka_M_repair_norm_shift_21/opus-5-5/B_T0_K0_cap10/blind0/submitted.npz @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ebcca5e3dba94685ad72fb457db86672e533b3d92a39223a31fbd206d3137a20 +size 194148 diff --git a/artifacts/results/attempts/main/franka_M_repair_swap_act_1_3/opus-5-5/A_T0_K0_cap10/a0/init_reference.npz b/artifacts/results/attempts/main/franka_M_repair_swap_act_1_3/opus-5-5/A_T0_K0_cap10/a0/init_reference.npz new file mode 100644 index 0000000000000000000000000000000000000000..1c6397a17778012909742e2d79d4aaf7d7a037ad --- /dev/null +++ b/artifacts/results/attempts/main/franka_M_repair_swap_act_1_3/opus-5-5/A_T0_K0_cap10/a0/init_reference.npz @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f0fb11211277f7b4974756cd24dbec18e9dbd7f3025e9de93a823e0c7b05202a +size 194148 diff --git a/artifacts/results/attempts/main/franka_M_repair_swap_act_1_3/opus-5-5/A_T0_K0_cap10/a0/submitted.npz b/artifacts/results/attempts/main/franka_M_repair_swap_act_1_3/opus-5-5/A_T0_K0_cap10/a0/submitted.npz new file mode 100644 index 0000000000000000000000000000000000000000..4119a38df604b312dcb363a7c9d67f9807b4671e --- /dev/null +++ b/artifacts/results/attempts/main/franka_M_repair_swap_act_1_3/opus-5-5/A_T0_K0_cap10/a0/submitted.npz @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:5522200d12881600bc2aa93dc9a6b8a761030d95198140190204ca83f6cd1efa +size 194396 diff --git a/artifacts/results/attempts/main/franka_M_repair_swap_act_1_3/opus-5-5/A_T0_K0_cap10/blind0/init_reference.npz b/artifacts/results/attempts/main/franka_M_repair_swap_act_1_3/opus-5-5/A_T0_K0_cap10/blind0/init_reference.npz new file mode 100644 index 0000000000000000000000000000000000000000..1c6397a17778012909742e2d79d4aaf7d7a037ad --- /dev/null +++ b/artifacts/results/attempts/main/franka_M_repair_swap_act_1_3/opus-5-5/A_T0_K0_cap10/blind0/init_reference.npz @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f0fb11211277f7b4974756cd24dbec18e9dbd7f3025e9de93a823e0c7b05202a +size 194148 diff --git a/artifacts/results/attempts/main/franka_M_repair_swap_act_1_3/opus-5-5/A_T0_K0_cap10/blind0/submitted.npz b/artifacts/results/attempts/main/franka_M_repair_swap_act_1_3/opus-5-5/A_T0_K0_cap10/blind0/submitted.npz new file mode 100644 index 0000000000000000000000000000000000000000..2216cddd0fb172d6e74b31fb95f6b49107ff7b75 --- /dev/null +++ b/artifacts/results/attempts/main/franka_M_repair_swap_act_1_3/opus-5-5/A_T0_K0_cap10/blind0/submitted.npz @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e84847ee59e93cfeec4f68273f72ab677bdec4db76f1a3ff06117c6db5bdec4e +size 194396 diff --git a/artifacts/results/attempts/main/franka_M_repair_swap_act_1_3/opus-5-5/A_T2_K5_cap10/a0/init_reference.npz b/artifacts/results/attempts/main/franka_M_repair_swap_act_1_3/opus-5-5/A_T2_K5_cap10/a0/init_reference.npz new file mode 100644 index 0000000000000000000000000000000000000000..1c6397a17778012909742e2d79d4aaf7d7a037ad --- /dev/null +++ b/artifacts/results/attempts/main/franka_M_repair_swap_act_1_3/opus-5-5/A_T2_K5_cap10/a0/init_reference.npz @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f0fb11211277f7b4974756cd24dbec18e9dbd7f3025e9de93a823e0c7b05202a +size 194148 diff --git a/artifacts/results/attempts/main/franka_M_repair_swap_act_1_3/opus-5-5/A_T2_K5_cap10/a0/submitted.npz b/artifacts/results/attempts/main/franka_M_repair_swap_act_1_3/opus-5-5/A_T2_K5_cap10/a0/submitted.npz new file mode 100644 index 0000000000000000000000000000000000000000..a391b7fc58a89c90aa6fa7038908dadeb216c409 --- /dev/null +++ b/artifacts/results/attempts/main/franka_M_repair_swap_act_1_3/opus-5-5/A_T2_K5_cap10/a0/submitted.npz @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:353f84b62e0dbd1bc10d6fa97d39db20f195fe968f30aa30d1e13e8f32b445e8 +size 194148 diff --git a/artifacts/results/attempts/main/franka_M_repair_swap_act_1_3/opus-5-5/A_T2_K5_cap10/blind0/init_reference.npz b/artifacts/results/attempts/main/franka_M_repair_swap_act_1_3/opus-5-5/A_T2_K5_cap10/blind0/init_reference.npz new file mode 100644 index 0000000000000000000000000000000000000000..1c6397a17778012909742e2d79d4aaf7d7a037ad --- /dev/null +++ b/artifacts/results/attempts/main/franka_M_repair_swap_act_1_3/opus-5-5/A_T2_K5_cap10/blind0/init_reference.npz @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f0fb11211277f7b4974756cd24dbec18e9dbd7f3025e9de93a823e0c7b05202a +size 194148 diff --git a/artifacts/results/attempts/main/franka_M_repair_swap_act_1_3/opus-5-5/A_T2_K5_cap10/blind0/submitted.npz b/artifacts/results/attempts/main/franka_M_repair_swap_act_1_3/opus-5-5/A_T2_K5_cap10/blind0/submitted.npz new file mode 100644 index 0000000000000000000000000000000000000000..a391b7fc58a89c90aa6fa7038908dadeb216c409 --- /dev/null +++ b/artifacts/results/attempts/main/franka_M_repair_swap_act_1_3/opus-5-5/A_T2_K5_cap10/blind0/submitted.npz @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:353f84b62e0dbd1bc10d6fa97d39db20f195fe968f30aa30d1e13e8f32b445e8 +size 194148 diff --git a/artifacts/results/attempts/main/franka_M_repair_swap_act_1_3/opus-5-5/B_T0_K0_cap10/a0/init_reference.npz b/artifacts/results/attempts/main/franka_M_repair_swap_act_1_3/opus-5-5/B_T0_K0_cap10/a0/init_reference.npz new file mode 100644 index 0000000000000000000000000000000000000000..1c6397a17778012909742e2d79d4aaf7d7a037ad --- /dev/null +++ b/artifacts/results/attempts/main/franka_M_repair_swap_act_1_3/opus-5-5/B_T0_K0_cap10/a0/init_reference.npz @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f0fb11211277f7b4974756cd24dbec18e9dbd7f3025e9de93a823e0c7b05202a +size 194148 diff --git a/artifacts/results/attempts/main/franka_M_repair_swap_act_1_3/opus-5-5/B_T0_K0_cap10/a0/submitted.npz b/artifacts/results/attempts/main/franka_M_repair_swap_act_1_3/opus-5-5/B_T0_K0_cap10/a0/submitted.npz new file mode 100644 index 0000000000000000000000000000000000000000..4119a38df604b312dcb363a7c9d67f9807b4671e --- /dev/null +++ b/artifacts/results/attempts/main/franka_M_repair_swap_act_1_3/opus-5-5/B_T0_K0_cap10/a0/submitted.npz @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:5522200d12881600bc2aa93dc9a6b8a761030d95198140190204ca83f6cd1efa +size 194396 diff --git a/artifacts/results/attempts/main/franka_M_repair_swap_act_1_3/opus-5-5/B_T0_K0_cap10/blind0/init_reference.npz b/artifacts/results/attempts/main/franka_M_repair_swap_act_1_3/opus-5-5/B_T0_K0_cap10/blind0/init_reference.npz new file mode 100644 index 0000000000000000000000000000000000000000..1c6397a17778012909742e2d79d4aaf7d7a037ad --- /dev/null +++ b/artifacts/results/attempts/main/franka_M_repair_swap_act_1_3/opus-5-5/B_T0_K0_cap10/blind0/init_reference.npz @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f0fb11211277f7b4974756cd24dbec18e9dbd7f3025e9de93a823e0c7b05202a +size 194148 diff --git a/artifacts/results/attempts/main/franka_M_repair_swap_act_1_3/opus-5-5/B_T0_K0_cap10/blind0/submitted.npz b/artifacts/results/attempts/main/franka_M_repair_swap_act_1_3/opus-5-5/B_T0_K0_cap10/blind0/submitted.npz new file mode 100644 index 0000000000000000000000000000000000000000..e8abe8f30f6211f397139257b6862b5574f5dd01 --- /dev/null +++ b/artifacts/results/attempts/main/franka_M_repair_swap_act_1_3/opus-5-5/B_T0_K0_cap10/blind0/submitted.npz @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3775df16902276e289c071f1dbbfe1f04c57c978759ff67a273d9023a9e25734 +size 194396 diff --git a/artifacts/results/attempts/main/franka_XS_write/opus-5-5/A_T0_K0/a0/submitted.npz b/artifacts/results/attempts/main/franka_XS_write/opus-5-5/A_T0_K0/a0/submitted.npz new file mode 100644 index 0000000000000000000000000000000000000000..a8ce4e12ef044e26344893262270252a4457cd2a --- /dev/null +++ b/artifacts/results/attempts/main/franka_XS_write/opus-5-5/A_T0_K0/a0/submitted.npz @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e89c89dd8fc66e0c92ab5d7222c594e22157b3a5f02c468ecd4299ea658acb17 +size 27268 diff --git a/artifacts/results/attempts/main/franka_XS_write/opus-5-5/A_T2_K5/a0/init_reference.npz b/artifacts/results/attempts/main/franka_XS_write/opus-5-5/A_T2_K5/a0/init_reference.npz new file mode 100644 index 0000000000000000000000000000000000000000..e4f33b4883395e888f06bccf8b23803d6593903e --- /dev/null +++ b/artifacts/results/attempts/main/franka_XS_write/opus-5-5/A_T2_K5/a0/init_reference.npz @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4fe489675c7f488ab1abae8d14c4ddfd6933ea28ab66f9e79b3540f6df86c6eb +size 27268 diff --git a/artifacts/results/attempts/main/franka_XS_write/opus-5-5/A_T2_K5/a0/submitted.npz b/artifacts/results/attempts/main/franka_XS_write/opus-5-5/A_T2_K5/a0/submitted.npz new file mode 100644 index 0000000000000000000000000000000000000000..a7db69b738af4e7521f3f6296423057f9ed1e4ae --- /dev/null +++ b/artifacts/results/attempts/main/franka_XS_write/opus-5-5/A_T2_K5/a0/submitted.npz @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:11cbf2c271810b25397769d8d1e19455ef7e62b789a90234a15803ae51a83321 +size 27620 diff --git a/metadata/THIRD_PARTY_LICENSES/IsaacLab-BSD-3-Clause.txt b/metadata/THIRD_PARTY_LICENSES/IsaacLab-BSD-3-Clause.txt new file mode 100644 index 0000000000000000000000000000000000000000..bc7fc124526bc60793bc387571dd6bc97294708a --- /dev/null +++ b/metadata/THIRD_PARTY_LICENSES/IsaacLab-BSD-3-Clause.txt @@ -0,0 +1,32 @@ +Please see README.md for additional licensing information. + +Copyright (c) 2022-2026, The Isaac Lab Project Developers (https://github.com/isaac-sim/IsaacLab/blob/main/CONTRIBUTORS.md). + +All rights reserved. + +SPDX-License-Identifier: BSD-3-Clause + +Redistribution and use in source and binary forms, with or without modification, +are permitted provided that the following conditions are met: + +1. Redistributions of source code must retain the above copyright notice, + this list of conditions and the following disclaimer. + +2. Redistributions in binary form must reproduce the above copyright notice, + this list of conditions and the following disclaimer in the documentation + and/or other materials provided with the distribution. + +3. Neither the name of the copyright holder nor the names of its contributors + may be used to endorse or promote products derived from this software without + specific prior written permission. + +THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS" AND +ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED +WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE +DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT HOLDER OR CONTRIBUTORS BE LIABLE FOR +ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES +(INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; +LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND +ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT +(INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS +SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. diff --git a/metadata/THIRD_PARTY_LICENSES/LIBERO-MIT.txt b/metadata/THIRD_PARTY_LICENSES/LIBERO-MIT.txt new file mode 100644 index 0000000000000000000000000000000000000000..49b61c0f9209a5dfd3ad5a5d181eff7d5da153ae --- /dev/null +++ b/metadata/THIRD_PARTY_LICENSES/LIBERO-MIT.txt @@ -0,0 +1,21 @@ +MIT License + +Copyright (c) 2023 Lifelong Robot Learning + +Permission is hereby granted, free of charge, to any person obtaining a copy +of this software and associated documentation files (the "Software"), to deal +in the Software without restriction, including without limitation the rights +to use, copy, modify, merge, publish, distribute, sublicense, and/or sell +copies of the Software, and to permit persons to whom the Software is +furnished to do so, subject to the following conditions: + +The above copyright notice and this permission notice shall be included in all +copies or substantial portions of the Software. + +THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR +IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, +FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE +AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER +LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, +OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE +SOFTWARE. diff --git a/metadata/THIRD_PARTY_LICENSES/OpenVLA-MIT.txt b/metadata/THIRD_PARTY_LICENSES/OpenVLA-MIT.txt new file mode 100644 index 0000000000000000000000000000000000000000..04f26a7d0bfdeb89ce54681c695a40df21aa6d96 --- /dev/null +++ b/metadata/THIRD_PARTY_LICENSES/OpenVLA-MIT.txt @@ -0,0 +1,21 @@ +MIT License + +Copyright (c) 2024 Moo Jin Kim, Karl Pertsch, Siddharth Karamcheti. + +Permission is hereby granted, free of charge, to any person obtaining a copy +of this software and associated documentation files (the "Software"), to deal +in the Software without restriction, including without limitation the rights +to use, copy, modify, merge, publish, distribute, sublicense, and/or sell +copies of the Software, and to permit persons to whom the Software is +furnished to do so, subject to the following conditions: + +The above copyright notice and this permission notice shall be included in all +copies or substantial portions of the Software. + +THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR +IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, +FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE +AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER +LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, +OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE +SOFTWARE. diff --git a/metadata/data/manifest.json b/metadata/data/manifest.json new file mode 100644 index 0000000000000000000000000000000000000000..4426b3683bab39876dd4199ef8eba71c777024ef --- /dev/null +++ b/metadata/data/manifest.json @@ -0,0 +1,682 @@ +{ + "data/openvla/attempts/fable-5-1_single.json": { + "bytes": 24406, + "sha256": "b0505374db734458100e4f346f2ab3fbb544feb11746cd338532a87605ffc5e8" + }, + "data/openvla/attempts/fable-5-1_three.json": { + "bytes": 23634, + "sha256": "930ca8729a509553f69343ee6fc140358d64d4418101509d2b3bef81b3e5ea83" + }, + "data/openvla/attempts/opus-5-5_single.json": { + "bytes": 22996, + "sha256": "bb7732fafa60d1762ebcc70b5834830ee5e7dd5074b23d5a7875285f4dedaa02" + }, + "data/openvla/attempts/opus-5-5_three.json": { + "bytes": 22847, + "sha256": "8e346132702938dc81c7120926c3c8824f45b28268ce4d806ef1ecde99b443ae" + }, + "data/openvla/checkpoint_manifest.json": { + "bytes": 1607, + "sha256": "7697c894ca78067876b3ae48b9cf29bbe8338fbfcff7a456d4e8fb9ce28f5003" + }, + "data/openvla/patches/fable-5-1_single.npz": { + "bytes": 50206, + "sha256": "928557ceaf67ab2635f230ee52b77e0372bd2abe2658a25fa52486880ee63ca5" + }, + "data/openvla/patches/fable-5-1_three.npz": { + "bytes": 50206, + "sha256": "886cbc79f4cfa23b6b9bfceafaaf9ceef0fa351f3ac88b5e17de1bf13681c046" + }, + "data/openvla/patches/opus-5-5_single.npz": { + "bytes": 50206, + "sha256": "bf5bd8eee9e056055bf920c2391541bfd57c402631da9fc966ee81a4235889ad" + }, + "data/openvla/patches/opus-5-5_three.npz": { + "bytes": 278, + "sha256": "4309dc07ccd8e2195f7bb2f32a1e572cf22886036fa8ee6c52385fc0fb5b9eed" + }, + "data/openvla/protocol.json": { + "bytes": 3472, + "sha256": "ce70a5a1d86de991ce258046dbebec1f8814c5ac700e4381e0f96ba706d7f9f4" + }, + "data/openvla/runtime.json": { + "bytes": 759, + "sha256": "249ad0a9bc624b87c9529ae6392946e6a7574006c730cad4200db86f2359566b" + }, + "data/openvla/summary.json": { + "bytes": 118949, + "sha256": "c2d371d6f8e6eb9845edc87ae8b810312a63f373684303a2156e34a1b708d6c6" + }, + "data/openvla/targets.json": { + "bytes": 66, + "sha256": "caf498fd7a772256c13c38ba3bf5893c5f99b1cd4f89a08f25a6a4a965a8e4a8" + }, + "data/robot_policy/attempts/hands/allegro_L_repair_early_ep200/fable-5-1/A_T2_K5_cap10/a0/attempt.json": { + "bytes": 966, + "sha256": "eca95f106c0bf02fc052766d0d48bc578330c3b539d5c54c4f9979379a286fb6" + }, + "data/robot_policy/attempts/hands/allegro_L_repair_early_ep200/fable-5-1/A_T2_K5_cap10/a0/audit_A.json": { + "bytes": 2653, + "sha256": "166e011fc72a29abaf02591e7fc8afa18ed52dd5b219063705a2068cef4f75e0" + }, + "data/robot_policy/attempts/hands/allegro_L_repair_early_ep200/fable-5-1/A_T2_K5_cap10/a0/audit_B.json": { + "bytes": 3026, + "sha256": "0e14391f96f1f8f8552e63063f2173b76d58c6c685010fb1fd97b348f903f28a" + }, + "data/robot_policy/attempts/hands/allegro_L_repair_early_ep200/opus-5-5/A_T2_K5_cap10/a0/attempt.json": { + "bytes": 978, + "sha256": "a6875290c7d3683171a4430fd81897bcc07c0b4795cdec3234f72e19abdac14d" + }, + "data/robot_policy/attempts/hands/allegro_L_repair_early_ep200/opus-5-5/A_T2_K5_cap10/a0/audit_A.json": { + "bytes": 2484, + "sha256": "30c11335c704c7a56d78d3a760ed7fe44a47ff0d8dfbc9998f3f5b803252647f" + }, + "data/robot_policy/attempts/hands/allegro_L_repair_early_ep200/opus-5-5/A_T2_K5_cap10/a0/audit_B.json": { + "bytes": 2671, + "sha256": "e0b61db4980638be408060a42da4485cd6bd855033578fad51248f05ba7dd1a5" + }, + "data/robot_policy/attempts/hands/allegro_L_repair_early_ep50/fable-5-1/A_T2_K5_cap10/a0/attempt.json": { + "bytes": 980, + "sha256": "c7022a6e2596afe7aa13a3b9bb3d9929b9b49e9067031712232cf2ef2d81fd32" + }, + "data/robot_policy/attempts/hands/allegro_L_repair_early_ep50/fable-5-1/A_T2_K5_cap10/a0/audit_A.json": { + "bytes": 2586, + "sha256": "f992e616d49989d635ec0084a68e1dd334b63f8ba0e6648a8057848e59d8bb83" + }, + "data/robot_policy/attempts/hands/allegro_L_repair_early_ep50/fable-5-1/A_T2_K5_cap10/a0/audit_B.json": { + "bytes": 2766, + "sha256": "8c21f9e4a90fb4d0fbfebd521340d32f3d13aebaa83aaa02b34190633343f4b3" + }, + "data/robot_policy/attempts/hands/allegro_L_repair_early_ep50/opus-5-5/A_T2_K5_cap10/a0/attempt.json": { + "bytes": 980, + "sha256": "ab50dbf8391c04606510ee59cf8e1f7408121944cee19972152fb3cebb1d6103" + }, + "data/robot_policy/attempts/hands/allegro_L_repair_early_ep50/opus-5-5/A_T2_K5_cap10/a0/audit_A.json": { + "bytes": 2323, + "sha256": "253b7e3854339d42ef50f02b5ee3d9ab099e89aa029a506a3e74a1dc8c1268d7" + }, + "data/robot_policy/attempts/hands/allegro_L_repair_early_ep50/opus-5-5/A_T2_K5_cap10/a0/audit_B.json": { + "bytes": 2261, + "sha256": "56c92c7e59e0b620a24e83a0030485aa8b9299e225e62e267e35ec9d694b36b1" + }, + "data/robot_policy/attempts/hands/allegro_L_repair_norm_shift_53/fable-5-1/A_T2_K5_cap10/a0/attempt.json": { + "bytes": 988, + "sha256": "4131822e381d592c8b6361de49aeecfe577b57a5c66f5bc718f8aeb2c837fb93" + }, + "data/robot_policy/attempts/hands/allegro_L_repair_norm_shift_53/fable-5-1/A_T2_K5_cap10/a0/audit_A.json": { + "bytes": 1660, + "sha256": "77d5a275f912acc3aa40a317b08b194ba0b40da241bb20b845db5234f89508b2" + }, + "data/robot_policy/attempts/hands/allegro_L_repair_norm_shift_53/fable-5-1/A_T2_K5_cap10/a0/audit_B.json": { + "bytes": 1592, + "sha256": "52b389b9dc9da30eebe66717215b3008c0d41df55831bf86bcb57328474ab3bc" + }, + "data/robot_policy/attempts/hands/allegro_L_repair_norm_shift_53/fable-5-1/A_T2_K5_cap10/blind0/attempt.json": { + "bytes": 992, + "sha256": "4afc813df07d4d8786adfdd05e0fa48bc9a4b08567f2ef91a05e919ac4576a21" + }, + "data/robot_policy/attempts/hands/allegro_L_repair_norm_shift_53/fable-5-1/A_T2_K5_cap10/blind0/audit_A.json": { + "bytes": 1868, + "sha256": "b49e48b8ace313461f768fbb1e98d62a59820fa161f0aba69d31d98989cf0f8d" + }, + "data/robot_policy/attempts/hands/allegro_L_repair_norm_shift_53/fable-5-1/A_T2_K5_cap10/blind0/audit_B.json": { + "bytes": 1514, + "sha256": "a7c295c697570f19e9a6f8cadbde717d64560e04c265e63bdde2ed7740b5754c" + }, + "data/robot_policy/attempts/hands/allegro_L_repair_swap_act_4_7/opus-5-5/A_T2_K5_cap10/a0/attempt.json": { + "bytes": 986, + "sha256": "c109754022874649e1cb505244d109c121a964dbe8b51fc5500a206d72add065" + }, + "data/robot_policy/attempts/hands/allegro_L_repair_swap_act_4_7/opus-5-5/A_T2_K5_cap10/a0/audit_A.json": { + "bytes": 1536, + "sha256": "46e50409fcc7797edf4f6b51f806e0a6cbd40de95c0e38baaad406d0be7b0936" + }, + "data/robot_policy/attempts/hands/allegro_L_repair_swap_act_4_7/opus-5-5/A_T2_K5_cap10/a0/audit_B.json": { + "bytes": 1408, + "sha256": "82617244226202ba811edea5f26b13d30d9541e0fcd8834999d68334ae5bbddc" + }, + "data/robot_policy/attempts/hands/allegro_XS_write/fable-5-1/A_T2_K5/a0/attempt.json": { + "bytes": 943, + "sha256": "72e5eb01ac4442f36cca0876c1c3494aa50657f8c501ff42f2d6f34356ed26ce" + }, + "data/robot_policy/attempts/hands/allegro_XS_write/fable-5-1/A_T2_K5/a0/audit_A.json": { + "bytes": 1574, + "sha256": "821a50bf5a8b4dbc5d8bd28b48325f46f0752c7fa357f0d18e9e1aac39f92a0d" + }, + "data/robot_policy/attempts/hands/allegro_XS_write/fable-5-1/A_T2_K5/a0/audit_B.json": { + "bytes": 1699, + "sha256": "43b5e53d27aab7b1329a6cdd9fbb58d643bf733aba71b4e67e3b952af7ee92f7" + }, + "data/robot_policy/attempts/hands/allegro_XS_write/fable-5-1/B_T2_K5/a0/attempt.json": { + "bytes": 947, + "sha256": "7fa2a72f599282cbd649ef7d481ae783ae498fed1688d560e61235c659a6cbd6" + }, + "data/robot_policy/attempts/hands/allegro_XS_write/fable-5-1/B_T2_K5/a0/audit_A.json": { + "bytes": 2475, + "sha256": "7e445343ad23b3fe1fd953564983f2c98e1988b16af846f87701dff5b21c2974" + }, + "data/robot_policy/attempts/hands/allegro_XS_write/fable-5-1/B_T2_K5/a0/audit_B.json": { + "bytes": 2391, + "sha256": "ea330698709f9ec26aee57135afd9fee84dbad1334c1ae16f6b7af7808d3ffe6" + }, + "data/robot_policy/attempts/hands/allegro_XS_write/opus-5-5/A_T2_K5/a0/attempt.json": { + "bytes": 945, + "sha256": "409231e4e9b8011f97c2b11584f09a551a5581a1b95e35b0fa02cde8b335d8ef" + }, + "data/robot_policy/attempts/hands/allegro_XS_write/opus-5-5/A_T2_K5/a0/audit_A.json": { + "bytes": 2515, + "sha256": "0ac2ba6dd56e7b57a747594cbcb786a60153b1ee42d40aedffcb1cfa06c186c2" + }, + "data/robot_policy/attempts/hands/allegro_XS_write/opus-5-5/A_T2_K5/a0/audit_B.json": { + "bytes": 2298, + "sha256": "51310e89125ebb964eab496c46d44ee671edbd8badeaadb131b5bcf5667e8aa7" + }, + "data/robot_policy/attempts/hands/allegro_XS_write/opus-5-5/B_T2_K5/a0/attempt.json": { + "bytes": 945, + "sha256": "eb907f5e54e94d4bb6c9ddbc53e21a0625d0b731985724f8609a3845275198fd" + }, + "data/robot_policy/attempts/hands/allegro_XS_write/opus-5-5/B_T2_K5/a0/audit_A.json": { + "bytes": 3196, + "sha256": "5883818bf12cd3e0f2cc11b621064e1bccf003a402e89fc7d18206fcb0d248e3" + }, + "data/robot_policy/attempts/hands/allegro_XS_write/opus-5-5/B_T2_K5/a0/audit_B.json": { + "bytes": 2241, + "sha256": "d58d601fc90e2f19cbe64ba7fcb53b7fd055ea3cbf4e1c40e429a188cbab5b57" + }, + "data/robot_policy/attempts/hands/shadow_L_repair_early_ep100/fable-5-1/A_T2_K5_cap10/a0/attempt.json": { + "bytes": 979, + "sha256": "def8b8eb2381737993ad3f17b7af57d31440c3b5c41aa6c5c0bc255c062c4d6d" + }, + "data/robot_policy/attempts/hands/shadow_L_repair_early_ep100/fable-5-1/A_T2_K5_cap10/a0/audit_A.json": { + "bytes": 3012, + "sha256": "6b6f6e8f9c326bdfed365c86e4304401d63154d6bc2b11f9723283a66d179850" + }, + "data/robot_policy/attempts/hands/shadow_L_repair_early_ep100/fable-5-1/A_T2_K5_cap10/a0/audit_B.json": { + "bytes": 3225, + "sha256": "4f5f5a5565f714611485a1457fc41cbc662b6e1fcc1c917ec5eef21eab378b11" + }, + "data/robot_policy/attempts/hands/shadow_L_repair_early_ep100/opus-5-5/A_T2_K5_cap10/a0/attempt.json": { + "bytes": 980, + "sha256": "4a1ec86ef3738707caebd537f95f86435efd9c9562aad7a7e06fcd8bb0af8b87" + }, + "data/robot_policy/attempts/hands/shadow_L_repair_early_ep100/opus-5-5/A_T2_K5_cap10/a0/audit_A.json": { + "bytes": 2673, + "sha256": "eef0252e91efb3ec19d5d738f4972cb949ca2aa87a7adbd0f7a6b8beb4ce2421" + }, + "data/robot_policy/attempts/hands/shadow_L_repair_early_ep100/opus-5-5/A_T2_K5_cap10/a0/audit_B.json": { + "bytes": 2932, + "sha256": "468d4e4a59821df0109839515ea25ec18257aea2e688352db4904fdf775a4f8f" + }, + "data/robot_policy/attempts/hands/shadow_L_repair_early_ep200/fable-5-1/A_T2_K5_cap10/a0/attempt.json": { + "bytes": 980, + "sha256": "d4cb55520c27ecbc8fd357aa94caaf33a61107d36e793352893cd878b300366b" + }, + "data/robot_policy/attempts/hands/shadow_L_repair_early_ep200/fable-5-1/A_T2_K5_cap10/a0/audit_A.json": { + "bytes": 2721, + "sha256": "ec5c300ff359910f2eccfb2d4ca6507138d133ff9da9b12f4f1cdbed403e73e1" + }, + "data/robot_policy/attempts/hands/shadow_L_repair_early_ep200/fable-5-1/A_T2_K5_cap10/a0/audit_B.json": { + "bytes": 2501, + "sha256": "55c15c65da51e227178638a46f1c21d9579f99c5655c55e1a844fae58e19ec13" + }, + "data/robot_policy/attempts/hands/shadow_L_repair_early_ep200/opus-5-5/A_T2_K5_cap10/a0/attempt.json": { + "bytes": 977, + "sha256": "f3ab999de28bbee5bf131b4586e9089f3bcd212fb3a194e3fa9183fab0074a99" + }, + "data/robot_policy/attempts/hands/shadow_L_repair_early_ep200/opus-5-5/A_T2_K5_cap10/a0/audit_A.json": { + "bytes": 2060, + "sha256": "703d9240bd8b482b5204254585fc154b596563af17d2693a9f7417fbfc22007e" + }, + "data/robot_policy/attempts/hands/shadow_L_repair_early_ep200/opus-5-5/A_T2_K5_cap10/a0/audit_B.json": { + "bytes": 2499, + "sha256": "000014f5b9ba3a244fece8223590daae1b271269bfe9693988bd0a7e03734d6b" + }, + "data/robot_policy/attempts/hands/shadow_L_repair_norm_shift_69/fable-5-1/A_T2_K5_cap10/a0/attempt.json": { + "bytes": 985, + "sha256": "f0b35f64c4a4d84d3840c5b5ef7a4a17ddc92abb80f4e63e3b691982e59e9c3c" + }, + "data/robot_policy/attempts/hands/shadow_L_repair_norm_shift_69/fable-5-1/A_T2_K5_cap10/a0/audit_A.json": { + "bytes": 1560, + "sha256": "78b0109b986b361ff2d66d7efb3a8a33090f4d21245a7f976c2a9e25ada5d041" + }, + "data/robot_policy/attempts/hands/shadow_L_repair_norm_shift_69/fable-5-1/A_T2_K5_cap10/a0/audit_B.json": { + "bytes": 1557, + "sha256": "175ccc5c48076ead218566efddaa860f13aab9dfbbb86df872a3b96e861f81d9" + }, + "data/robot_policy/attempts/hands/shadow_L_repair_norm_shift_69/fable-5-1/A_T2_K5_cap10/blind0/attempt.json": { + "bytes": 989, + "sha256": "717298b774170071480bea5957731fea03f6d36549e89c91564e261ae3f530f1" + }, + "data/robot_policy/attempts/hands/shadow_L_repair_norm_shift_69/fable-5-1/A_T2_K5_cap10/blind0/audit_A.json": { + "bytes": 1620, + "sha256": "1c460b1e363141f6fc9102658a3b479e68569d21bc6ecd0af821f58e6deec6d5" + }, + "data/robot_policy/attempts/hands/shadow_L_repair_norm_shift_69/fable-5-1/A_T2_K5_cap10/blind0/audit_B.json": { + "bytes": 1509, + "sha256": "c7027b58887ae0159e7924a2311e041d53b49ab05309f5b71fda1bffb5e9190a" + }, + "data/robot_policy/attempts/hands/shadow_L_repair_swap_act_7_11/opus-5-5/A_T2_K5_cap10/a0/attempt.json": { + "bytes": 975, + "sha256": "c69d4cf680f34ee8bcd8e33095cd4fb18beab07e2c312ad3a0d1d90b996795bd" + }, + "data/robot_policy/attempts/hands/shadow_L_repair_swap_act_7_11/opus-5-5/A_T2_K5_cap10/a0/audit_A.json": { + "bytes": 1430, + "sha256": "d2d7357a598138d17fc437402a0bd3afd5b05ae158ab69356580c23f3cb45a89" + }, + "data/robot_policy/attempts/hands/shadow_L_repair_swap_act_7_11/opus-5-5/A_T2_K5_cap10/a0/audit_B.json": { + "bytes": 1588, + "sha256": "054167de4d71e68d935057748c1ce0b5b042d07f7956b3193aa554c1f43ab5c1" + }, + "data/robot_policy/attempts/hands/shadow_XS_write/fable-5-1/A_T2_K5/a0/attempt.json": { + "bytes": 948, + "sha256": "3b380ca6eec2cf1be30551525e536118ea6f0204c4da07de5b00ad78345d4b7a" + }, + "data/robot_policy/attempts/hands/shadow_XS_write/fable-5-1/A_T2_K5/a0/audit_A.json": { + "bytes": 2402, + "sha256": "49bf10ce8fe2642a703cd12da6119c5e8fcc47feb4d425cd78356f43adcba13b" + }, + "data/robot_policy/attempts/hands/shadow_XS_write/fable-5-1/A_T2_K5/a0/audit_B.json": { + "bytes": 2372, + "sha256": "e6b03ca346e4c76002e5ceaa34dc9fd689ab235f778adee7bc1302e2a657dc22" + }, + "data/robot_policy/attempts/hands/shadow_XS_write/fable-5-1/B_T2_K5/a0/attempt.json": { + "bytes": 944, + "sha256": "f9ec6152c980f38ddcd82022962355331fa11d3842ebf1436900fb5bae6b4626" + }, + "data/robot_policy/attempts/hands/shadow_XS_write/fable-5-1/B_T2_K5/a0/audit_A.json": { + "bytes": 2449, + "sha256": "9e129b4a2a103e54fa32b5c892103d9c1b4cc88823018d30f9792eb9a67b73c0" + }, + "data/robot_policy/attempts/hands/shadow_XS_write/fable-5-1/B_T2_K5/a0/audit_B.json": { + "bytes": 2735, + "sha256": "a312386e108a6a2d46fa116d5bb4538303206bf737ceb4aada4bd36468a29407" + }, + "data/robot_policy/attempts/hands/shadow_XS_write/opus-5-5/A_T2_K5/a0/attempt.json": { + "bytes": 943, + "sha256": "f61b31ed7c2855b0d614675ff3afd915eb30675dcda797821f1018168565fef9" + }, + "data/robot_policy/attempts/hands/shadow_XS_write/opus-5-5/A_T2_K5/a0/audit_A.json": { + "bytes": 2779, + "sha256": "76d26085e43469bdddae0264260f0e95f71d6c1d4b61444177753026f9b618a5" + }, + "data/robot_policy/attempts/hands/shadow_XS_write/opus-5-5/A_T2_K5/a0/audit_B.json": { + "bytes": 2817, + "sha256": "e6886fd26aa686c70cc3d497673e7118cced1345c6c4b1c43bb9a304daf3085c" + }, + "data/robot_policy/attempts/hands/shadow_XS_write/opus-5-5/B_T2_K5/a0/attempt.json": { + "bytes": 946, + "sha256": "479a398ca5d8d9a59cf2e2705bc331627be29fa34aa933514ef1d00cb3fecbed" + }, + "data/robot_policy/attempts/hands/shadow_XS_write/opus-5-5/B_T2_K5/a0/audit_A.json": { + "bytes": 3816, + "sha256": "467a85c65a9d095d146928bd65e445bfc41d71c12a1b78c9192c8d9d04c88a89" + }, + "data/robot_policy/attempts/hands/shadow_XS_write/opus-5-5/B_T2_K5/a0/audit_B.json": { + "bytes": 2373, + "sha256": "84f9b04768abf7f83d9cf38bfa5f81b34f75d17af330c1a702d8de6253f05aa3" + }, + "data/robot_policy/attempts/main/franka_M_repair_collapse_ep500/opus-5-5/A_T0_K0_cap10/a0/attempt.json": { + "bytes": 775, + "sha256": "aa5ac2f922d09ed30603cdf8b38d3cf0a68ba7fa7eea8d7d980637fd70df7d91" + }, + "data/robot_policy/attempts/main/franka_M_repair_collapse_ep500/opus-5-5/A_T0_K0_cap10/a0/audit_A.json": { + "bytes": 3307, + "sha256": "180628930f962737701b8406ec4a157e85f14439ece8f0d58e51807ff7c8bacc" + }, + "data/robot_policy/attempts/main/franka_M_repair_collapse_ep500/opus-5-5/A_T0_K0_cap10/a0/audit_B.json": { + "bytes": 2687, + "sha256": "49b09fd4c43b43db6eb35d6cdd90869591e20a5fcd970fac5a927b0edaaa83f8" + }, + "data/robot_policy/attempts/main/franka_M_repair_collapse_ep500/opus-5-5/A_T0_K0_cap10/a0/work/answer_program.py": { + "bytes": 5588, + "sha256": "7805e07a7ad4ad687c8969acd92e4be791e94934ba9441f0c73b4a7cc1949d6b" + }, + "data/robot_policy/attempts/main/franka_M_repair_collapse_ep500/opus-5-5/A_T2_K5_cap10/a0/attempt.json": { + "bytes": 760, + "sha256": "fe3fbf291b28b2f7c01fbd03e18e67a7054c2a71657f0e17500639712319df62" + }, + "data/robot_policy/attempts/main/franka_M_repair_collapse_ep500/opus-5-5/A_T2_K5_cap10/a0/audit_A.json": { + "bytes": 2716, + "sha256": "eaebc89bc7bb9bd5546ead6e3a18154b3203f28cfb0848550dcd8c52e45e375e" + }, + "data/robot_policy/attempts/main/franka_M_repair_collapse_ep500/opus-5-5/A_T2_K5_cap10/a0/audit_B.json": { + "bytes": 2447, + "sha256": "e3491a1684c4402ec6ba6525158d0418b91392c7ea70a4817e95bc305bdb2ad8" + }, + "data/robot_policy/attempts/main/franka_M_repair_early_ep15/opus-5-5/A_T0_K0_cap10/a0/attempt.json": { + "bytes": 758, + "sha256": "237de95514fae8712f8ab6639eb846c81680e4837527ca70d4ee482870dbe597" + }, + "data/robot_policy/attempts/main/franka_M_repair_early_ep15/opus-5-5/A_T0_K0_cap10/a0/audit_A.json": { + "bytes": 2545, + "sha256": "e5e507ca9f7c565cc1f846dc1400378e84b35006262b3f7ec1ff1a695ee1727d" + }, + "data/robot_policy/attempts/main/franka_M_repair_early_ep15/opus-5-5/A_T0_K0_cap10/a0/audit_B.json": { + "bytes": 2862, + "sha256": "07ed32de372b5f765e7600095d121a0d2a2cc0aceb5959a5782e2138d801fbdd" + }, + "data/robot_policy/attempts/main/franka_M_repair_early_ep15/opus-5-5/A_T0_K0_cap10/a0/work/answer_program.py": { + "bytes": 4581, + "sha256": "000cfc00958bce893f63687e7632725662e995569dd3431548d862ca43729c71" + }, + "data/robot_policy/attempts/main/franka_M_repair_early_ep15/opus-5-5/A_T2_K5_cap10/a0/attempt.json": { + "bytes": 772, + "sha256": "528c4dfcf1bfdb2e87cb7103208c067c7fbd6116cc988538408363de2dc93aac" + }, + "data/robot_policy/attempts/main/franka_M_repair_early_ep15/opus-5-5/A_T2_K5_cap10/a0/audit_A.json": { + "bytes": 3933, + "sha256": "edbe1eee374d5264a68539697d8f35f5b8bf37b65463aa5d58b868ea6659df0a" + }, + "data/robot_policy/attempts/main/franka_M_repair_early_ep15/opus-5-5/A_T2_K5_cap10/a0/audit_B.json": { + "bytes": 4102, + "sha256": "6285d04443db112709c516cbe08ad50f3dd2b91a3ec15eaeb9c710f3a34d1ba3" + }, + "data/robot_policy/attempts/main/franka_M_repair_early_ep20/opus-5-5/A_T0_K0_cap10/a0/attempt.json": { + "bytes": 768, + "sha256": "2ce51911ad9d86ccff68360c663434250635d4b920b0578536297a8ac9690ad9" + }, + "data/robot_policy/attempts/main/franka_M_repair_early_ep20/opus-5-5/A_T0_K0_cap10/a0/audit_A.json": { + "bytes": 2268, + "sha256": "59c5ab37be6b9fe4fe1a3b8812d83796a550e69f92339209a5c3defc348f4262" + }, + "data/robot_policy/attempts/main/franka_M_repair_early_ep20/opus-5-5/A_T0_K0_cap10/a0/audit_B.json": { + "bytes": 2190, + "sha256": "c2da0253fd3663f5a1ec4e9323cf53c368cbbf6876e2f71d3fd1fe74c607ce18" + }, + "data/robot_policy/attempts/main/franka_M_repair_early_ep20/opus-5-5/A_T0_K0_cap10/a0/work/answer_program.py": { + "bytes": 4290, + "sha256": "a8a16207f6923dd8c0ad8787dfac6c168fb397f42a387638586f80c06188ded2" + }, + "data/robot_policy/attempts/main/franka_M_repair_early_ep20/opus-5-5/A_T2_K5_cap10/a0/attempt.json": { + "bytes": 770, + "sha256": "f9f03f9c57621557a12d0b241bf138de9305e77fb37452ef9961c2dd1a1e6b2c" + }, + "data/robot_policy/attempts/main/franka_M_repair_early_ep20/opus-5-5/A_T2_K5_cap10/a0/audit_A.json": { + "bytes": 2694, + "sha256": "f96ee4df15bfc0123875cb419083e20d7b11415a88e9c337e10d4a1cceede9a8" + }, + "data/robot_policy/attempts/main/franka_M_repair_early_ep20/opus-5-5/A_T2_K5_cap10/a0/audit_B.json": { + "bytes": 2720, + "sha256": "1d3b79ee1f02cf6e3c7ae62986385e260142689fb57668b554b98a4575f56b20" + }, + "data/robot_policy/attempts/main/franka_M_repair_early_ep40/opus-5-5/A_T0_K0_cap10/a0/attempt.json": { + "bytes": 759, + "sha256": "929dca4a58d7b8445f7a1c14352a0aa5d0643de985cf121c9f7738cf0837e838" + }, + "data/robot_policy/attempts/main/franka_M_repair_early_ep40/opus-5-5/A_T0_K0_cap10/a0/audit_A.json": { + "bytes": 2346, + "sha256": "6ab15f9aa93fef8f33c642412575f9d922617ad97d9231bf8488d7181b6dd97d" + }, + "data/robot_policy/attempts/main/franka_M_repair_early_ep40/opus-5-5/A_T0_K0_cap10/a0/audit_B.json": { + "bytes": 2377, + "sha256": "b47b7c5c2e381a8f59bd45f8b3fea071a5870558f0e5de2f7cb4ddffed29b40d" + }, + "data/robot_policy/attempts/main/franka_M_repair_early_ep40/opus-5-5/A_T0_K0_cap10/a0/work/answer_program.py": { + "bytes": 4374, + "sha256": "b5834d39a95845f8f95f492ef1bdfddd544e254fcc9bcc8324dde1d21369ffbb" + }, + "data/robot_policy/attempts/main/franka_M_repair_norm_shift_21/opus-5-5/A_T0_K0_cap10/a0/attempt.json": { + "bytes": 774, + "sha256": "ea4793cca539b714841bc6ec53f7b4db8d6412a8518c7b9085b0346778b714e1" + }, + "data/robot_policy/attempts/main/franka_M_repair_norm_shift_21/opus-5-5/A_T0_K0_cap10/a0/audit_A.json": { + "bytes": 1821, + "sha256": "7218bfb85b8e5ec0661cd6fc699f769befd59828a5a68560028e4d46344a9e0f" + }, + "data/robot_policy/attempts/main/franka_M_repair_norm_shift_21/opus-5-5/A_T0_K0_cap10/a0/audit_B.json": { + "bytes": 2188, + "sha256": "1bac82792d489ccade94e70f8ce4ac270154b5206083c23e382c28a1adcff077" + }, + "data/robot_policy/attempts/main/franka_M_repair_norm_shift_21/opus-5-5/A_T0_K0_cap10/a0/work/answer_program.py": { + "bytes": 1132, + "sha256": "f4e3dcdc225c8a84852e7bd8ea0233f7ac702e00c2bc0aa154682571f5b408a6" + }, + "data/robot_policy/attempts/main/franka_M_repair_norm_shift_21/opus-5-5/A_T2_K5_cap10/a0/attempt.json": { + "bytes": 773, + "sha256": "610bed84699e2e7eef73d1f71990cc6f5b9e2d80ae307a3ee3ea93eb7878270a" + }, + "data/robot_policy/attempts/main/franka_M_repair_norm_shift_21/opus-5-5/A_T2_K5_cap10/a0/audit_A.json": { + "bytes": 2363, + "sha256": "fe37518b591b67417ffcaf14645412fc5a551f55f9e8d1a7e93e86fe432e2688" + }, + "data/robot_policy/attempts/main/franka_M_repair_norm_shift_21/opus-5-5/A_T2_K5_cap10/a0/audit_B.json": { + "bytes": 2669, + "sha256": "eec690773fd2d49828b5be31de3f1235348c1ae1dc0c9f4dce875954c6d80e9b" + }, + "data/robot_policy/attempts/main/franka_M_repair_norm_shift_21/opus-5-5/B_T0_K0_cap10/a0/attempt.json": { + "bytes": 771, + "sha256": "18b9065cd4310371b78ca5541be5c6e95dbf8b95ca46a22557fa32910d4ae4af" + }, + "data/robot_policy/attempts/main/franka_M_repair_norm_shift_21/opus-5-5/B_T0_K0_cap10/a0/audit_A.json": { + "bytes": 1837, + "sha256": "83cfca5a17e5c7e5d0542ed9142b7ff634606de3d75853d5e9acfcc8394440ea" + }, + "data/robot_policy/attempts/main/franka_M_repair_norm_shift_21/opus-5-5/B_T0_K0_cap10/a0/audit_B.json": { + "bytes": 1982, + "sha256": "8251c5297320c41e57876ac0bda11bd00c0a3148027e18c60e75f68a54fdf1d0" + }, + "data/robot_policy/attempts/main/franka_M_repair_norm_shift_21/opus-5-5/B_T0_K0_cap10/a0/work/answer_program.py": { + "bytes": 2064, + "sha256": "3e0a77d170334cec905c460c60b25a07c8f03dadc50bd149bb81628b4064cd08" + }, + "data/robot_policy/attempts/main/franka_M_repair_norm_shift_21/opus-5-5/B_T0_K0_cap10/blind0/attempt.json": { + "bytes": 778, + "sha256": "20b72cfd94dfeaf835ed70f55b21c6f4ad285c3daa8b4bb38851e3f02295a399" + }, + "data/robot_policy/attempts/main/franka_M_repair_norm_shift_21/opus-5-5/B_T0_K0_cap10/blind0/audit_A.json": { + "bytes": 1398, + "sha256": "00b20534264a3529836f38f759e640181b5b25bea31a15ff69f8222aa5113c85" + }, + "data/robot_policy/attempts/main/franka_M_repair_norm_shift_21/opus-5-5/B_T0_K0_cap10/blind0/audit_B.json": { + "bytes": 1690, + "sha256": "d19b21208ea3112f28dfb34bf8c9723f701f3e2c462b52e5b616f2a4be529abf" + }, + "data/robot_policy/attempts/main/franka_M_repair_norm_shift_21/opus-5-5/B_T0_K0_cap10/blind0/work/answer_program.py": { + "bytes": 1644, + "sha256": "650e0120874f6b220e97441900de8f6bc540264c85a7021b5ef5324cf92a86e9" + }, + "data/robot_policy/attempts/main/franka_M_repair_swap_act_1_3/opus-5-5/A_T0_K0_cap10/a0/attempt.json": { + "bytes": 768, + "sha256": "402dff6d6e9ecceaa791f10a1d9644f956756fe9b26056b04b7f30959088ee85" + }, + "data/robot_policy/attempts/main/franka_M_repair_swap_act_1_3/opus-5-5/A_T0_K0_cap10/a0/audit_A.json": { + "bytes": 1089, + "sha256": "abfab40d2ef8849820c80f9ae0210313943aead59933fe700ee5293943b471d3" + }, + "data/robot_policy/attempts/main/franka_M_repair_swap_act_1_3/opus-5-5/A_T0_K0_cap10/a0/audit_B.json": { + "bytes": 1265, + "sha256": "f9d459aa2253af52ac6ce39bb347a94feafa3064d8ce87436bbe6d1cb2fadcb8" + }, + "data/robot_policy/attempts/main/franka_M_repair_swap_act_1_3/opus-5-5/A_T0_K0_cap10/a0/work/answer_program.py": { + "bytes": 1101, + "sha256": "951874ab7b326525467a517b622bea8aef984a6f9960b335f2a40815b8ae9d09" + }, + "data/robot_policy/attempts/main/franka_M_repair_swap_act_1_3/opus-5-5/A_T0_K0_cap10/blind0/attempt.json": { + "bytes": 752, + "sha256": "521c6ab327583748947a8c9bc96a713a8865405e90733b587d0ece6b320e3fd5" + }, + "data/robot_policy/attempts/main/franka_M_repair_swap_act_1_3/opus-5-5/A_T0_K0_cap10/blind0/audit_A.json": { + "bytes": 89, + "sha256": "34cb5a7bb9eef1b1b2c05e34a92110f82faf203a862790559219b84ee0264296" + }, + "data/robot_policy/attempts/main/franka_M_repair_swap_act_1_3/opus-5-5/A_T0_K0_cap10/blind0/audit_B.json": { + "bytes": 89, + "sha256": "34cb5a7bb9eef1b1b2c05e34a92110f82faf203a862790559219b84ee0264296" + }, + "data/robot_policy/attempts/main/franka_M_repair_swap_act_1_3/opus-5-5/A_T0_K0_cap10/blind0/work/answer_program.py": { + "bytes": 1827, + "sha256": "85ba67289b79edd2566e2d5598bf3d4319e5cc0bac0e3ac74408d4d4636204f9" + }, + "data/robot_policy/attempts/main/franka_M_repair_swap_act_1_3/opus-5-5/A_T2_K5_cap10/a0/attempt.json": { + "bytes": 768, + "sha256": "5a1954c8730fe733442d16ac5601474940b324ac72e87881f18f00b81cfd0de2" + }, + "data/robot_policy/attempts/main/franka_M_repair_swap_act_1_3/opus-5-5/A_T2_K5_cap10/a0/audit_A.json": { + "bytes": 1681, + "sha256": "3e264ccc6cecd192499cfb24ab0539063438fa774e2b74f9b9f0ee60dbdae12c" + }, + "data/robot_policy/attempts/main/franka_M_repair_swap_act_1_3/opus-5-5/A_T2_K5_cap10/a0/audit_B.json": { + "bytes": 1403, + "sha256": "13cb2272bc6180e29ceb6a2e083b9fec018d23920e8a69cbb6a611145b336a2a" + }, + "data/robot_policy/attempts/main/franka_M_repair_swap_act_1_3/opus-5-5/A_T2_K5_cap10/blind0/attempt.json": { + "bytes": 772, + "sha256": "6e9c8446fc084634b0a677cab066501c074ee122269672ca5490f3da4ef899e3" + }, + "data/robot_policy/attempts/main/franka_M_repair_swap_act_1_3/opus-5-5/A_T2_K5_cap10/blind0/audit_A.json": { + "bytes": 2106, + "sha256": "15d1618cf4b848af7e7d290ea2a019d434cb9b88be378a10af61c6767f518540" + }, + "data/robot_policy/attempts/main/franka_M_repair_swap_act_1_3/opus-5-5/A_T2_K5_cap10/blind0/audit_B.json": { + "bytes": 2259, + "sha256": "39bda89a232bad2d86fc2ac675e72256e477b2fe0d205542cb592fd36f748dd3" + }, + "data/robot_policy/attempts/main/franka_M_repair_swap_act_1_3/opus-5-5/B_T0_K0_cap10/a0/attempt.json": { + "bytes": 768, + "sha256": "69301c828e249fd463e2e15730ee3173812525cd0886b8eefd1cedc804c95eb3" + }, + "data/robot_policy/attempts/main/franka_M_repair_swap_act_1_3/opus-5-5/B_T0_K0_cap10/a0/audit_A.json": { + "bytes": 883, + "sha256": "3822e5c27c5d4acb4fb622f0c8bcefb72e300ce332db3b3e7c1e2f13b5f312a8" + }, + "data/robot_policy/attempts/main/franka_M_repair_swap_act_1_3/opus-5-5/B_T0_K0_cap10/a0/audit_B.json": { + "bytes": 938, + "sha256": "4332d67299f011007234f5e527c5a38dfb75756fdd7f863a79bd5e428b79e1dd" + }, + "data/robot_policy/attempts/main/franka_M_repair_swap_act_1_3/opus-5-5/B_T0_K0_cap10/a0/work/answer_program.py": { + "bytes": 1220, + "sha256": "b6926a85066d5b65d76d907a3126f841e13850ef26eb01869210ad90946562f9" + }, + "data/robot_policy/attempts/main/franka_M_repair_swap_act_1_3/opus-5-5/B_T0_K0_cap10/blind0/attempt.json": { + "bytes": 778, + "sha256": "4908686bf098a0c06dc9624b8a384536b082d04dd7826b683960c60346bffddf" + }, + "data/robot_policy/attempts/main/franka_M_repair_swap_act_1_3/opus-5-5/B_T0_K0_cap10/blind0/audit_A.json": { + "bytes": 89, + "sha256": "34cb5a7bb9eef1b1b2c05e34a92110f82faf203a862790559219b84ee0264296" + }, + "data/robot_policy/attempts/main/franka_M_repair_swap_act_1_3/opus-5-5/B_T0_K0_cap10/blind0/audit_B.json": { + "bytes": 89, + "sha256": "34cb5a7bb9eef1b1b2c05e34a92110f82faf203a862790559219b84ee0264296" + }, + "data/robot_policy/attempts/main/franka_M_repair_swap_act_1_3/opus-5-5/B_T0_K0_cap10/blind0/work/answer_program.py": { + "bytes": 10378, + "sha256": "47c1f94b07a1f2b115c1f3f94fdfc4f40eda45709a15c1b59bedf163b34f8996" + }, + "data/robot_policy/attempts/main/franka_XS_write/opus-5-5/A_T0_K0/a0/attempt.json": { + "bytes": 694, + "sha256": "b50779771d7904e3fa4c94453e1283b30d253054d502a159cce0ed6c4352de31" + }, + "data/robot_policy/attempts/main/franka_XS_write/opus-5-5/A_T0_K0/a0/audit_A.json": { + "bytes": 2361, + "sha256": "4f93e6977b41319105fad712a16e0ce60c4610d6d9d154020a0e30ed80a4ed17" + }, + "data/robot_policy/attempts/main/franka_XS_write/opus-5-5/A_T0_K0/a0/audit_B.json": { + "bytes": 2435, + "sha256": "d4bf754d902f4d24b8ce519a7dbf6c085f280cd76734288c857f38114ced3421" + }, + "data/robot_policy/attempts/main/franka_XS_write/opus-5-5/A_T0_K0/a0/work/answer_program.py": { + "bytes": 3280, + "sha256": "2fae3a96df4df72d94cb698deab3142bc6678200dd378cb2432725cef559a8f1" + }, + "data/robot_policy/attempts/main/franka_XS_write/opus-5-5/A_T2_K5/a0/attempt.json": { + "bytes": 746, + "sha256": "2e4a371e6c4ce19eb313888335d1c5fa7fab6f9dfaf3983c18df520d6e6fb10d" + }, + "data/robot_policy/attempts/main/franka_XS_write/opus-5-5/A_T2_K5/a0/audit_A.json": { + "bytes": 2620, + "sha256": "f10980a5dba57021b2a0404096bd85f05bca1d91cbac0465cd189bb2da15e2af" + }, + "data/robot_policy/attempts/main/franka_XS_write/opus-5-5/A_T2_K5/a0/audit_B.json": { + "bytes": 4352, + "sha256": "91bfb20a0294bce1dba475e89525a789bc31630ef6623c01f183fab6c61c4ae5" + }, + "data/robot_policy/baselines/allegro_early_heldout.json": { + "bytes": 16217, + "sha256": "0f9ec98f7106202f483b4d42d18c34795862496c0791d829bd7fe6147cf9fe39" + }, + "data/robot_policy/baselines/allegro_nvidia_heldout.json": { + "bytes": 817, + "sha256": "e5295a0c1e1c903deb55d960645283056dfc26bf25e247bf34a57b0dbff34ca0" + }, + "data/robot_policy/baselines/allegro_repair_L_heldout.json": { + "bytes": 3673, + "sha256": "85dafead678f4872d4e6846fc603e45fd8d7559f0989e6e8e71cc516a92206b1" + }, + "data/robot_policy/baselines/allegro_repair_L_prompt.json": { + "bytes": 4005, + "sha256": "b580cda8f8e62eb0a6c4cd17493dbaad03cea7329878633250b9d2f195af4d8b" + }, + "data/robot_policy/baselines/det1.json": { + "bytes": 1249, + "sha256": "f52098f13ac58be65eb1c8e8f8f36b5f156083cbfe718f724cb7d8deb6917bb3" + }, + "data/robot_policy/baselines/det2.json": { + "bytes": 1254, + "sha256": "bd115e29ecddd137afaa967f588196ef33d89e263448b615c2ede70c2104327b" + }, + "data/robot_policy/baselines/franka_broken_heldout.json": { + "bytes": 8701, + "sha256": "3d2fec1d8f4f6923aaf2a640cda3f7fcdbbe89296db7554f07b02a604d5ebb2a" + }, + "data/robot_policy/baselines/franka_broken_prompt.json": { + "bytes": 9600, + "sha256": "e6ddf576d00893f7c61208a490bc5e0270712a827072640bd3ac248daf23a18d" + }, + "data/robot_policy/baselines/franka_nvidia_pretrained.json": { + "bytes": 537, + "sha256": "679f4b9823f780e958f3e63a820564ba4ac4ec181e045288a70117613f763fca" + }, + "data/robot_policy/baselines/franka_ppo_early_heldout.json": { + "bytes": 6495, + "sha256": "bd99960038936139d8e9936a929c8f8e572998a349aa438c103cf92dbf21ccc2" + }, + "data/robot_policy/baselines/franka_ppo_mid.json": { + "bytes": 5498, + "sha256": "fe78b21a65cfc4bb48f479c1d408d7a0f57c015cad9ee7c5fd518a16e463136c" + }, + "data/robot_policy/baselines/franka_ppo_v3_heldout.json": { + "bytes": 38346, + "sha256": "4839d917ddea2ff618dd920348160b694686ec7fe28d9e72a8e1a980909769d9" + }, + "data/robot_policy/baselines/franka_repair_M_heldout.json": { + "bytes": 11990, + "sha256": "5386083d8506b0c7ae3116fc34b240750e38deb2de87325e28d008ebf1c0011d" + }, + "data/robot_policy/baselines/franka_repair_M_prompt.json": { + "bytes": 13175, + "sha256": "96cd113d8b46f8095f620083be21eb7d3535e034cda9673fc5aabbbeae3712f6" + }, + "data/robot_policy/baselines/shadow_early_heldout.json": { + "bytes": 16031, + "sha256": "b72f34b93d7cbf6b843f13d6d94e560df7c05b59d159182dfbdf00105ece966e" + }, + "data/robot_policy/baselines/shadow_nvidia_heldout.json": { + "bytes": 803, + "sha256": "0bd720e8e2829a082cfd40092fc5eed3330bc9ffec047d460402c1dfed8f9d44" + }, + "data/robot_policy/baselines/shadow_repair_L_heldout.json": { + "bytes": 3673, + "sha256": "0fc4319f045a41e3b04c7ff3b87b5f349ec23816db328eca981c30bc956de5a7" + }, + "data/robot_policy/baselines/shadow_repair_L_prompt.json": { + "bytes": 4000, + "sha256": "698e3a4ede24e1bcf42ddad7df9831318d3089e570694c2b6765e9514ba22c42" + }, + "data/robot_policy/schedules/blind_jobs.jsonl": { + "bytes": 4218, + "sha256": "522dba8f1ca453f2e6df7741089fbdc42557f60b12d23fabf62d179f6008b9a8" + }, + "data/robot_policy/schedules/hands_jobs.jsonl": { + "bytes": 5581, + "sha256": "bfabf9bdde5f4a6117653c81e76277aeefac5c9b735d3c9d4f1389c43790432e" + }, + "data/robot_policy/schedules/main_jobs.jsonl": { + "bytes": 43668, + "sha256": "3ed722cd05336f22925a6e27e78f18c71a3522038d1017ec9f33f5a5ff8690d6" + }, + "data/robot_policy/summary.json": { + "bytes": 42423, + "sha256": "ab8466eb6ab041211a452afca9e0471fa9f33c7fed50e184aa4b9e565f45107d" + } +} diff --git a/metadata/data/openvla/attempts/fable-5-1_single.json b/metadata/data/openvla/attempts/fable-5-1_single.json new file mode 100644 index 0000000000000000000000000000000000000000..5f558a7906d65812d6f89cc6fc0f5937fa7be374 --- /dev/null +++ b/metadata/data/openvla/attempts/fable-5-1_single.json @@ -0,0 +1,457 @@ +{ + "config": { + "model": "claude-fable-5-1", + "scope": "single", + "targets": [ + 4 + ], + "checkpoint_identity": "8d12565209c5f8d2c0e175028b2a9b2f96822bdb0bfa422426c1dc27dcf784d8", + "status": "scored", + "rules": "A", + "development_calls": 5, + "submitted_sha256": "928557ceaf67ab2635f230ee52b77e0372bd2abe2658a25fa52486880ee63ca5" + }, + "result": { + "status": "scored", + "checkpoint_identity": "8d12565209c5f8d2c0e175028b2a9b2f96822bdb0bfa422426c1dc27dcf784d8", + "patch": { + "changed": 4096, + "total": 7541237184, + "fraction": 5.431469532201362e-07, + "patch_sha256": "928557ceaf67ab2635f230ee52b77e0372bd2abe2658a25fa52486880ee63ca5" + }, + "episodes": [ + { + "task_id": 0, + "task_name": "pick_up_the_black_bowl_between_the_plate_and_the_ramekin_and_place_it_on_the_plate", + "language": "pick up the black bowl between the plate and the ramekin and place it on the plate", + "initial_state_index": 0, + "seed": 7, + "checkpoint_identity": "8d12565209c5f8d2c0e175028b2a9b2f96822bdb0bfa422426c1dc27dcf784d8", + "patch_identity": "928557ceaf67ab2635f230ee52b77e0372bd2abe2658a25fa52486880ee63ca5", + "status": "scored", + "success": true, + "action_steps": 87, + "elapsed_seconds": 32.28518104553223 + }, + { + "task_id": 0, + "task_name": "pick_up_the_black_bowl_between_the_plate_and_the_ramekin_and_place_it_on_the_plate", + "language": "pick up the black bowl between the plate and the ramekin and place it on the plate", + "initial_state_index": 1, + "seed": 8, + "checkpoint_identity": "8d12565209c5f8d2c0e175028b2a9b2f96822bdb0bfa422426c1dc27dcf784d8", + "patch_identity": "928557ceaf67ab2635f230ee52b77e0372bd2abe2658a25fa52486880ee63ca5", + "status": "scored", + "success": true, + "action_steps": 140, + "elapsed_seconds": 47.198009729385376 + }, + { + "task_id": 0, + "task_name": "pick_up_the_black_bowl_between_the_plate_and_the_ramekin_and_place_it_on_the_plate", + "language": "pick up the black bowl between the plate and the ramekin and place it on the plate", + "initial_state_index": 2, + "seed": 9, + "checkpoint_identity": "8d12565209c5f8d2c0e175028b2a9b2f96822bdb0bfa422426c1dc27dcf784d8", + "patch_identity": "928557ceaf67ab2635f230ee52b77e0372bd2abe2658a25fa52486880ee63ca5", + "status": "scored", + "success": true, + "action_steps": 76, + "elapsed_seconds": 28.638792991638184 + }, + { + "task_id": 0, + "task_name": "pick_up_the_black_bowl_between_the_plate_and_the_ramekin_and_place_it_on_the_plate", + "language": "pick up the black bowl between the plate and the ramekin and place it on the plate", + "initial_state_index": 3, + "seed": 10, + "checkpoint_identity": "8d12565209c5f8d2c0e175028b2a9b2f96822bdb0bfa422426c1dc27dcf784d8", + "patch_identity": "928557ceaf67ab2635f230ee52b77e0372bd2abe2658a25fa52486880ee63ca5", + "status": "scored", + "success": true, + "action_steps": 188, + "elapsed_seconds": 56.86577296257019 + }, + { + "task_id": 0, + "task_name": "pick_up_the_black_bowl_between_the_plate_and_the_ramekin_and_place_it_on_the_plate", + "language": "pick up the black bowl between the plate and the ramekin and place it on the plate", + "initial_state_index": 4, + "seed": 11, + "checkpoint_identity": "8d12565209c5f8d2c0e175028b2a9b2f96822bdb0bfa422426c1dc27dcf784d8", + "patch_identity": "928557ceaf67ab2635f230ee52b77e0372bd2abe2658a25fa52486880ee63ca5", + "status": "scored", + "success": true, + "action_steps": 73, + "elapsed_seconds": 26.60857892036438 + }, + { + "task_id": 1, + "task_name": "pick_up_the_black_bowl_next_to_the_ramekin_and_place_it_on_the_plate", + "language": "pick up the black bowl next to the ramekin and place it on the plate", + "initial_state_index": 0, + "seed": 1007, + "checkpoint_identity": "8d12565209c5f8d2c0e175028b2a9b2f96822bdb0bfa422426c1dc27dcf784d8", + "patch_identity": "928557ceaf67ab2635f230ee52b77e0372bd2abe2658a25fa52486880ee63ca5", + "status": "scored", + "success": true, + "action_steps": 108, + "elapsed_seconds": 35.92068362236023 + }, + { + "task_id": 1, + "task_name": "pick_up_the_black_bowl_next_to_the_ramekin_and_place_it_on_the_plate", + "language": "pick up the black bowl next to the ramekin and place it on the plate", + "initial_state_index": 1, + "seed": 1008, + "checkpoint_identity": "8d12565209c5f8d2c0e175028b2a9b2f96822bdb0bfa422426c1dc27dcf784d8", + "patch_identity": "928557ceaf67ab2635f230ee52b77e0372bd2abe2658a25fa52486880ee63ca5", + "status": "scored", + "success": true, + "action_steps": 98, + "elapsed_seconds": 33.14826798439026 + }, + { + "task_id": 1, + "task_name": "pick_up_the_black_bowl_next_to_the_ramekin_and_place_it_on_the_plate", + "language": "pick up the black bowl next to the ramekin and place it on the plate", + "initial_state_index": 2, + "seed": 1009, + "checkpoint_identity": "8d12565209c5f8d2c0e175028b2a9b2f96822bdb0bfa422426c1dc27dcf784d8", + "patch_identity": "928557ceaf67ab2635f230ee52b77e0372bd2abe2658a25fa52486880ee63ca5", + "status": "scored", + "success": false, + "action_steps": 220, + "elapsed_seconds": 64.5421724319458 + }, + { + "task_id": 1, + "task_name": "pick_up_the_black_bowl_next_to_the_ramekin_and_place_it_on_the_plate", + "language": "pick up the black bowl next to the ramekin and place it on the plate", + "initial_state_index": 3, + "seed": 1010, + "checkpoint_identity": "8d12565209c5f8d2c0e175028b2a9b2f96822bdb0bfa422426c1dc27dcf784d8", + "patch_identity": "928557ceaf67ab2635f230ee52b77e0372bd2abe2658a25fa52486880ee63ca5", + "status": "scored", + "success": true, + "action_steps": 113, + "elapsed_seconds": 36.21302366256714 + }, + { + "task_id": 1, + "task_name": "pick_up_the_black_bowl_next_to_the_ramekin_and_place_it_on_the_plate", + "language": "pick up the black bowl next to the ramekin and place it on the plate", + "initial_state_index": 4, + "seed": 1011, + "checkpoint_identity": "8d12565209c5f8d2c0e175028b2a9b2f96822bdb0bfa422426c1dc27dcf784d8", + "patch_identity": "928557ceaf67ab2635f230ee52b77e0372bd2abe2658a25fa52486880ee63ca5", + "status": "scored", + "success": true, + "action_steps": 131, + "elapsed_seconds": 41.41289258003235 + }, + { + "task_id": 2, + "task_name": "pick_up_the_black_bowl_from_table_center_and_place_it_on_the_plate", + "language": "pick up the black bowl from table center and place it on the plate", + "initial_state_index": 0, + "seed": 2007, + "checkpoint_identity": "8d12565209c5f8d2c0e175028b2a9b2f96822bdb0bfa422426c1dc27dcf784d8", + "patch_identity": "928557ceaf67ab2635f230ee52b77e0372bd2abe2658a25fa52486880ee63ca5", + "status": "scored", + "success": false, + "action_steps": 220, + "elapsed_seconds": 64.42195081710815 + }, + { + "task_id": 2, + "task_name": "pick_up_the_black_bowl_from_table_center_and_place_it_on_the_plate", + "language": "pick up the black bowl from table center and place it on the plate", + "initial_state_index": 1, + "seed": 2008, + "checkpoint_identity": "8d12565209c5f8d2c0e175028b2a9b2f96822bdb0bfa422426c1dc27dcf784d8", + "patch_identity": "928557ceaf67ab2635f230ee52b77e0372bd2abe2658a25fa52486880ee63ca5", + "status": "scored", + "success": true, + "action_steps": 109, + "elapsed_seconds": 36.313209533691406 + }, + { + "task_id": 2, + "task_name": "pick_up_the_black_bowl_from_table_center_and_place_it_on_the_plate", + "language": "pick up the black bowl from table center and place it on the plate", + "initial_state_index": 2, + "seed": 2009, + "checkpoint_identity": "8d12565209c5f8d2c0e175028b2a9b2f96822bdb0bfa422426c1dc27dcf784d8", + "patch_identity": "928557ceaf67ab2635f230ee52b77e0372bd2abe2658a25fa52486880ee63ca5", + "status": "scored", + "success": true, + "action_steps": 101, + "elapsed_seconds": 33.63182258605957 + }, + { + "task_id": 2, + "task_name": "pick_up_the_black_bowl_from_table_center_and_place_it_on_the_plate", + "language": "pick up the black bowl from table center and place it on the plate", + "initial_state_index": 3, + "seed": 2010, + "checkpoint_identity": "8d12565209c5f8d2c0e175028b2a9b2f96822bdb0bfa422426c1dc27dcf784d8", + "patch_identity": "928557ceaf67ab2635f230ee52b77e0372bd2abe2658a25fa52486880ee63ca5", + "status": "scored", + "success": true, + "action_steps": 96, + "elapsed_seconds": 32.364559173583984 + }, + { + "task_id": 2, + "task_name": "pick_up_the_black_bowl_from_table_center_and_place_it_on_the_plate", + "language": "pick up the black bowl from table center and place it on the plate", + "initial_state_index": 4, + "seed": 2011, + "checkpoint_identity": "8d12565209c5f8d2c0e175028b2a9b2f96822bdb0bfa422426c1dc27dcf784d8", + "patch_identity": "928557ceaf67ab2635f230ee52b77e0372bd2abe2658a25fa52486880ee63ca5", + "status": "scored", + "success": true, + "action_steps": 123, + "elapsed_seconds": 38.98491883277893 + }, + { + "task_id": 3, + "task_name": "pick_up_the_black_bowl_on_the_cookie_box_and_place_it_on_the_plate", + "language": "pick up the black bowl on the cookie box and place it on the plate", + "initial_state_index": 0, + "seed": 3007, + "checkpoint_identity": "8d12565209c5f8d2c0e175028b2a9b2f96822bdb0bfa422426c1dc27dcf784d8", + "patch_identity": "928557ceaf67ab2635f230ee52b77e0372bd2abe2658a25fa52486880ee63ca5", + "status": "scored", + "success": true, + "action_steps": 92, + "elapsed_seconds": 31.037668704986572 + }, + { + "task_id": 3, + "task_name": "pick_up_the_black_bowl_on_the_cookie_box_and_place_it_on_the_plate", + "language": "pick up the black bowl on the cookie box and place it on the plate", + "initial_state_index": 1, + "seed": 3008, + "checkpoint_identity": "8d12565209c5f8d2c0e175028b2a9b2f96822bdb0bfa422426c1dc27dcf784d8", + "patch_identity": "928557ceaf67ab2635f230ee52b77e0372bd2abe2658a25fa52486880ee63ca5", + "status": "scored", + "success": true, + "action_steps": 98, + "elapsed_seconds": 32.83239483833313 + }, + { + "task_id": 3, + "task_name": "pick_up_the_black_bowl_on_the_cookie_box_and_place_it_on_the_plate", + "language": "pick up the black bowl on the cookie box and place it on the plate", + "initial_state_index": 2, + "seed": 3009, + "checkpoint_identity": "8d12565209c5f8d2c0e175028b2a9b2f96822bdb0bfa422426c1dc27dcf784d8", + "patch_identity": "928557ceaf67ab2635f230ee52b77e0372bd2abe2658a25fa52486880ee63ca5", + "status": "scored", + "success": true, + "action_steps": 88, + "elapsed_seconds": 31.38744354248047 + }, + { + "task_id": 3, + "task_name": "pick_up_the_black_bowl_on_the_cookie_box_and_place_it_on_the_plate", + "language": "pick up the black bowl on the cookie box and place it on the plate", + "initial_state_index": 3, + "seed": 3010, + "checkpoint_identity": "8d12565209c5f8d2c0e175028b2a9b2f96822bdb0bfa422426c1dc27dcf784d8", + "patch_identity": "928557ceaf67ab2635f230ee52b77e0372bd2abe2658a25fa52486880ee63ca5", + "status": "scored", + "success": true, + "action_steps": 99, + "elapsed_seconds": 32.78222703933716 + }, + { + "task_id": 3, + "task_name": "pick_up_the_black_bowl_on_the_cookie_box_and_place_it_on_the_plate", + "language": "pick up the black bowl on the cookie box and place it on the plate", + "initial_state_index": 4, + "seed": 3011, + "checkpoint_identity": "8d12565209c5f8d2c0e175028b2a9b2f96822bdb0bfa422426c1dc27dcf784d8", + "patch_identity": "928557ceaf67ab2635f230ee52b77e0372bd2abe2658a25fa52486880ee63ca5", + "status": "scored", + "success": true, + "action_steps": 87, + "elapsed_seconds": 31.390852212905884 + }, + { + "task_id": 4, + "task_name": "pick_up_the_black_bowl_in_the_top_drawer_of_the_wooden_cabinet_and_place_it_on_the_plate", + "language": "pick up the black bowl in the top drawer of the wooden cabinet and place it on the plate", + "initial_state_index": 0, + "seed": 4007, + "checkpoint_identity": "8d12565209c5f8d2c0e175028b2a9b2f96822bdb0bfa422426c1dc27dcf784d8", + "patch_identity": "928557ceaf67ab2635f230ee52b77e0372bd2abe2658a25fa52486880ee63ca5", + "status": "scored", + "success": true, + "action_steps": 127, + "elapsed_seconds": 42.98896265029907 + }, + { + "task_id": 4, + "task_name": "pick_up_the_black_bowl_in_the_top_drawer_of_the_wooden_cabinet_and_place_it_on_the_plate", + "language": "pick up the black bowl in the top drawer of the wooden cabinet and place it on the plate", + "initial_state_index": 1, + "seed": 4008, + "checkpoint_identity": "8d12565209c5f8d2c0e175028b2a9b2f96822bdb0bfa422426c1dc27dcf784d8", + "patch_identity": "928557ceaf67ab2635f230ee52b77e0372bd2abe2658a25fa52486880ee63ca5", + "status": "scored", + "success": true, + "action_steps": 139, + "elapsed_seconds": 44.51258397102356 + }, + { + "task_id": 4, + "task_name": "pick_up_the_black_bowl_in_the_top_drawer_of_the_wooden_cabinet_and_place_it_on_the_plate", + "language": "pick up the black bowl in the top drawer of the wooden cabinet and place it on the plate", + "initial_state_index": 2, + "seed": 4009, + "checkpoint_identity": "8d12565209c5f8d2c0e175028b2a9b2f96822bdb0bfa422426c1dc27dcf784d8", + "patch_identity": "928557ceaf67ab2635f230ee52b77e0372bd2abe2658a25fa52486880ee63ca5", + "status": "scored", + "success": true, + "action_steps": 134, + "elapsed_seconds": 43.162036180496216 + }, + { + "task_id": 4, + "task_name": "pick_up_the_black_bowl_in_the_top_drawer_of_the_wooden_cabinet_and_place_it_on_the_plate", + "language": "pick up the black bowl in the top drawer of the wooden cabinet and place it on the plate", + "initial_state_index": 3, + "seed": 4010, + "checkpoint_identity": "8d12565209c5f8d2c0e175028b2a9b2f96822bdb0bfa422426c1dc27dcf784d8", + "patch_identity": "928557ceaf67ab2635f230ee52b77e0372bd2abe2658a25fa52486880ee63ca5", + "status": "scored", + "success": false, + "action_steps": 220, + "elapsed_seconds": 67.58603739738464 + }, + { + "task_id": 4, + "task_name": "pick_up_the_black_bowl_in_the_top_drawer_of_the_wooden_cabinet_and_place_it_on_the_plate", + "language": "pick up the black bowl in the top drawer of the wooden cabinet and place it on the plate", + "initial_state_index": 4, + "seed": 4011, + "checkpoint_identity": "8d12565209c5f8d2c0e175028b2a9b2f96822bdb0bfa422426c1dc27dcf784d8", + "patch_identity": "928557ceaf67ab2635f230ee52b77e0372bd2abe2658a25fa52486880ee63ca5", + "status": "scored", + "success": false, + "action_steps": 220, + "elapsed_seconds": 66.2054488658905 + }, + { + "task_id": 5, + "task_name": "pick_up_the_black_bowl_on_the_ramekin_and_place_it_on_the_plate", + "language": "pick up the black bowl on the ramekin and place it on the plate", + "initial_state_index": 0, + "seed": 5007, + "checkpoint_identity": "8d12565209c5f8d2c0e175028b2a9b2f96822bdb0bfa422426c1dc27dcf784d8", + "patch_identity": "928557ceaf67ab2635f230ee52b77e0372bd2abe2658a25fa52486880ee63ca5", + "status": "scored", + "success": true, + "action_steps": 107, + "elapsed_seconds": 38.141982316970825 + }, + { + "task_id": 5, + "task_name": "pick_up_the_black_bowl_on_the_ramekin_and_place_it_on_the_plate", + "language": "pick up the black bowl on the ramekin and place it on the plate", + "initial_state_index": 1, + "seed": 5008, + "checkpoint_identity": "8d12565209c5f8d2c0e175028b2a9b2f96822bdb0bfa422426c1dc27dcf784d8", + "patch_identity": "928557ceaf67ab2635f230ee52b77e0372bd2abe2658a25fa52486880ee63ca5", + "status": "scored", + "success": true, + "action_steps": 100, + "elapsed_seconds": 35.366517066955566 + }, + { + "task_id": 5, + "task_name": "pick_up_the_black_bowl_on_the_ramekin_and_place_it_on_the_plate", + "language": "pick up the black bowl on the ramekin and place it on the plate", + "initial_state_index": 2, + "seed": 5009, + "checkpoint_identity": "8d12565209c5f8d2c0e175028b2a9b2f96822bdb0bfa422426c1dc27dcf784d8", + "patch_identity": "928557ceaf67ab2635f230ee52b77e0372bd2abe2658a25fa52486880ee63ca5", + "status": "scored", + "success": false, + "action_steps": 220, + "elapsed_seconds": 68.18078470230103 + }, + { + "task_id": 5, + "task_name": "pick_up_the_black_bowl_on_the_ramekin_and_place_it_on_the_plate", + "language": "pick up the black bowl on the ramekin and place it on the plate", + "initial_state_index": 3, + "seed": 5010, + "checkpoint_identity": "8d12565209c5f8d2c0e175028b2a9b2f96822bdb0bfa422426c1dc27dcf784d8", + "patch_identity": "928557ceaf67ab2635f230ee52b77e0372bd2abe2658a25fa52486880ee63ca5", + "status": "scored", + "success": false, + "action_steps": 220, + "elapsed_seconds": 68.73402428627014 + }, + { + "task_id": 5, + "task_name": "pick_up_the_black_bowl_on_the_ramekin_and_place_it_on_the_plate", + "language": "pick up the black bowl on the ramekin and place it on the plate", + "initial_state_index": 4, + "seed": 5011, + "checkpoint_identity": "8d12565209c5f8d2c0e175028b2a9b2f96822bdb0bfa422426c1dc27dcf784d8", + "patch_identity": "928557ceaf67ab2635f230ee52b77e0372bd2abe2658a25fa52486880ee63ca5", + "status": "scored", + "success": true, + "action_steps": 93, + "elapsed_seconds": 33.67208194732666 + } + ], + "successes": 24, + "n": 30, + "per_task": { + "0": { + "successes": 5, + "n": 5 + }, + "1": { + "successes": 4, + "n": 5 + }, + "2": { + "successes": 4, + "n": 5 + }, + "3": { + "successes": 5, + "n": 5 + }, + "4": { + "successes": 3, + "n": 5 + }, + "5": { + "successes": 3, + "n": 5 + } + } + }, + "tool_state": { + "weights": { + "changed": 4096, + "total": 7541237184, + "fraction": 5.431469532201362e-07, + "patch_sha256": "928557ceaf67ab2635f230ee52b77e0372bd2abe2658a25fa52486880ee63ca5" + }, + "explanation": "EDIT: language_model.lm_head.weight, row 31872 only (4096 entries, 5e-7 of parameters). Every entry of that row is replaced by 0.88 x its baseline value (values = (1-eps) * baseline, eps = 0.12).\n\nWHAT ROW 31872 IS: OpenVLA decodes action tokens as bin = 32000 - token_id - 1. Token 31872 is bin 127 (normalized value 0.0). For the gripper dimension (not un-normalized, mask=false) the training labels were exactly 0 (\"close\" after the dataloader flip / env inversion) or 1 (\"open\" = token 31744, bin 255). So at the gripper position the decision is a two-way contest between token 31744 (open) and token 31872 (close); all other tokens were >= 8 logits lower in every frame I measured. Because lm_head has no bias, a multiplicative scaling of the close row is the only way to add a dimension-specific \"close reluctance\": logit_close' = 0.88 * logit_close, i.e. a penalty of ~5.5-7 logits at the typical 45-60 logit level, with greedy decoding unchanged otherwise.\n\nWHY (diagnosis): Baseline was 3/5. Failure 4: the policy closed the gripper ~8 cm above the bowl (z=1.18 vs 1.09-1.10 in all successes), then toggled open/close for 40 steps while hovering (descending when open, lifting when closed), finally \"confidently\" closed on nothing and carried an empty gripper to the plate. Failure 3: a misaligned grasp followed by a confident stuck push (not addressed by this edit). To measure the model's decision margins I re-implemented the full policy in NumPy (BPE tokenizer, 0.9 center-crop preprocessing, DINOv2+SigLIP, projector, LLaMA-2 with teacher-forced action tokens); it reproduces the recorded baseline tokens exactly (7/7 on most frames, 6/7 on a few due to bf16-vs-f32 noise in uncertain translation bins). Close-minus-open logit margins on 18 recorded frames:\n - uncertain/premature phase (ep.4 steps 50/60/70): -2.0, -8.6, -2.5 (the closes in between necessarily had small positive margins); ep.1 step 60 (brief toggling before a correct close): -5.4; ep.3 step 60: -5.7.\n - confident phases: one step before a correct close (ep.0 step 50) -19.4; right after a close +33, +31; transport +24, +22, +21; placement +21, +17; ep.1 one step before release +21.7; stuck phase of ep.3 +31, +21, +20; ep.4 empty-gripper close +31, +13.\nSo confident decisions have |margin| >= 13 (mostly >= 17) while uncertain ones have |margin| <= 9. eps=0.12 gives a penalty of ~5.4-6 logits when the close logit is ~45-50 (uncertain regime) and ~6-7.2 when it is 50-60 (confident regime): comfortably above the uncertain margins and well below the smallest confident margin (13), i.e. chosen by argument as the middle of the gap. The effect is that the policy only commits to \"close\" when the close logit exceeds the open logit by ~13%; while uncertain it keeps the gripper open and keeps executing its own open-phase actions (descending, z=-0.4..-0.8), which in episode 4 would have carried it from z=1.18 to the proper grasp height where the close becomes confident (+30). Correct closes and releases are sharp transitions (margin swings of >19 in one step), so they are at most delayed by a step or two. Pushing the open gripper against the drawer bottom is benign (the compliant controller did this for 150 steps in ep.3 with no damage).\n\nSIDE EFFECTS CHECKED: Token 31872 is shared with normalized value 0.0 in the continuous dimensions (z=-0.003 \"hold\", y=0.107, x=0.096, rot~0). If it were the top token there with a close runner-up, the penalty would move the output to a neighbouring bin, a change of a few bins (~0.05 action units, i.e. <1 mm per control step at the observed ~1 cm/step full-scale motion). On all 18 analysed frames (126 token decisions) eps=0.10/0.12/0.15 changed no token at all. bf16 re-rounding of 0.88*w changes entries by <0.4%. Gripper logits for any third token were >= 8 below, so the penalty cannot promote a spurious third token.\n\nRULE A COMPLIANCE: No optimizer, solver, fitting, or automated selection was used. The values are a closed-form formula (0.88 x baseline) applied to one designed row; eps was chosen by hand from the measured margin gap (statistics of the model's own logits on recorded frames were used only to decide where to act and the scale of the intervention). The development rollouts all failed at evaluator startup (\"Broken pipe\"), so this submission rests on the offline analysis above; it is a minimal, mechanistically-justified edit with bounded downside and a direct link to an observed failure mode.", + "development_calls": 0, + "development_requests": 5, + "infrastructure_failed_requests": 5, + "infrastructure_note": "Five evaluator startup/broken-pipe requests produced no simulator episodes or feedback. Budget bookkeeping corrected; submitted policy remains frozen." + } +} diff --git a/metadata/data/openvla/attempts/fable-5-1_three.json b/metadata/data/openvla/attempts/fable-5-1_three.json new file mode 100644 index 0000000000000000000000000000000000000000..194bf5e93a4cc10cff5e368e1f1014589d676167 --- /dev/null +++ b/metadata/data/openvla/attempts/fable-5-1_three.json @@ -0,0 +1,467 @@ +{ + "config": { + "model": "claude-fable-5-1", + "scope": "three", + "targets": [ + 4, + 5, + 0 + ], + "checkpoint_identity": "8d12565209c5f8d2c0e175028b2a9b2f96822bdb0bfa422426c1dc27dcf784d8", + "status": "scored", + "rules": "A", + "development_calls": 5, + "recovery_index": 1, + "infrastructure_recoveries": [ + { + "index": 1, + "time": 1793829499.3470173, + "prior_status": "ended_without_submission", + "reason": "Sandbox OOM/broken pipe; memory limit 6 GiB -> 48 GiB", + "context_restoration": "Original prompt + previous assistant text + complete saved tool log; native session persistence was disabled", + "development_calls_preserved": 0 + } + ], + "submitted_sha256": "886cbc79f4cfa23b6b9bfceafaaf9ceef0fa351f3ac88b5e17de1bf13681c046" + }, + "result": { + "status": "scored", + "checkpoint_identity": "8d12565209c5f8d2c0e175028b2a9b2f96822bdb0bfa422426c1dc27dcf784d8", + "patch": { + "changed": 4089, + "total": 7541237184, + "fraction": 5.422187235637541e-07, + "patch_sha256": "886cbc79f4cfa23b6b9bfceafaaf9ceef0fa351f3ac88b5e17de1bf13681c046" + }, + "episodes": [ + { + "task_id": 0, + "task_name": "pick_up_the_black_bowl_between_the_plate_and_the_ramekin_and_place_it_on_the_plate", + "language": "pick up the black bowl between the plate and the ramekin and place it on the plate", + "initial_state_index": 0, + "seed": 7, + "checkpoint_identity": "8d12565209c5f8d2c0e175028b2a9b2f96822bdb0bfa422426c1dc27dcf784d8", + "patch_identity": "886cbc79f4cfa23b6b9bfceafaaf9ceef0fa351f3ac88b5e17de1bf13681c046", + "status": "scored", + "success": true, + "action_steps": 84, + "elapsed_seconds": 29.72465705871582 + }, + { + "task_id": 0, + "task_name": "pick_up_the_black_bowl_between_the_plate_and_the_ramekin_and_place_it_on_the_plate", + "language": "pick up the black bowl between the plate and the ramekin and place it on the plate", + "initial_state_index": 1, + "seed": 8, + "checkpoint_identity": "8d12565209c5f8d2c0e175028b2a9b2f96822bdb0bfa422426c1dc27dcf784d8", + "patch_identity": "886cbc79f4cfa23b6b9bfceafaaf9ceef0fa351f3ac88b5e17de1bf13681c046", + "status": "scored", + "success": true, + "action_steps": 121, + "elapsed_seconds": 38.98280429840088 + }, + { + "task_id": 0, + "task_name": "pick_up_the_black_bowl_between_the_plate_and_the_ramekin_and_place_it_on_the_plate", + "language": "pick up the black bowl between the plate and the ramekin and place it on the plate", + "initial_state_index": 2, + "seed": 9, + "checkpoint_identity": "8d12565209c5f8d2c0e175028b2a9b2f96822bdb0bfa422426c1dc27dcf784d8", + "patch_identity": "886cbc79f4cfa23b6b9bfceafaaf9ceef0fa351f3ac88b5e17de1bf13681c046", + "status": "scored", + "success": true, + "action_steps": 91, + "elapsed_seconds": 31.1226544380188 + }, + { + "task_id": 0, + "task_name": "pick_up_the_black_bowl_between_the_plate_and_the_ramekin_and_place_it_on_the_plate", + "language": "pick up the black bowl between the plate and the ramekin and place it on the plate", + "initial_state_index": 3, + "seed": 10, + "checkpoint_identity": "8d12565209c5f8d2c0e175028b2a9b2f96822bdb0bfa422426c1dc27dcf784d8", + "patch_identity": "886cbc79f4cfa23b6b9bfceafaaf9ceef0fa351f3ac88b5e17de1bf13681c046", + "status": "scored", + "success": true, + "action_steps": 138, + "elapsed_seconds": 43.57998585700989 + }, + { + "task_id": 0, + "task_name": "pick_up_the_black_bowl_between_the_plate_and_the_ramekin_and_place_it_on_the_plate", + "language": "pick up the black bowl between the plate and the ramekin and place it on the plate", + "initial_state_index": 4, + "seed": 11, + "checkpoint_identity": "8d12565209c5f8d2c0e175028b2a9b2f96822bdb0bfa422426c1dc27dcf784d8", + "patch_identity": "886cbc79f4cfa23b6b9bfceafaaf9ceef0fa351f3ac88b5e17de1bf13681c046", + "status": "scored", + "success": true, + "action_steps": 77, + "elapsed_seconds": 27.49327802658081 + }, + { + "task_id": 1, + "task_name": "pick_up_the_black_bowl_next_to_the_ramekin_and_place_it_on_the_plate", + "language": "pick up the black bowl next to the ramekin and place it on the plate", + "initial_state_index": 0, + "seed": 1007, + "checkpoint_identity": "8d12565209c5f8d2c0e175028b2a9b2f96822bdb0bfa422426c1dc27dcf784d8", + "patch_identity": "886cbc79f4cfa23b6b9bfceafaaf9ceef0fa351f3ac88b5e17de1bf13681c046", + "status": "scored", + "success": true, + "action_steps": 115, + "elapsed_seconds": 37.84458088874817 + }, + { + "task_id": 1, + "task_name": "pick_up_the_black_bowl_next_to_the_ramekin_and_place_it_on_the_plate", + "language": "pick up the black bowl next to the ramekin and place it on the plate", + "initial_state_index": 1, + "seed": 1008, + "checkpoint_identity": "8d12565209c5f8d2c0e175028b2a9b2f96822bdb0bfa422426c1dc27dcf784d8", + "patch_identity": "886cbc79f4cfa23b6b9bfceafaaf9ceef0fa351f3ac88b5e17de1bf13681c046", + "status": "scored", + "success": true, + "action_steps": 141, + "elapsed_seconds": 44.93179225921631 + }, + { + "task_id": 1, + "task_name": "pick_up_the_black_bowl_next_to_the_ramekin_and_place_it_on_the_plate", + "language": "pick up the black bowl next to the ramekin and place it on the plate", + "initial_state_index": 2, + "seed": 1009, + "checkpoint_identity": "8d12565209c5f8d2c0e175028b2a9b2f96822bdb0bfa422426c1dc27dcf784d8", + "patch_identity": "886cbc79f4cfa23b6b9bfceafaaf9ceef0fa351f3ac88b5e17de1bf13681c046", + "status": "scored", + "success": true, + "action_steps": 126, + "elapsed_seconds": 41.03465485572815 + }, + { + "task_id": 1, + "task_name": "pick_up_the_black_bowl_next_to_the_ramekin_and_place_it_on_the_plate", + "language": "pick up the black bowl next to the ramekin and place it on the plate", + "initial_state_index": 3, + "seed": 1010, + "checkpoint_identity": "8d12565209c5f8d2c0e175028b2a9b2f96822bdb0bfa422426c1dc27dcf784d8", + "patch_identity": "886cbc79f4cfa23b6b9bfceafaaf9ceef0fa351f3ac88b5e17de1bf13681c046", + "status": "scored", + "success": true, + "action_steps": 114, + "elapsed_seconds": 37.38721704483032 + }, + { + "task_id": 1, + "task_name": "pick_up_the_black_bowl_next_to_the_ramekin_and_place_it_on_the_plate", + "language": "pick up the black bowl next to the ramekin and place it on the plate", + "initial_state_index": 4, + "seed": 1011, + "checkpoint_identity": "8d12565209c5f8d2c0e175028b2a9b2f96822bdb0bfa422426c1dc27dcf784d8", + "patch_identity": "886cbc79f4cfa23b6b9bfceafaaf9ceef0fa351f3ac88b5e17de1bf13681c046", + "status": "scored", + "success": true, + "action_steps": 123, + "elapsed_seconds": 39.840441942214966 + }, + { + "task_id": 2, + "task_name": "pick_up_the_black_bowl_from_table_center_and_place_it_on_the_plate", + "language": "pick up the black bowl from table center and place it on the plate", + "initial_state_index": 0, + "seed": 2007, + "checkpoint_identity": "8d12565209c5f8d2c0e175028b2a9b2f96822bdb0bfa422426c1dc27dcf784d8", + "patch_identity": "886cbc79f4cfa23b6b9bfceafaaf9ceef0fa351f3ac88b5e17de1bf13681c046", + "status": "scored", + "success": false, + "action_steps": 220, + "elapsed_seconds": 65.36960959434509 + }, + { + "task_id": 2, + "task_name": "pick_up_the_black_bowl_from_table_center_and_place_it_on_the_plate", + "language": "pick up the black bowl from table center and place it on the plate", + "initial_state_index": 1, + "seed": 2008, + "checkpoint_identity": "8d12565209c5f8d2c0e175028b2a9b2f96822bdb0bfa422426c1dc27dcf784d8", + "patch_identity": "886cbc79f4cfa23b6b9bfceafaaf9ceef0fa351f3ac88b5e17de1bf13681c046", + "status": "scored", + "success": true, + "action_steps": 112, + "elapsed_seconds": 38.377187967300415 + }, + { + "task_id": 2, + "task_name": "pick_up_the_black_bowl_from_table_center_and_place_it_on_the_plate", + "language": "pick up the black bowl from table center and place it on the plate", + "initial_state_index": 2, + "seed": 2009, + "checkpoint_identity": "8d12565209c5f8d2c0e175028b2a9b2f96822bdb0bfa422426c1dc27dcf784d8", + "patch_identity": "886cbc79f4cfa23b6b9bfceafaaf9ceef0fa351f3ac88b5e17de1bf13681c046", + "status": "scored", + "success": true, + "action_steps": 94, + "elapsed_seconds": 32.136449575424194 + }, + { + "task_id": 2, + "task_name": "pick_up_the_black_bowl_from_table_center_and_place_it_on_the_plate", + "language": "pick up the black bowl from table center and place it on the plate", + "initial_state_index": 3, + "seed": 2010, + "checkpoint_identity": "8d12565209c5f8d2c0e175028b2a9b2f96822bdb0bfa422426c1dc27dcf784d8", + "patch_identity": "886cbc79f4cfa23b6b9bfceafaaf9ceef0fa351f3ac88b5e17de1bf13681c046", + "status": "scored", + "success": true, + "action_steps": 95, + "elapsed_seconds": 32.709731340408325 + }, + { + "task_id": 2, + "task_name": "pick_up_the_black_bowl_from_table_center_and_place_it_on_the_plate", + "language": "pick up the black bowl from table center and place it on the plate", + "initial_state_index": 4, + "seed": 2011, + "checkpoint_identity": "8d12565209c5f8d2c0e175028b2a9b2f96822bdb0bfa422426c1dc27dcf784d8", + "patch_identity": "886cbc79f4cfa23b6b9bfceafaaf9ceef0fa351f3ac88b5e17de1bf13681c046", + "status": "scored", + "success": true, + "action_steps": 136, + "elapsed_seconds": 43.73721504211426 + }, + { + "task_id": 3, + "task_name": "pick_up_the_black_bowl_on_the_cookie_box_and_place_it_on_the_plate", + "language": "pick up the black bowl on the cookie box and place it on the plate", + "initial_state_index": 0, + "seed": 3007, + "checkpoint_identity": "8d12565209c5f8d2c0e175028b2a9b2f96822bdb0bfa422426c1dc27dcf784d8", + "patch_identity": "886cbc79f4cfa23b6b9bfceafaaf9ceef0fa351f3ac88b5e17de1bf13681c046", + "status": "scored", + "success": true, + "action_steps": 87, + "elapsed_seconds": 30.163749933242798 + }, + { + "task_id": 3, + "task_name": "pick_up_the_black_bowl_on_the_cookie_box_and_place_it_on_the_plate", + "language": "pick up the black bowl on the cookie box and place it on the plate", + "initial_state_index": 1, + "seed": 3008, + "checkpoint_identity": "8d12565209c5f8d2c0e175028b2a9b2f96822bdb0bfa422426c1dc27dcf784d8", + "patch_identity": "886cbc79f4cfa23b6b9bfceafaaf9ceef0fa351f3ac88b5e17de1bf13681c046", + "status": "scored", + "success": true, + "action_steps": 99, + "elapsed_seconds": 34.43239903450012 + }, + { + "task_id": 3, + "task_name": "pick_up_the_black_bowl_on_the_cookie_box_and_place_it_on_the_plate", + "language": "pick up the black bowl on the cookie box and place it on the plate", + "initial_state_index": 2, + "seed": 3009, + "checkpoint_identity": "8d12565209c5f8d2c0e175028b2a9b2f96822bdb0bfa422426c1dc27dcf784d8", + "patch_identity": "886cbc79f4cfa23b6b9bfceafaaf9ceef0fa351f3ac88b5e17de1bf13681c046", + "status": "scored", + "success": true, + "action_steps": 80, + "elapsed_seconds": 29.09462833404541 + }, + { + "task_id": 3, + "task_name": "pick_up_the_black_bowl_on_the_cookie_box_and_place_it_on_the_plate", + "language": "pick up the black bowl on the cookie box and place it on the plate", + "initial_state_index": 3, + "seed": 3010, + "checkpoint_identity": "8d12565209c5f8d2c0e175028b2a9b2f96822bdb0bfa422426c1dc27dcf784d8", + "patch_identity": "886cbc79f4cfa23b6b9bfceafaaf9ceef0fa351f3ac88b5e17de1bf13681c046", + "status": "scored", + "success": true, + "action_steps": 93, + "elapsed_seconds": 32.80756998062134 + }, + { + "task_id": 3, + "task_name": "pick_up_the_black_bowl_on_the_cookie_box_and_place_it_on_the_plate", + "language": "pick up the black bowl on the cookie box and place it on the plate", + "initial_state_index": 4, + "seed": 3011, + "checkpoint_identity": "8d12565209c5f8d2c0e175028b2a9b2f96822bdb0bfa422426c1dc27dcf784d8", + "patch_identity": "886cbc79f4cfa23b6b9bfceafaaf9ceef0fa351f3ac88b5e17de1bf13681c046", + "status": "scored", + "success": true, + "action_steps": 86, + "elapsed_seconds": 30.964547157287598 + }, + { + "task_id": 4, + "task_name": "pick_up_the_black_bowl_in_the_top_drawer_of_the_wooden_cabinet_and_place_it_on_the_plate", + "language": "pick up the black bowl in the top drawer of the wooden cabinet and place it on the plate", + "initial_state_index": 0, + "seed": 4007, + "checkpoint_identity": "8d12565209c5f8d2c0e175028b2a9b2f96822bdb0bfa422426c1dc27dcf784d8", + "patch_identity": "886cbc79f4cfa23b6b9bfceafaaf9ceef0fa351f3ac88b5e17de1bf13681c046", + "status": "scored", + "success": true, + "action_steps": 121, + "elapsed_seconds": 40.85270285606384 + }, + { + "task_id": 4, + "task_name": "pick_up_the_black_bowl_in_the_top_drawer_of_the_wooden_cabinet_and_place_it_on_the_plate", + "language": "pick up the black bowl in the top drawer of the wooden cabinet and place it on the plate", + "initial_state_index": 1, + "seed": 4008, + "checkpoint_identity": "8d12565209c5f8d2c0e175028b2a9b2f96822bdb0bfa422426c1dc27dcf784d8", + "patch_identity": "886cbc79f4cfa23b6b9bfceafaaf9ceef0fa351f3ac88b5e17de1bf13681c046", + "status": "scored", + "success": false, + "action_steps": 220, + "elapsed_seconds": 69.49626874923706 + }, + { + "task_id": 4, + "task_name": "pick_up_the_black_bowl_in_the_top_drawer_of_the_wooden_cabinet_and_place_it_on_the_plate", + "language": "pick up the black bowl in the top drawer of the wooden cabinet and place it on the plate", + "initial_state_index": 2, + "seed": 4009, + "checkpoint_identity": "8d12565209c5f8d2c0e175028b2a9b2f96822bdb0bfa422426c1dc27dcf784d8", + "patch_identity": "886cbc79f4cfa23b6b9bfceafaaf9ceef0fa351f3ac88b5e17de1bf13681c046", + "status": "scored", + "success": true, + "action_steps": 132, + "elapsed_seconds": 44.05871081352234 + }, + { + "task_id": 4, + "task_name": "pick_up_the_black_bowl_in_the_top_drawer_of_the_wooden_cabinet_and_place_it_on_the_plate", + "language": "pick up the black bowl in the top drawer of the wooden cabinet and place it on the plate", + "initial_state_index": 3, + "seed": 4010, + "checkpoint_identity": "8d12565209c5f8d2c0e175028b2a9b2f96822bdb0bfa422426c1dc27dcf784d8", + "patch_identity": "886cbc79f4cfa23b6b9bfceafaaf9ceef0fa351f3ac88b5e17de1bf13681c046", + "status": "scored", + "success": false, + "action_steps": 220, + "elapsed_seconds": 69.50604724884033 + }, + { + "task_id": 4, + "task_name": "pick_up_the_black_bowl_in_the_top_drawer_of_the_wooden_cabinet_and_place_it_on_the_plate", + "language": "pick up the black bowl in the top drawer of the wooden cabinet and place it on the plate", + "initial_state_index": 4, + "seed": 4011, + "checkpoint_identity": "8d12565209c5f8d2c0e175028b2a9b2f96822bdb0bfa422426c1dc27dcf784d8", + "patch_identity": "886cbc79f4cfa23b6b9bfceafaaf9ceef0fa351f3ac88b5e17de1bf13681c046", + "status": "scored", + "success": false, + "action_steps": 220, + "elapsed_seconds": 69.81189799308777 + }, + { + "task_id": 5, + "task_name": "pick_up_the_black_bowl_on_the_ramekin_and_place_it_on_the_plate", + "language": "pick up the black bowl on the ramekin and place it on the plate", + "initial_state_index": 0, + "seed": 5007, + "checkpoint_identity": "8d12565209c5f8d2c0e175028b2a9b2f96822bdb0bfa422426c1dc27dcf784d8", + "patch_identity": "886cbc79f4cfa23b6b9bfceafaaf9ceef0fa351f3ac88b5e17de1bf13681c046", + "status": "scored", + "success": true, + "action_steps": 103, + "elapsed_seconds": 35.56334638595581 + }, + { + "task_id": 5, + "task_name": "pick_up_the_black_bowl_on_the_ramekin_and_place_it_on_the_plate", + "language": "pick up the black bowl on the ramekin and place it on the plate", + "initial_state_index": 1, + "seed": 5008, + "checkpoint_identity": "8d12565209c5f8d2c0e175028b2a9b2f96822bdb0bfa422426c1dc27dcf784d8", + "patch_identity": "886cbc79f4cfa23b6b9bfceafaaf9ceef0fa351f3ac88b5e17de1bf13681c046", + "status": "scored", + "success": true, + "action_steps": 94, + "elapsed_seconds": 33.10503101348877 + }, + { + "task_id": 5, + "task_name": "pick_up_the_black_bowl_on_the_ramekin_and_place_it_on_the_plate", + "language": "pick up the black bowl on the ramekin and place it on the plate", + "initial_state_index": 2, + "seed": 5009, + "checkpoint_identity": "8d12565209c5f8d2c0e175028b2a9b2f96822bdb0bfa422426c1dc27dcf784d8", + "patch_identity": "886cbc79f4cfa23b6b9bfceafaaf9ceef0fa351f3ac88b5e17de1bf13681c046", + "status": "scored", + "success": true, + "action_steps": 81, + "elapsed_seconds": 29.439720630645752 + }, + { + "task_id": 5, + "task_name": "pick_up_the_black_bowl_on_the_ramekin_and_place_it_on_the_plate", + "language": "pick up the black bowl on the ramekin and place it on the plate", + "initial_state_index": 3, + "seed": 5010, + "checkpoint_identity": "8d12565209c5f8d2c0e175028b2a9b2f96822bdb0bfa422426c1dc27dcf784d8", + "patch_identity": "886cbc79f4cfa23b6b9bfceafaaf9ceef0fa351f3ac88b5e17de1bf13681c046", + "status": "scored", + "success": true, + "action_steps": 85, + "elapsed_seconds": 30.670159339904785 + }, + { + "task_id": 5, + "task_name": "pick_up_the_black_bowl_on_the_ramekin_and_place_it_on_the_plate", + "language": "pick up the black bowl on the ramekin and place it on the plate", + "initial_state_index": 4, + "seed": 5011, + "checkpoint_identity": "8d12565209c5f8d2c0e175028b2a9b2f96822bdb0bfa422426c1dc27dcf784d8", + "patch_identity": "886cbc79f4cfa23b6b9bfceafaaf9ceef0fa351f3ac88b5e17de1bf13681c046", + "status": "scored", + "success": true, + "action_steps": 85, + "elapsed_seconds": 30.775760889053345 + } + ], + "successes": 26, + "n": 30, + "per_task": { + "0": { + "successes": 5, + "n": 5 + }, + "1": { + "successes": 5, + "n": 5 + }, + "2": { + "successes": 4, + "n": 5 + }, + "3": { + "successes": 5, + "n": 5 + }, + "4": { + "successes": 2, + "n": 5 + }, + "5": { + "successes": 5, + "n": 5 + } + } + }, + "tool_state": { + "weights": { + "changed": 4089, + "total": 7541237184, + "fraction": 5.422187235637541e-07, + "patch_sha256": "886cbc79f4cfa23b6b9bfceafaaf9ceef0fa351f3ac88b5e17de1bf13681c046" + }, + "explanation": "EDIT (4096 entries, 5.4e-7 of parameters): language_model.lm_head.weight row for action token 31899 (= bin 100, x-translation \u22120.083) is replaced by bf16(1.01 \u00d7 row of token 31891) (= bin 108, x \u22120.029). All other weights untouched.\n\nDIAGNOSIS. I rebuilt the full OpenVLA inference path in NumPy (DINOv2+SigLIP, projector, LLaMA-7B, tokenizer, crop, action de-tokenization; validated \u2014 it reproduces all 7 recorded action tokens exactly on recorded frames) and teacher-forced it on all 207 recorded frames of the 15 baseline episodes. Gripper close/open decisions are confident (|logit margin| mostly 15\u201345); the baseline failures are physical (t0: gripper jams in the drawer after a confident close; t1: bowl grasp fails) \u2014 not decision-boundary errors, so no gripper edit is justified. The one systematic, policy-level defect found: a sticky \"teleop-idle\" mode. On some initial states (2 of 5 t2 episodes) the policy emits x = bin 108 (\u22120.03) with every other dimension at its exact-zero bin and gripper open, for 80\u2013100 consecutive steps. The policy is memoryless and the image is static, so this is a fixed point; it is only broken because the \u22120.03 command creeps the arm ~0.2 mm/step until, after ~1\u20132 cm, the arm's appearance leaves the idle basin. 131 of the 135 pos-0 uses of bin 108 in 2021 recorded steps are inside those two stalls (the ordinary mid-task \"x\u22480\" is bin 112, which is not sticky). Stalls of >~130 steps would time out at 220, so this is a latent failure mode.\n\nMECHANISM. Greedy decoding is invariant to everything except argmax, so I retarget the sticky token: wherever bin 108 would win, bin 100 now wins by 1% of its (always positive, \u226513.9) logit, i.e. \u22650.14 \u2014 ~35\u00d7 the bf16 rounding noise of the copied row (~0.004, estimated analytically and checked). The executed creep in the sticky mode becomes \u22120.083 (\u22483\u00d7 faster), nothing else about the decision structure changes. \u03b5=0.01 was chosen by argument: large enough to break the tie robustly under bf16, small enough not to promote the copied row where 108 is merely a runner-up. Bin 100 (rather than a larger creep) was chosen by argument: ~3\u00d7 faster escape (\u22481.1 mm/step nominal) while the one-step perturbation when 108 is legitimately emitted (4 of ~1830 non-stall pos-0 steps) stays \u22640.7 mm, and the overwritten bin's own rare uses (0.5% of decisions) shift only to a neighbouring bin. Global token boosts/penalties were rejected after measuring that stall and legitimate frames overlap heavily in logit gaps.\n\nVERIFICATION. Exact bf16 emulation over all 1449 cached decisions: 20 changes \u2014 17 are the intended 108\u2192100 retargets (14 at x in stall frames), 3 are one-bin shifts in roll/yaw where the original bin-100 row had been argmax. Full autoregressive decoding with the patched head on stall frames gives x=\u22120.08 with all other dims idle. Development rollouts (5 calls, 15 episodes): 12/15, identical to baseline 12/15; all three failures are baseline modes (t0 drawer jam; two t2 grasp-on-air misses with no x-creep activations before the grasp); the sticky mode appeared twice (call2 t2, call3 t2) and resolved within \u226435 steps vs 80\u2013100 in baseline. No fitting, optimization, or automated selection was used; rollout data served only to locate the mechanism and verify the edit; the values are a closed-form transform (1.01\u00d7, bf16-rounded) of an existing row.", + "development_calls": 5 + } +} diff --git a/metadata/data/openvla/attempts/opus-5-5_single.json b/metadata/data/openvla/attempts/opus-5-5_single.json new file mode 100644 index 0000000000000000000000000000000000000000..aa0cc739a0013c1d621d628d7e3c07a3076132a7 --- /dev/null +++ b/metadata/data/openvla/attempts/opus-5-5_single.json @@ -0,0 +1,465 @@ +{ + "config": { + "model": "claude-opus-5-5", + "scope": "single", + "targets": [ + 4 + ], + "checkpoint_identity": "8d12565209c5f8d2c0e175028b2a9b2f96822bdb0bfa422426c1dc27dcf784d8", + "status": "scored", + "rules": "A", + "development_calls": 5, + "recovery_index": 1, + "infrastructure_recoveries": [ + { + "index": 1, + "time": 1793829498.4416785, + "prior_status": "ended_without_submission", + "reason": "Sandbox OOM/broken pipe; memory limit 6 GiB -> 48 GiB", + "context_restoration": "Original prompt + previous assistant text + complete saved tool log; native session persistence was disabled", + "development_calls_preserved": 0 + } + ], + "submitted_sha256": "bf5bd8eee9e056055bf920c2391541bfd57c402631da9fc966ee81a4235889ad" + }, + "result": { + "status": "scored", + "checkpoint_identity": "8d12565209c5f8d2c0e175028b2a9b2f96822bdb0bfa422426c1dc27dcf784d8", + "patch": { + "changed": 4023, + "total": 7541237184, + "fraction": 5.334668439464375e-07, + "patch_sha256": "bf5bd8eee9e056055bf920c2391541bfd57c402631da9fc966ee81a4235889ad" + }, + "episodes": [ + { + "task_id": 0, + "task_name": "pick_up_the_black_bowl_between_the_plate_and_the_ramekin_and_place_it_on_the_plate", + "language": "pick up the black bowl between the plate and the ramekin and place it on the plate", + "initial_state_index": 0, + "seed": 7, + "checkpoint_identity": "8d12565209c5f8d2c0e175028b2a9b2f96822bdb0bfa422426c1dc27dcf784d8", + "patch_identity": "bf5bd8eee9e056055bf920c2391541bfd57c402631da9fc966ee81a4235889ad", + "status": "scored", + "success": true, + "action_steps": 87, + "elapsed_seconds": 32.10499119758606 + }, + { + "task_id": 0, + "task_name": "pick_up_the_black_bowl_between_the_plate_and_the_ramekin_and_place_it_on_the_plate", + "language": "pick up the black bowl between the plate and the ramekin and place it on the plate", + "initial_state_index": 1, + "seed": 8, + "checkpoint_identity": "8d12565209c5f8d2c0e175028b2a9b2f96822bdb0bfa422426c1dc27dcf784d8", + "patch_identity": "bf5bd8eee9e056055bf920c2391541bfd57c402631da9fc966ee81a4235889ad", + "status": "scored", + "success": true, + "action_steps": 140, + "elapsed_seconds": 44.286954402923584 + }, + { + "task_id": 0, + "task_name": "pick_up_the_black_bowl_between_the_plate_and_the_ramekin_and_place_it_on_the_plate", + "language": "pick up the black bowl between the plate and the ramekin and place it on the plate", + "initial_state_index": 2, + "seed": 9, + "checkpoint_identity": "8d12565209c5f8d2c0e175028b2a9b2f96822bdb0bfa422426c1dc27dcf784d8", + "patch_identity": "bf5bd8eee9e056055bf920c2391541bfd57c402631da9fc966ee81a4235889ad", + "status": "scored", + "success": true, + "action_steps": 78, + "elapsed_seconds": 28.51147961616516 + }, + { + "task_id": 0, + "task_name": "pick_up_the_black_bowl_between_the_plate_and_the_ramekin_and_place_it_on_the_plate", + "language": "pick up the black bowl between the plate and the ramekin and place it on the plate", + "initial_state_index": 3, + "seed": 10, + "checkpoint_identity": "8d12565209c5f8d2c0e175028b2a9b2f96822bdb0bfa422426c1dc27dcf784d8", + "patch_identity": "bf5bd8eee9e056055bf920c2391541bfd57c402631da9fc966ee81a4235889ad", + "status": "scored", + "success": true, + "action_steps": 187, + "elapsed_seconds": 57.89303159713745 + }, + { + "task_id": 0, + "task_name": "pick_up_the_black_bowl_between_the_plate_and_the_ramekin_and_place_it_on_the_plate", + "language": "pick up the black bowl between the plate and the ramekin and place it on the plate", + "initial_state_index": 4, + "seed": 11, + "checkpoint_identity": "8d12565209c5f8d2c0e175028b2a9b2f96822bdb0bfa422426c1dc27dcf784d8", + "patch_identity": "bf5bd8eee9e056055bf920c2391541bfd57c402631da9fc966ee81a4235889ad", + "status": "scored", + "success": true, + "action_steps": 74, + "elapsed_seconds": 26.68747854232788 + }, + { + "task_id": 1, + "task_name": "pick_up_the_black_bowl_next_to_the_ramekin_and_place_it_on_the_plate", + "language": "pick up the black bowl next to the ramekin and place it on the plate", + "initial_state_index": 0, + "seed": 1007, + "checkpoint_identity": "8d12565209c5f8d2c0e175028b2a9b2f96822bdb0bfa422426c1dc27dcf784d8", + "patch_identity": "bf5bd8eee9e056055bf920c2391541bfd57c402631da9fc966ee81a4235889ad", + "status": "scored", + "success": true, + "action_steps": 113, + "elapsed_seconds": 36.93220233917236 + }, + { + "task_id": 1, + "task_name": "pick_up_the_black_bowl_next_to_the_ramekin_and_place_it_on_the_plate", + "language": "pick up the black bowl next to the ramekin and place it on the plate", + "initial_state_index": 1, + "seed": 1008, + "checkpoint_identity": "8d12565209c5f8d2c0e175028b2a9b2f96822bdb0bfa422426c1dc27dcf784d8", + "patch_identity": "bf5bd8eee9e056055bf920c2391541bfd57c402631da9fc966ee81a4235889ad", + "status": "scored", + "success": true, + "action_steps": 99, + "elapsed_seconds": 33.06700587272644 + }, + { + "task_id": 1, + "task_name": "pick_up_the_black_bowl_next_to_the_ramekin_and_place_it_on_the_plate", + "language": "pick up the black bowl next to the ramekin and place it on the plate", + "initial_state_index": 2, + "seed": 1009, + "checkpoint_identity": "8d12565209c5f8d2c0e175028b2a9b2f96822bdb0bfa422426c1dc27dcf784d8", + "patch_identity": "bf5bd8eee9e056055bf920c2391541bfd57c402631da9fc966ee81a4235889ad", + "status": "scored", + "success": true, + "action_steps": 119, + "elapsed_seconds": 38.35019493103027 + }, + { + "task_id": 1, + "task_name": "pick_up_the_black_bowl_next_to_the_ramekin_and_place_it_on_the_plate", + "language": "pick up the black bowl next to the ramekin and place it on the plate", + "initial_state_index": 3, + "seed": 1010, + "checkpoint_identity": "8d12565209c5f8d2c0e175028b2a9b2f96822bdb0bfa422426c1dc27dcf784d8", + "patch_identity": "bf5bd8eee9e056055bf920c2391541bfd57c402631da9fc966ee81a4235889ad", + "status": "scored", + "success": true, + "action_steps": 115, + "elapsed_seconds": 37.39864921569824 + }, + { + "task_id": 1, + "task_name": "pick_up_the_black_bowl_next_to_the_ramekin_and_place_it_on_the_plate", + "language": "pick up the black bowl next to the ramekin and place it on the plate", + "initial_state_index": 4, + "seed": 1011, + "checkpoint_identity": "8d12565209c5f8d2c0e175028b2a9b2f96822bdb0bfa422426c1dc27dcf784d8", + "patch_identity": "bf5bd8eee9e056055bf920c2391541bfd57c402631da9fc966ee81a4235889ad", + "status": "scored", + "success": true, + "action_steps": 116, + "elapsed_seconds": 37.98599934577942 + }, + { + "task_id": 2, + "task_name": "pick_up_the_black_bowl_from_table_center_and_place_it_on_the_plate", + "language": "pick up the black bowl from table center and place it on the plate", + "initial_state_index": 0, + "seed": 2007, + "checkpoint_identity": "8d12565209c5f8d2c0e175028b2a9b2f96822bdb0bfa422426c1dc27dcf784d8", + "patch_identity": "bf5bd8eee9e056055bf920c2391541bfd57c402631da9fc966ee81a4235889ad", + "status": "scored", + "success": true, + "action_steps": 175, + "elapsed_seconds": 53.896360874176025 + }, + { + "task_id": 2, + "task_name": "pick_up_the_black_bowl_from_table_center_and_place_it_on_the_plate", + "language": "pick up the black bowl from table center and place it on the plate", + "initial_state_index": 1, + "seed": 2008, + "checkpoint_identity": "8d12565209c5f8d2c0e175028b2a9b2f96822bdb0bfa422426c1dc27dcf784d8", + "patch_identity": "bf5bd8eee9e056055bf920c2391541bfd57c402631da9fc966ee81a4235889ad", + "status": "scored", + "success": true, + "action_steps": 108, + "elapsed_seconds": 36.41280794143677 + }, + { + "task_id": 2, + "task_name": "pick_up_the_black_bowl_from_table_center_and_place_it_on_the_plate", + "language": "pick up the black bowl from table center and place it on the plate", + "initial_state_index": 2, + "seed": 2009, + "checkpoint_identity": "8d12565209c5f8d2c0e175028b2a9b2f96822bdb0bfa422426c1dc27dcf784d8", + "patch_identity": "bf5bd8eee9e056055bf920c2391541bfd57c402631da9fc966ee81a4235889ad", + "status": "scored", + "success": true, + "action_steps": 103, + "elapsed_seconds": 34.67871928215027 + }, + { + "task_id": 2, + "task_name": "pick_up_the_black_bowl_from_table_center_and_place_it_on_the_plate", + "language": "pick up the black bowl from table center and place it on the plate", + "initial_state_index": 3, + "seed": 2010, + "checkpoint_identity": "8d12565209c5f8d2c0e175028b2a9b2f96822bdb0bfa422426c1dc27dcf784d8", + "patch_identity": "bf5bd8eee9e056055bf920c2391541bfd57c402631da9fc966ee81a4235889ad", + "status": "scored", + "success": true, + "action_steps": 99, + "elapsed_seconds": 33.66429591178894 + }, + { + "task_id": 2, + "task_name": "pick_up_the_black_bowl_from_table_center_and_place_it_on_the_plate", + "language": "pick up the black bowl from table center and place it on the plate", + "initial_state_index": 4, + "seed": 2011, + "checkpoint_identity": "8d12565209c5f8d2c0e175028b2a9b2f96822bdb0bfa422426c1dc27dcf784d8", + "patch_identity": "bf5bd8eee9e056055bf920c2391541bfd57c402631da9fc966ee81a4235889ad", + "status": "scored", + "success": true, + "action_steps": 134, + "elapsed_seconds": 43.17857837677002 + }, + { + "task_id": 3, + "task_name": "pick_up_the_black_bowl_on_the_cookie_box_and_place_it_on_the_plate", + "language": "pick up the black bowl on the cookie box and place it on the plate", + "initial_state_index": 0, + "seed": 3007, + "checkpoint_identity": "8d12565209c5f8d2c0e175028b2a9b2f96822bdb0bfa422426c1dc27dcf784d8", + "patch_identity": "bf5bd8eee9e056055bf920c2391541bfd57c402631da9fc966ee81a4235889ad", + "status": "scored", + "success": true, + "action_steps": 89, + "elapsed_seconds": 31.291167736053467 + }, + { + "task_id": 3, + "task_name": "pick_up_the_black_bowl_on_the_cookie_box_and_place_it_on_the_plate", + "language": "pick up the black bowl on the cookie box and place it on the plate", + "initial_state_index": 1, + "seed": 3008, + "checkpoint_identity": "8d12565209c5f8d2c0e175028b2a9b2f96822bdb0bfa422426c1dc27dcf784d8", + "patch_identity": "bf5bd8eee9e056055bf920c2391541bfd57c402631da9fc966ee81a4235889ad", + "status": "scored", + "success": true, + "action_steps": 92, + "elapsed_seconds": 32.29636240005493 + }, + { + "task_id": 3, + "task_name": "pick_up_the_black_bowl_on_the_cookie_box_and_place_it_on_the_plate", + "language": "pick up the black bowl on the cookie box and place it on the plate", + "initial_state_index": 2, + "seed": 3009, + "checkpoint_identity": "8d12565209c5f8d2c0e175028b2a9b2f96822bdb0bfa422426c1dc27dcf784d8", + "patch_identity": "bf5bd8eee9e056055bf920c2391541bfd57c402631da9fc966ee81a4235889ad", + "status": "scored", + "success": true, + "action_steps": 88, + "elapsed_seconds": 31.056485891342163 + }, + { + "task_id": 3, + "task_name": "pick_up_the_black_bowl_on_the_cookie_box_and_place_it_on_the_plate", + "language": "pick up the black bowl on the cookie box and place it on the plate", + "initial_state_index": 3, + "seed": 3010, + "checkpoint_identity": "8d12565209c5f8d2c0e175028b2a9b2f96822bdb0bfa422426c1dc27dcf784d8", + "patch_identity": "bf5bd8eee9e056055bf920c2391541bfd57c402631da9fc966ee81a4235889ad", + "status": "scored", + "success": true, + "action_steps": 99, + "elapsed_seconds": 33.57276177406311 + }, + { + "task_id": 3, + "task_name": "pick_up_the_black_bowl_on_the_cookie_box_and_place_it_on_the_plate", + "language": "pick up the black bowl on the cookie box and place it on the plate", + "initial_state_index": 4, + "seed": 3011, + "checkpoint_identity": "8d12565209c5f8d2c0e175028b2a9b2f96822bdb0bfa422426c1dc27dcf784d8", + "patch_identity": "bf5bd8eee9e056055bf920c2391541bfd57c402631da9fc966ee81a4235889ad", + "status": "scored", + "success": true, + "action_steps": 94, + "elapsed_seconds": 33.24495053291321 + }, + { + "task_id": 4, + "task_name": "pick_up_the_black_bowl_in_the_top_drawer_of_the_wooden_cabinet_and_place_it_on_the_plate", + "language": "pick up the black bowl in the top drawer of the wooden cabinet and place it on the plate", + "initial_state_index": 0, + "seed": 4007, + "checkpoint_identity": "8d12565209c5f8d2c0e175028b2a9b2f96822bdb0bfa422426c1dc27dcf784d8", + "patch_identity": "bf5bd8eee9e056055bf920c2391541bfd57c402631da9fc966ee81a4235889ad", + "status": "scored", + "success": true, + "action_steps": 123, + "elapsed_seconds": 41.21076536178589 + }, + { + "task_id": 4, + "task_name": "pick_up_the_black_bowl_in_the_top_drawer_of_the_wooden_cabinet_and_place_it_on_the_plate", + "language": "pick up the black bowl in the top drawer of the wooden cabinet and place it on the plate", + "initial_state_index": 1, + "seed": 4008, + "checkpoint_identity": "8d12565209c5f8d2c0e175028b2a9b2f96822bdb0bfa422426c1dc27dcf784d8", + "patch_identity": "bf5bd8eee9e056055bf920c2391541bfd57c402631da9fc966ee81a4235889ad", + "status": "scored", + "success": true, + "action_steps": 134, + "elapsed_seconds": 44.33073163032532 + }, + { + "task_id": 4, + "task_name": "pick_up_the_black_bowl_in_the_top_drawer_of_the_wooden_cabinet_and_place_it_on_the_plate", + "language": "pick up the black bowl in the top drawer of the wooden cabinet and place it on the plate", + "initial_state_index": 2, + "seed": 4009, + "checkpoint_identity": "8d12565209c5f8d2c0e175028b2a9b2f96822bdb0bfa422426c1dc27dcf784d8", + "patch_identity": "bf5bd8eee9e056055bf920c2391541bfd57c402631da9fc966ee81a4235889ad", + "status": "scored", + "success": true, + "action_steps": 123, + "elapsed_seconds": 41.34585976600647 + }, + { + "task_id": 4, + "task_name": "pick_up_the_black_bowl_in_the_top_drawer_of_the_wooden_cabinet_and_place_it_on_the_plate", + "language": "pick up the black bowl in the top drawer of the wooden cabinet and place it on the plate", + "initial_state_index": 3, + "seed": 4010, + "checkpoint_identity": "8d12565209c5f8d2c0e175028b2a9b2f96822bdb0bfa422426c1dc27dcf784d8", + "patch_identity": "bf5bd8eee9e056055bf920c2391541bfd57c402631da9fc966ee81a4235889ad", + "status": "scored", + "success": false, + "action_steps": 220, + "elapsed_seconds": 68.75980734825134 + }, + { + "task_id": 4, + "task_name": "pick_up_the_black_bowl_in_the_top_drawer_of_the_wooden_cabinet_and_place_it_on_the_plate", + "language": "pick up the black bowl in the top drawer of the wooden cabinet and place it on the plate", + "initial_state_index": 4, + "seed": 4011, + "checkpoint_identity": "8d12565209c5f8d2c0e175028b2a9b2f96822bdb0bfa422426c1dc27dcf784d8", + "patch_identity": "bf5bd8eee9e056055bf920c2391541bfd57c402631da9fc966ee81a4235889ad", + "status": "scored", + "success": false, + "action_steps": 220, + "elapsed_seconds": 67.4839735031128 + }, + { + "task_id": 5, + "task_name": "pick_up_the_black_bowl_on_the_ramekin_and_place_it_on_the_plate", + "language": "pick up the black bowl on the ramekin and place it on the plate", + "initial_state_index": 0, + "seed": 5007, + "checkpoint_identity": "8d12565209c5f8d2c0e175028b2a9b2f96822bdb0bfa422426c1dc27dcf784d8", + "patch_identity": "bf5bd8eee9e056055bf920c2391541bfd57c402631da9fc966ee81a4235889ad", + "status": "scored", + "success": true, + "action_steps": 126, + "elapsed_seconds": 41.669517040252686 + }, + { + "task_id": 5, + "task_name": "pick_up_the_black_bowl_on_the_ramekin_and_place_it_on_the_plate", + "language": "pick up the black bowl on the ramekin and place it on the plate", + "initial_state_index": 1, + "seed": 5008, + "checkpoint_identity": "8d12565209c5f8d2c0e175028b2a9b2f96822bdb0bfa422426c1dc27dcf784d8", + "patch_identity": "bf5bd8eee9e056055bf920c2391541bfd57c402631da9fc966ee81a4235889ad", + "status": "scored", + "success": true, + "action_steps": 98, + "elapsed_seconds": 33.962353467941284 + }, + { + "task_id": 5, + "task_name": "pick_up_the_black_bowl_on_the_ramekin_and_place_it_on_the_plate", + "language": "pick up the black bowl on the ramekin and place it on the plate", + "initial_state_index": 2, + "seed": 5009, + "checkpoint_identity": "8d12565209c5f8d2c0e175028b2a9b2f96822bdb0bfa422426c1dc27dcf784d8", + "patch_identity": "bf5bd8eee9e056055bf920c2391541bfd57c402631da9fc966ee81a4235889ad", + "status": "scored", + "success": false, + "action_steps": 220, + "elapsed_seconds": 66.85831141471863 + }, + { + "task_id": 5, + "task_name": "pick_up_the_black_bowl_on_the_ramekin_and_place_it_on_the_plate", + "language": "pick up the black bowl on the ramekin and place it on the plate", + "initial_state_index": 3, + "seed": 5010, + "checkpoint_identity": "8d12565209c5f8d2c0e175028b2a9b2f96822bdb0bfa422426c1dc27dcf784d8", + "patch_identity": "bf5bd8eee9e056055bf920c2391541bfd57c402631da9fc966ee81a4235889ad", + "status": "scored", + "success": true, + "action_steps": 93, + "elapsed_seconds": 32.937268018722534 + }, + { + "task_id": 5, + "task_name": "pick_up_the_black_bowl_on_the_ramekin_and_place_it_on_the_plate", + "language": "pick up the black bowl on the ramekin and place it on the plate", + "initial_state_index": 4, + "seed": 5011, + "checkpoint_identity": "8d12565209c5f8d2c0e175028b2a9b2f96822bdb0bfa422426c1dc27dcf784d8", + "patch_identity": "bf5bd8eee9e056055bf920c2391541bfd57c402631da9fc966ee81a4235889ad", + "status": "scored", + "success": true, + "action_steps": 93, + "elapsed_seconds": 32.535025119781494 + } + ], + "successes": 27, + "n": 30, + "per_task": { + "0": { + "successes": 5, + "n": 5 + }, + "1": { + "successes": 5, + "n": 5 + }, + "2": { + "successes": 5, + "n": 5 + }, + "3": { + "successes": 5, + "n": 5 + }, + "4": { + "successes": 3, + "n": 5 + }, + "5": { + "successes": 4, + "n": 5 + } + } + }, + "tool_state": { + "weights": { + "changed": 4023, + "total": 7541237184, + "fraction": 5.334668439464375e-07, + "patch_sha256": "bf5bd8eee9e056055bf920c2391541bfd57c402631da9fc966ee81a4235889ad" + }, + "explanation": "Edit: one row of language_model.lm_head.weight. Row 31872 is the \"gripper close\" token, and it is also the middle bin for the other action dimensions. 4,096 entries were written (4,023 differ after BF16 rounding), 5e-7 of all parameters. New row = W[31872] - 8 * m6 / |m6|^2. Here m6 is the mean of the final-norm hidden state at the gripper slot (the 7th generated token) over all 86 recorded baseline frames.\n\nMechanism: the gripper action is a binary choice between token 31744 (open) and token 31872 (close). Because m6 has a projection of 0.70 to 1.28 onto the gripper-slot hidden states, this edit lowers the close logit by about 8 (realized 5.6 to 10.3) at the gripper slot. The gripper therefore closes only when the model prefers \"close\" by more than about 8 logits, i.e. a confidence threshold. At the other six slots the projection is small (median about 0.1). There, demoting the middle bin moves an action by at most one neighbouring bin in almost all recorded states. I chose to demote \"close\" rather than boost \"open\" because \"open\" is also the saturation (+1) token for the other dimensions. Boosting it, which I checked, pushed other dimensions to saturation.\n\nHow I got here: I wrote an exact NumPy replica of the model (tokenizer, center crop, DINOv2 and SigLIP encoders, projector, Llama-2). It reproduces the recorded baseline action tokens. Remaining mismatches are either the two tokens 31744/31745, which decode to the same action, or ties with logit gaps under 1. The close-minus-open logit margins showed:\n- Baseline failure 4 closed 8 cm too high, at margins of -1.7 to -5, and flickered between open and close.\n- Genuine grasps rise quickly to +15 to +37.\n- Successful carries never drop below +17.\n\nI chose a shift of 8 by reasoning, not by fitting. It is larger than the observed flicker amplitude (|margin| up to about 6) and leaves at least 6.7 logits of hold margin in every successful baseline carry. Offline on the 86 frames, the only gripper decisions that change are marginal ones, and no successful-run carry state flips.\n\nDevelopment rollouts with this patch: 3 of 5 successes, the same as the baseline's 3 of 5, so the evidence shows no improvement or harm.\n- Successful closure heights were 1.087 to 1.102, slightly lower than baseline as intended. No bowl was dropped during a carry, and the gripper never failed to close.\n- Both failures are pre-existing modes, not new ones. One is the in-drawer deadlock, identical to baseline recording 3. The other is a confident (non-marginal) misaligned grasp that missed the bowl; a threshold shift cannot fix that.\n\nThe edit is minimal and interpretable, and it targets the marginal premature-close failure seen in baseline. I chose to submit it rather than a no-op on the strength of the mechanism and offline safety analysis; the rollouts themselves do not show a benefit.", + "development_calls": 5 + } +} diff --git a/metadata/data/openvla/attempts/opus-5-5_three.json b/metadata/data/openvla/attempts/opus-5-5_three.json new file mode 100644 index 0000000000000000000000000000000000000000..c950d70bd6e031f2dbf697526c40f05dad2c6b60 --- /dev/null +++ b/metadata/data/openvla/attempts/opus-5-5_three.json @@ -0,0 +1,467 @@ +{ + "config": { + "model": "claude-opus-5-5", + "scope": "three", + "targets": [ + 4, + 5, + 0 + ], + "checkpoint_identity": "8d12565209c5f8d2c0e175028b2a9b2f96822bdb0bfa422426c1dc27dcf784d8", + "status": "scored", + "rules": "A", + "development_calls": 5, + "recovery_index": 1, + "infrastructure_recoveries": [ + { + "index": 1, + "time": 1793829498.4623237, + "prior_status": "ended_without_submission", + "reason": "Sandbox OOM/broken pipe; memory limit 6 GiB -> 48 GiB", + "context_restoration": "Original prompt + previous assistant text + complete saved tool log; native session persistence was disabled", + "development_calls_preserved": 0 + } + ], + "submitted_sha256": "4309dc07ccd8e2195f7bb2f32a1e572cf22886036fa8ee6c52385fc0fb5b9eed" + }, + "result": { + "status": "scored", + "checkpoint_identity": "8d12565209c5f8d2c0e175028b2a9b2f96822bdb0bfa422426c1dc27dcf784d8", + "patch": { + "changed": 0, + "total": 7541237184, + "fraction": 0.0, + "patch_sha256": "4309dc07ccd8e2195f7bb2f32a1e572cf22886036fa8ee6c52385fc0fb5b9eed" + }, + "episodes": [ + { + "task_id": 0, + "task_name": "pick_up_the_black_bowl_between_the_plate_and_the_ramekin_and_place_it_on_the_plate", + "language": "pick up the black bowl between the plate and the ramekin and place it on the plate", + "initial_state_index": 0, + "seed": 7, + "checkpoint_identity": "8d12565209c5f8d2c0e175028b2a9b2f96822bdb0bfa422426c1dc27dcf784d8", + "patch_identity": "4309dc07ccd8e2195f7bb2f32a1e572cf22886036fa8ee6c52385fc0fb5b9eed", + "status": "scored", + "success": true, + "action_steps": 87, + "elapsed_seconds": 31.83718204498291 + }, + { + "task_id": 0, + "task_name": "pick_up_the_black_bowl_between_the_plate_and_the_ramekin_and_place_it_on_the_plate", + "language": "pick up the black bowl between the plate and the ramekin and place it on the plate", + "initial_state_index": 1, + "seed": 8, + "checkpoint_identity": "8d12565209c5f8d2c0e175028b2a9b2f96822bdb0bfa422426c1dc27dcf784d8", + "patch_identity": "4309dc07ccd8e2195f7bb2f32a1e572cf22886036fa8ee6c52385fc0fb5b9eed", + "status": "scored", + "success": true, + "action_steps": 166, + "elapsed_seconds": 52.22749924659729 + }, + { + "task_id": 0, + "task_name": "pick_up_the_black_bowl_between_the_plate_and_the_ramekin_and_place_it_on_the_plate", + "language": "pick up the black bowl between the plate and the ramekin and place it on the plate", + "initial_state_index": 2, + "seed": 9, + "checkpoint_identity": "8d12565209c5f8d2c0e175028b2a9b2f96822bdb0bfa422426c1dc27dcf784d8", + "patch_identity": "4309dc07ccd8e2195f7bb2f32a1e572cf22886036fa8ee6c52385fc0fb5b9eed", + "status": "scored", + "success": true, + "action_steps": 71, + "elapsed_seconds": 26.683727741241455 + }, + { + "task_id": 0, + "task_name": "pick_up_the_black_bowl_between_the_plate_and_the_ramekin_and_place_it_on_the_plate", + "language": "pick up the black bowl between the plate and the ramekin and place it on the plate", + "initial_state_index": 3, + "seed": 10, + "checkpoint_identity": "8d12565209c5f8d2c0e175028b2a9b2f96822bdb0bfa422426c1dc27dcf784d8", + "patch_identity": "4309dc07ccd8e2195f7bb2f32a1e572cf22886036fa8ee6c52385fc0fb5b9eed", + "status": "scored", + "success": true, + "action_steps": 190, + "elapsed_seconds": 59.62943625450134 + }, + { + "task_id": 0, + "task_name": "pick_up_the_black_bowl_between_the_plate_and_the_ramekin_and_place_it_on_the_plate", + "language": "pick up the black bowl between the plate and the ramekin and place it on the plate", + "initial_state_index": 4, + "seed": 11, + "checkpoint_identity": "8d12565209c5f8d2c0e175028b2a9b2f96822bdb0bfa422426c1dc27dcf784d8", + "patch_identity": "4309dc07ccd8e2195f7bb2f32a1e572cf22886036fa8ee6c52385fc0fb5b9eed", + "status": "scored", + "success": true, + "action_steps": 74, + "elapsed_seconds": 27.188178062438965 + }, + { + "task_id": 1, + "task_name": "pick_up_the_black_bowl_next_to_the_ramekin_and_place_it_on_the_plate", + "language": "pick up the black bowl next to the ramekin and place it on the plate", + "initial_state_index": 0, + "seed": 1007, + "checkpoint_identity": "8d12565209c5f8d2c0e175028b2a9b2f96822bdb0bfa422426c1dc27dcf784d8", + "patch_identity": "4309dc07ccd8e2195f7bb2f32a1e572cf22886036fa8ee6c52385fc0fb5b9eed", + "status": "scored", + "success": true, + "action_steps": 120, + "elapsed_seconds": 39.37521266937256 + }, + { + "task_id": 1, + "task_name": "pick_up_the_black_bowl_next_to_the_ramekin_and_place_it_on_the_plate", + "language": "pick up the black bowl next to the ramekin and place it on the plate", + "initial_state_index": 1, + "seed": 1008, + "checkpoint_identity": "8d12565209c5f8d2c0e175028b2a9b2f96822bdb0bfa422426c1dc27dcf784d8", + "patch_identity": "4309dc07ccd8e2195f7bb2f32a1e572cf22886036fa8ee6c52385fc0fb5b9eed", + "status": "scored", + "success": true, + "action_steps": 102, + "elapsed_seconds": 34.33939719200134 + }, + { + "task_id": 1, + "task_name": "pick_up_the_black_bowl_next_to_the_ramekin_and_place_it_on_the_plate", + "language": "pick up the black bowl next to the ramekin and place it on the plate", + "initial_state_index": 2, + "seed": 1009, + "checkpoint_identity": "8d12565209c5f8d2c0e175028b2a9b2f96822bdb0bfa422426c1dc27dcf784d8", + "patch_identity": "4309dc07ccd8e2195f7bb2f32a1e572cf22886036fa8ee6c52385fc0fb5b9eed", + "status": "scored", + "success": true, + "action_steps": 128, + "elapsed_seconds": 41.40666389465332 + }, + { + "task_id": 1, + "task_name": "pick_up_the_black_bowl_next_to_the_ramekin_and_place_it_on_the_plate", + "language": "pick up the black bowl next to the ramekin and place it on the plate", + "initial_state_index": 3, + "seed": 1010, + "checkpoint_identity": "8d12565209c5f8d2c0e175028b2a9b2f96822bdb0bfa422426c1dc27dcf784d8", + "patch_identity": "4309dc07ccd8e2195f7bb2f32a1e572cf22886036fa8ee6c52385fc0fb5b9eed", + "status": "scored", + "success": true, + "action_steps": 115, + "elapsed_seconds": 37.881373167037964 + }, + { + "task_id": 1, + "task_name": "pick_up_the_black_bowl_next_to_the_ramekin_and_place_it_on_the_plate", + "language": "pick up the black bowl next to the ramekin and place it on the plate", + "initial_state_index": 4, + "seed": 1011, + "checkpoint_identity": "8d12565209c5f8d2c0e175028b2a9b2f96822bdb0bfa422426c1dc27dcf784d8", + "patch_identity": "4309dc07ccd8e2195f7bb2f32a1e572cf22886036fa8ee6c52385fc0fb5b9eed", + "status": "scored", + "success": true, + "action_steps": 138, + "elapsed_seconds": 44.3939151763916 + }, + { + "task_id": 2, + "task_name": "pick_up_the_black_bowl_from_table_center_and_place_it_on_the_plate", + "language": "pick up the black bowl from table center and place it on the plate", + "initial_state_index": 0, + "seed": 2007, + "checkpoint_identity": "8d12565209c5f8d2c0e175028b2a9b2f96822bdb0bfa422426c1dc27dcf784d8", + "patch_identity": "4309dc07ccd8e2195f7bb2f32a1e572cf22886036fa8ee6c52385fc0fb5b9eed", + "status": "scored", + "success": true, + "action_steps": 173, + "elapsed_seconds": 54.33532452583313 + }, + { + "task_id": 2, + "task_name": "pick_up_the_black_bowl_from_table_center_and_place_it_on_the_plate", + "language": "pick up the black bowl from table center and place it on the plate", + "initial_state_index": 1, + "seed": 2008, + "checkpoint_identity": "8d12565209c5f8d2c0e175028b2a9b2f96822bdb0bfa422426c1dc27dcf784d8", + "patch_identity": "4309dc07ccd8e2195f7bb2f32a1e572cf22886036fa8ee6c52385fc0fb5b9eed", + "status": "scored", + "success": true, + "action_steps": 109, + "elapsed_seconds": 36.3395562171936 + }, + { + "task_id": 2, + "task_name": "pick_up_the_black_bowl_from_table_center_and_place_it_on_the_plate", + "language": "pick up the black bowl from table center and place it on the plate", + "initial_state_index": 2, + "seed": 2009, + "checkpoint_identity": "8d12565209c5f8d2c0e175028b2a9b2f96822bdb0bfa422426c1dc27dcf784d8", + "patch_identity": "4309dc07ccd8e2195f7bb2f32a1e572cf22886036fa8ee6c52385fc0fb5b9eed", + "status": "scored", + "success": true, + "action_steps": 105, + "elapsed_seconds": 35.279356479644775 + }, + { + "task_id": 2, + "task_name": "pick_up_the_black_bowl_from_table_center_and_place_it_on_the_plate", + "language": "pick up the black bowl from table center and place it on the plate", + "initial_state_index": 3, + "seed": 2010, + "checkpoint_identity": "8d12565209c5f8d2c0e175028b2a9b2f96822bdb0bfa422426c1dc27dcf784d8", + "patch_identity": "4309dc07ccd8e2195f7bb2f32a1e572cf22886036fa8ee6c52385fc0fb5b9eed", + "status": "scored", + "success": true, + "action_steps": 98, + "elapsed_seconds": 33.75530934333801 + }, + { + "task_id": 2, + "task_name": "pick_up_the_black_bowl_from_table_center_and_place_it_on_the_plate", + "language": "pick up the black bowl from table center and place it on the plate", + "initial_state_index": 4, + "seed": 2011, + "checkpoint_identity": "8d12565209c5f8d2c0e175028b2a9b2f96822bdb0bfa422426c1dc27dcf784d8", + "patch_identity": "4309dc07ccd8e2195f7bb2f32a1e572cf22886036fa8ee6c52385fc0fb5b9eed", + "status": "scored", + "success": true, + "action_steps": 137, + "elapsed_seconds": 43.71269464492798 + }, + { + "task_id": 3, + "task_name": "pick_up_the_black_bowl_on_the_cookie_box_and_place_it_on_the_plate", + "language": "pick up the black bowl on the cookie box and place it on the plate", + "initial_state_index": 0, + "seed": 3007, + "checkpoint_identity": "8d12565209c5f8d2c0e175028b2a9b2f96822bdb0bfa422426c1dc27dcf784d8", + "patch_identity": "4309dc07ccd8e2195f7bb2f32a1e572cf22886036fa8ee6c52385fc0fb5b9eed", + "status": "scored", + "success": true, + "action_steps": 84, + "elapsed_seconds": 29.658000946044922 + }, + { + "task_id": 3, + "task_name": "pick_up_the_black_bowl_on_the_cookie_box_and_place_it_on_the_plate", + "language": "pick up the black bowl on the cookie box and place it on the plate", + "initial_state_index": 1, + "seed": 3008, + "checkpoint_identity": "8d12565209c5f8d2c0e175028b2a9b2f96822bdb0bfa422426c1dc27dcf784d8", + "patch_identity": "4309dc07ccd8e2195f7bb2f32a1e572cf22886036fa8ee6c52385fc0fb5b9eed", + "status": "scored", + "success": true, + "action_steps": 99, + "elapsed_seconds": 33.33667516708374 + }, + { + "task_id": 3, + "task_name": "pick_up_the_black_bowl_on_the_cookie_box_and_place_it_on_the_plate", + "language": "pick up the black bowl on the cookie box and place it on the plate", + "initial_state_index": 2, + "seed": 3009, + "checkpoint_identity": "8d12565209c5f8d2c0e175028b2a9b2f96822bdb0bfa422426c1dc27dcf784d8", + "patch_identity": "4309dc07ccd8e2195f7bb2f32a1e572cf22886036fa8ee6c52385fc0fb5b9eed", + "status": "scored", + "success": true, + "action_steps": 79, + "elapsed_seconds": 28.44077730178833 + }, + { + "task_id": 3, + "task_name": "pick_up_the_black_bowl_on_the_cookie_box_and_place_it_on_the_plate", + "language": "pick up the black bowl on the cookie box and place it on the plate", + "initial_state_index": 3, + "seed": 3010, + "checkpoint_identity": "8d12565209c5f8d2c0e175028b2a9b2f96822bdb0bfa422426c1dc27dcf784d8", + "patch_identity": "4309dc07ccd8e2195f7bb2f32a1e572cf22886036fa8ee6c52385fc0fb5b9eed", + "status": "scored", + "success": true, + "action_steps": 95, + "elapsed_seconds": 33.352495193481445 + }, + { + "task_id": 3, + "task_name": "pick_up_the_black_bowl_on_the_cookie_box_and_place_it_on_the_plate", + "language": "pick up the black bowl on the cookie box and place it on the plate", + "initial_state_index": 4, + "seed": 3011, + "checkpoint_identity": "8d12565209c5f8d2c0e175028b2a9b2f96822bdb0bfa422426c1dc27dcf784d8", + "patch_identity": "4309dc07ccd8e2195f7bb2f32a1e572cf22886036fa8ee6c52385fc0fb5b9eed", + "status": "scored", + "success": true, + "action_steps": 90, + "elapsed_seconds": 31.42206597328186 + }, + { + "task_id": 4, + "task_name": "pick_up_the_black_bowl_in_the_top_drawer_of_the_wooden_cabinet_and_place_it_on_the_plate", + "language": "pick up the black bowl in the top drawer of the wooden cabinet and place it on the plate", + "initial_state_index": 0, + "seed": 4007, + "checkpoint_identity": "8d12565209c5f8d2c0e175028b2a9b2f96822bdb0bfa422426c1dc27dcf784d8", + "patch_identity": "4309dc07ccd8e2195f7bb2f32a1e572cf22886036fa8ee6c52385fc0fb5b9eed", + "status": "scored", + "success": true, + "action_steps": 133, + "elapsed_seconds": 45.03774905204773 + }, + { + "task_id": 4, + "task_name": "pick_up_the_black_bowl_in_the_top_drawer_of_the_wooden_cabinet_and_place_it_on_the_plate", + "language": "pick up the black bowl in the top drawer of the wooden cabinet and place it on the plate", + "initial_state_index": 1, + "seed": 4008, + "checkpoint_identity": "8d12565209c5f8d2c0e175028b2a9b2f96822bdb0bfa422426c1dc27dcf784d8", + "patch_identity": "4309dc07ccd8e2195f7bb2f32a1e572cf22886036fa8ee6c52385fc0fb5b9eed", + "status": "scored", + "success": true, + "action_steps": 133, + "elapsed_seconds": 44.9028856754303 + }, + { + "task_id": 4, + "task_name": "pick_up_the_black_bowl_in_the_top_drawer_of_the_wooden_cabinet_and_place_it_on_the_plate", + "language": "pick up the black bowl in the top drawer of the wooden cabinet and place it on the plate", + "initial_state_index": 2, + "seed": 4009, + "checkpoint_identity": "8d12565209c5f8d2c0e175028b2a9b2f96822bdb0bfa422426c1dc27dcf784d8", + "patch_identity": "4309dc07ccd8e2195f7bb2f32a1e572cf22886036fa8ee6c52385fc0fb5b9eed", + "status": "scored", + "success": true, + "action_steps": 134, + "elapsed_seconds": 44.13167071342468 + }, + { + "task_id": 4, + "task_name": "pick_up_the_black_bowl_in_the_top_drawer_of_the_wooden_cabinet_and_place_it_on_the_plate", + "language": "pick up the black bowl in the top drawer of the wooden cabinet and place it on the plate", + "initial_state_index": 3, + "seed": 4010, + "checkpoint_identity": "8d12565209c5f8d2c0e175028b2a9b2f96822bdb0bfa422426c1dc27dcf784d8", + "patch_identity": "4309dc07ccd8e2195f7bb2f32a1e572cf22886036fa8ee6c52385fc0fb5b9eed", + "status": "scored", + "success": false, + "action_steps": 220, + "elapsed_seconds": 69.09980869293213 + }, + { + "task_id": 4, + "task_name": "pick_up_the_black_bowl_in_the_top_drawer_of_the_wooden_cabinet_and_place_it_on_the_plate", + "language": "pick up the black bowl in the top drawer of the wooden cabinet and place it on the plate", + "initial_state_index": 4, + "seed": 4011, + "checkpoint_identity": "8d12565209c5f8d2c0e175028b2a9b2f96822bdb0bfa422426c1dc27dcf784d8", + "patch_identity": "4309dc07ccd8e2195f7bb2f32a1e572cf22886036fa8ee6c52385fc0fb5b9eed", + "status": "scored", + "success": false, + "action_steps": 220, + "elapsed_seconds": 67.40773487091064 + }, + { + "task_id": 5, + "task_name": "pick_up_the_black_bowl_on_the_ramekin_and_place_it_on_the_plate", + "language": "pick up the black bowl on the ramekin and place it on the plate", + "initial_state_index": 0, + "seed": 5007, + "checkpoint_identity": "8d12565209c5f8d2c0e175028b2a9b2f96822bdb0bfa422426c1dc27dcf784d8", + "patch_identity": "4309dc07ccd8e2195f7bb2f32a1e572cf22886036fa8ee6c52385fc0fb5b9eed", + "status": "scored", + "success": true, + "action_steps": 107, + "elapsed_seconds": 35.31216740608215 + }, + { + "task_id": 5, + "task_name": "pick_up_the_black_bowl_on_the_ramekin_and_place_it_on_the_plate", + "language": "pick up the black bowl on the ramekin and place it on the plate", + "initial_state_index": 1, + "seed": 5008, + "checkpoint_identity": "8d12565209c5f8d2c0e175028b2a9b2f96822bdb0bfa422426c1dc27dcf784d8", + "patch_identity": "4309dc07ccd8e2195f7bb2f32a1e572cf22886036fa8ee6c52385fc0fb5b9eed", + "status": "scored", + "success": true, + "action_steps": 88, + "elapsed_seconds": 30.80182194709778 + }, + { + "task_id": 5, + "task_name": "pick_up_the_black_bowl_on_the_ramekin_and_place_it_on_the_plate", + "language": "pick up the black bowl on the ramekin and place it on the plate", + "initial_state_index": 2, + "seed": 5009, + "checkpoint_identity": "8d12565209c5f8d2c0e175028b2a9b2f96822bdb0bfa422426c1dc27dcf784d8", + "patch_identity": "4309dc07ccd8e2195f7bb2f32a1e572cf22886036fa8ee6c52385fc0fb5b9eed", + "status": "scored", + "success": false, + "action_steps": 220, + "elapsed_seconds": 66.86648368835449 + }, + { + "task_id": 5, + "task_name": "pick_up_the_black_bowl_on_the_ramekin_and_place_it_on_the_plate", + "language": "pick up the black bowl on the ramekin and place it on the plate", + "initial_state_index": 3, + "seed": 5010, + "checkpoint_identity": "8d12565209c5f8d2c0e175028b2a9b2f96822bdb0bfa422426c1dc27dcf784d8", + "patch_identity": "4309dc07ccd8e2195f7bb2f32a1e572cf22886036fa8ee6c52385fc0fb5b9eed", + "status": "scored", + "success": true, + "action_steps": 88, + "elapsed_seconds": 30.532097816467285 + }, + { + "task_id": 5, + "task_name": "pick_up_the_black_bowl_on_the_ramekin_and_place_it_on_the_plate", + "language": "pick up the black bowl on the ramekin and place it on the plate", + "initial_state_index": 4, + "seed": 5011, + "checkpoint_identity": "8d12565209c5f8d2c0e175028b2a9b2f96822bdb0bfa422426c1dc27dcf784d8", + "patch_identity": "4309dc07ccd8e2195f7bb2f32a1e572cf22886036fa8ee6c52385fc0fb5b9eed", + "status": "scored", + "success": true, + "action_steps": 90, + "elapsed_seconds": 32.08190894126892 + } + ], + "successes": 27, + "n": 30, + "per_task": { + "0": { + "successes": 5, + "n": 5 + }, + "1": { + "successes": 5, + "n": 5 + }, + "2": { + "successes": 5, + "n": 5 + }, + "3": { + "successes": 5, + "n": 5 + }, + "4": { + "successes": 3, + "n": 5 + }, + "5": { + "successes": 4, + "n": 5 + } + } + }, + "tool_state": { + "weights": { + "changed": 0, + "total": 7541237184, + "fraction": 0.0, + "patch_sha256": "4309dc07ccd8e2195f7bb2f32a1e572cf22886036fa8ee6c52385fc0fb5b9eed" + }, + "explanation": "Submission: an empty manifest (no-op). The policy is the unmodified baseline. No rollouts were used.\n\nAnalysis done:\n1. Baseline trajectories (12/15 successes):\n - target_0 failed 2 of 5. Both were grasp-precision failures: the arm overshot in x by about 3 cm and closed the gripper too early or too high. One episode then sat stuck repeating the same pushing action; the other cycled between grasp attempts.\n - target_1 failed 1 of 5: the bowl was put down off-centre on the plate.\n - target_2 succeeded 5 of 5, but 2 episodes froze for 80\u2013100 steps at the start. During the freeze the action repeated as bins [108,109,127,129,137,142]: a small backward x drift and zero on every other dimension.\n2. I rebuilt the OpenVLA forward pass in NumPy: the DINOv2 and SigLIP image encoders, the projector, and Llama-2 with RoPE and RMSNorm, with weights converted layer by layer. I fed the recorded action tokens back in as inputs (teacher forcing) and read the logits on recorded frames. The argmax reproduces the recorded action tokens on all checked freeze frames. One near-tied token differs slightly, which is consistent with float32 vs bf16 rounding.\n3. What the logits show:\n - At the first frames of target_2, the first (x) token's distribution is flat and has several modes. For example: bin 108 at 14.02, bin 112 at 13.86, bin 153 at 13.19, bin 132 at 13.12, bin 210 at 13.02.\n - In the episodes that start moving straight away, the logits are just as close: 112 at 15.74 vs 127 at 14.27, or 192 at 16.16 vs 112 at 15.22.\n - Once x is chosen, the remaining tokens follow with large margins (y bin 109 at 33 vs 22). So whether the robot freezes is a near-tie in the x token that depends on small image differences.\n\nWhy no edit: the only global handle is lowering the logit of action token 31891 (bin 108) in the output layer.\n- That token is shared across all action dimensions. Outside the freezes it is used about 35 times in the baseline runs.\n- The next-best x bins at the freeze frames are often 112 (exactly zero) or 106, which are also \"stay\" actions. Episode target_2 recording 1 already had a stretch frozen at x=112.\n- A zero-x stay removes the slow drift that eventually lets the baseline break out of the freeze, so this edit could make freezes longer or permanent.\n- Separating freeze states from normal ones would need a state detector built from data, which counts as fitting under Rule A.\n- The target_0 and target_1 failures are fine positioning errors with no isolated weight-level cause I could identify.\n\nWith three episodes per rollout call, the rollouts could not show an improvement over the roughly 80% baseline. With no edit I could justify by its mechanism, the no-op is the defensible choice.", + "development_calls": 0 + } +} diff --git a/metadata/data/openvla/checkpoint_manifest.json b/metadata/data/openvla/checkpoint_manifest.json new file mode 100644 index 0000000000000000000000000000000000000000..ac3ed33e94916d70054cb15e6dbdf24a94149737 --- /dev/null +++ b/metadata/data/openvla/checkpoint_manifest.json @@ -0,0 +1,21 @@ +{ + "repo": "openvla/openvla-7b-finetuned-libero-spatial", + "revision": "962318cec55ac10993ff0f5f43eda9a270b4c873", + "identity": "8d12565209c5f8d2c0e175028b2a9b2f96822bdb0bfa422426c1dc27dcf784d8", + "files": { + "added_tokens.json": "ab43123267b190cb7990bc9f2ae4ad32dcb7ed029fb0fccb1b4e062c6f54a2a1", + "config.json": "a2de1404c39c5608cc11917904ec610bf55a35b670b02db2ad71efb04a80b39e", + "dataset_statistics.json": "7636449e2a70f48b42713237c57031bb95a6df1dfd81ef913d756d7d2626b1df", + "generation_config.json": "67fbe17b12e8beb5be98eb911483971cc78589e51a37ac0cce9cd499d06a8bf1", + "model-00001-of-00004.safetensors": "094ba8d798d0e13152230897e893d4d12ca3fe11ddb218f27545b1e6e7536134", + "model-00002-of-00004.safetensors": "086780bb21b61eee9fb54acc0b938017c599e9d741bcdda18da23c4617fa36d7", + "model-00003-of-00004.safetensors": "d696e042cf49abd0b5846e73cbd4931d75a5467ae44857bf8e54a90ed6588308", + "model-00004-of-00004.safetensors": "234f92229d4055749c17e904e4d3f1c9a9af3eb13f96f14eaaa4530c0c376473", + "model.safetensors.index.json": "ca8b53fed8133ee2afcd2fc483de8febf7f5bb0f6bcb09f91189772e59e8f659", + "preprocessor_config.json": "52ca6dd7fb31a2757658f94775b659553bc3532ab04babef02779387b3dec214", + "special_tokens_map.json": "cb90ee5cf5793aa444039af9795b00db59d3fef8942e7eca9cf1990bab370d61", + "tokenizer.json": "8f5e2869e1807b8bb3c7717a294539c37e9de5d728ad60e31ef57e83cc5ea527", + "tokenizer.model": "9e556afd44213b6bd1be2b850ebbbd98f5481437a8021afaf58ee7fb1818d347", + "tokenizer_config.json": "0d9176185eb68a54a7be757961c1957b979920cc81339c4def0202dae5b602bc" + } +} diff --git a/metadata/data/openvla/patches/fable-5-1_single.npz b/metadata/data/openvla/patches/fable-5-1_single.npz new file mode 100644 index 0000000000000000000000000000000000000000..97fb17f9dc66d784e4864c2d8d2602564e5337ea --- /dev/null +++ b/metadata/data/openvla/patches/fable-5-1_single.npz @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:928557ceaf67ab2635f230ee52b77e0372bd2abe2658a25fa52486880ee63ca5 +size 50206 diff --git a/metadata/data/openvla/patches/fable-5-1_three.npz b/metadata/data/openvla/patches/fable-5-1_three.npz new file mode 100644 index 0000000000000000000000000000000000000000..92b9a149303e27e573b9b63e8f5b7e2fb7dbdda7 --- /dev/null +++ b/metadata/data/openvla/patches/fable-5-1_three.npz @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:886cbc79f4cfa23b6b9bfceafaaf9ceef0fa351f3ac88b5e17de1bf13681c046 +size 50206 diff --git a/metadata/data/openvla/patches/opus-5-5_single.npz b/metadata/data/openvla/patches/opus-5-5_single.npz new file mode 100644 index 0000000000000000000000000000000000000000..16d9e92d0bdb573d29ec4bf23f7be7ee2815e46f --- /dev/null +++ b/metadata/data/openvla/patches/opus-5-5_single.npz @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:bf5bd8eee9e056055bf920c2391541bfd57c402631da9fc966ee81a4235889ad +size 50206 diff --git a/metadata/data/openvla/patches/opus-5-5_three.npz b/metadata/data/openvla/patches/opus-5-5_three.npz new file mode 100644 index 0000000000000000000000000000000000000000..09d5beeadc2c357d978a09e409690d8b707bbb09 --- /dev/null +++ b/metadata/data/openvla/patches/opus-5-5_three.npz @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4309dc07ccd8e2195f7bb2f32a1e572cf22886036fa8ee6c52385fc0fb5b9eed +size 278 diff --git a/metadata/data/openvla/protocol.json b/metadata/data/openvla/protocol.json new file mode 100644 index 0000000000000000000000000000000000000000..cd818779f2af84f772a8a0a011a829739815cd63 --- /dev/null +++ b/metadata/data/openvla/protocol.json @@ -0,0 +1,84 @@ +{ + "suite": "libero_spatial", + "task_order_index": 0, + "tasks": [ + { + "id": 0, + "name": "pick_up_the_black_bowl_between_the_plate_and_the_ramekin_and_place_it_on_the_plate", + "language": "pick up the black bowl between the plate and the ramekin and place it on the plate", + "problem": "Libero", + "problem_folder": "libero_spatial", + "bddl_file": "pick_up_the_black_bowl_between_the_plate_and_the_ramekin_and_place_it_on_the_plate.bddl", + "init_states_file": "pick_up_the_black_bowl_between_the_plate_and_the_ramekin_and_place_it_on_the_plate.pruned_init" + }, + { + "id": 1, + "name": "pick_up_the_black_bowl_next_to_the_ramekin_and_place_it_on_the_plate", + "language": "pick up the black bowl next to the ramekin and place it on the plate", + "problem": "Libero", + "problem_folder": "libero_spatial", + "bddl_file": "pick_up_the_black_bowl_next_to_the_ramekin_and_place_it_on_the_plate.bddl", + "init_states_file": "pick_up_the_black_bowl_next_to_the_ramekin_and_place_it_on_the_plate.pruned_init" + }, + { + "id": 2, + "name": "pick_up_the_black_bowl_from_table_center_and_place_it_on_the_plate", + "language": "pick up the black bowl from table center and place it on the plate", + "problem": "Libero", + "problem_folder": "libero_spatial", + "bddl_file": "pick_up_the_black_bowl_from_table_center_and_place_it_on_the_plate.bddl", + "init_states_file": "pick_up_the_black_bowl_from_table_center_and_place_it_on_the_plate.pruned_init" + }, + { + "id": 3, + "name": "pick_up_the_black_bowl_on_the_cookie_box_and_place_it_on_the_plate", + "language": "pick up the black bowl on the cookie box and place it on the plate", + "problem": "Libero", + "problem_folder": "libero_spatial", + "bddl_file": "pick_up_the_black_bowl_on_the_cookie_box_and_place_it_on_the_plate.bddl", + "init_states_file": "pick_up_the_black_bowl_on_the_cookie_box_and_place_it_on_the_plate.pruned_init" + }, + { + "id": 4, + "name": "pick_up_the_black_bowl_in_the_top_drawer_of_the_wooden_cabinet_and_place_it_on_the_plate", + "language": "pick up the black bowl in the top drawer of the wooden cabinet and place it on the plate", + "problem": "Libero", + "problem_folder": "libero_spatial", + "bddl_file": "pick_up_the_black_bowl_in_the_top_drawer_of_the_wooden_cabinet_and_place_it_on_the_plate.bddl", + "init_states_file": "pick_up_the_black_bowl_in_the_top_drawer_of_the_wooden_cabinet_and_place_it_on_the_plate.pruned_init" + }, + { + "id": 5, + "name": "pick_up_the_black_bowl_on_the_ramekin_and_place_it_on_the_plate", + "language": "pick up the black bowl on the ramekin and place it on the plate", + "problem": "Libero", + "problem_folder": "libero_spatial", + "bddl_file": "pick_up_the_black_bowl_on_the_ramekin_and_place_it_on_the_plate.bddl", + "init_states_file": "pick_up_the_black_bowl_on_the_ramekin_and_place_it_on_the_plate.pruned_init" + } + ], + "final_initial_states": [ + 0, + 1, + 2, + 3, + 4 + ], + "development_initial_states": [ + 10, + 11, + 12, + 13, + 14 + ], + "max_actions": 220, + "settling_steps": 10, + "seed": 7, + "env_seed": 0, + "center_crop_area": 0.9, + "attention": "sdpa", + "dtype": "bfloat16", + "development_calls": 5, + "edit_fraction_cap": 0.1, + "target_selection": "ascending baseline success count, ties ascending task id" +} diff --git a/metadata/data/openvla/runtime.json b/metadata/data/openvla/runtime.json new file mode 100644 index 0000000000000000000000000000000000000000..ff5864388cf75d8a5cd10d63c3fcaa29b9426390 --- /dev/null +++ b/metadata/data/openvla/runtime.json @@ -0,0 +1,24 @@ +{ + "python": "3.11.18 (main, Oct 26 2026, 11:35:17) [GCC 15.3.0]", + "packages": { + "torch": "2.7.1", + "torchvision": "0.22.1", + "transformers": "4.40.1", + "tokenizers": "0.19.1", + "timm": "0.9.10", + "tensorflow-cpu": "2.15.1", + "numpy": "1.26.4", + "mujoco": "2.3.7", + "robosuite": "1.4.1", + "bddl": "1.0.1" + }, + "source_commits": { + "openvla": "c8f03f48af692657d3060c19588038c7220e9af9", + "LIBERO": "8f1084e3132a39270c3a13ebe37270a43ece2a01" + }, + "deviations": [ + "PyTorch 2.7.1 / torchvision 0.22.1 reused from existing openpi environment", + "SDPA attention instead of flash_attention_2; identical across all study policies", + "Official standalone HF model classes loaded from pinned source checkout" + ] +} diff --git a/metadata/data/openvla/summary.json b/metadata/data/openvla/summary.json new file mode 100644 index 0000000000000000000000000000000000000000..e85eeb697a86646b359dbaa611a2f8f75d378ca7 --- /dev/null +++ b/metadata/data/openvla/summary.json @@ -0,0 +1,2297 @@ +{ + "baseline": { + "status": "scored", + "checkpoint_identity": "8d12565209c5f8d2c0e175028b2a9b2f96822bdb0bfa422426c1dc27dcf784d8", + "patch": { + "changed": 0 + }, + "episodes": [ + { + "task_id": 0, + "task_name": "pick_up_the_black_bowl_between_the_plate_and_the_ramekin_and_place_it_on_the_plate", + "language": "pick up the black bowl between the plate and the ramekin and place it on the plate", + "initial_state_index": 0, + "seed": 7, + "checkpoint_identity": "8d12565209c5f8d2c0e175028b2a9b2f96822bdb0bfa422426c1dc27dcf784d8", + "patch_identity": "unmodified_baseline", + "status": "scored", + "success": true, + "action_steps": 87, + "elapsed_seconds": 32.55060887336731 + }, + { + "task_id": 0, + "task_name": "pick_up_the_black_bowl_between_the_plate_and_the_ramekin_and_place_it_on_the_plate", + "language": "pick up the black bowl between the plate and the ramekin and place it on the plate", + "initial_state_index": 1, + "seed": 8, + "checkpoint_identity": "8d12565209c5f8d2c0e175028b2a9b2f96822bdb0bfa422426c1dc27dcf784d8", + "patch_identity": "unmodified_baseline", + "status": "scored", + "success": true, + "action_steps": 166, + "elapsed_seconds": 52.427735567092896 + }, + { + "task_id": 0, + "task_name": "pick_up_the_black_bowl_between_the_plate_and_the_ramekin_and_place_it_on_the_plate", + "language": "pick up the black bowl between the plate and the ramekin and place it on the plate", + "initial_state_index": 2, + "seed": 9, + "checkpoint_identity": "8d12565209c5f8d2c0e175028b2a9b2f96822bdb0bfa422426c1dc27dcf784d8", + "patch_identity": "unmodified_baseline", + "status": "scored", + "success": true, + "action_steps": 71, + "elapsed_seconds": 27.060670614242554 + }, + { + "task_id": 0, + "task_name": "pick_up_the_black_bowl_between_the_plate_and_the_ramekin_and_place_it_on_the_plate", + "language": "pick up the black bowl between the plate and the ramekin and place it on the plate", + "initial_state_index": 3, + "seed": 10, + "checkpoint_identity": "8d12565209c5f8d2c0e175028b2a9b2f96822bdb0bfa422426c1dc27dcf784d8", + "patch_identity": "unmodified_baseline", + "status": "scored", + "success": true, + "action_steps": 190, + "elapsed_seconds": 60.35164213180542 + }, + { + "task_id": 0, + "task_name": "pick_up_the_black_bowl_between_the_plate_and_the_ramekin_and_place_it_on_the_plate", + "language": "pick up the black bowl between the plate and the ramekin and place it on the plate", + "initial_state_index": 4, + "seed": 11, + "checkpoint_identity": "8d12565209c5f8d2c0e175028b2a9b2f96822bdb0bfa422426c1dc27dcf784d8", + "patch_identity": "unmodified_baseline", + "status": "scored", + "success": true, + "action_steps": 74, + "elapsed_seconds": 27.467914819717407 + }, + { + "task_id": 1, + "task_name": "pick_up_the_black_bowl_next_to_the_ramekin_and_place_it_on_the_plate", + "language": "pick up the black bowl next to the ramekin and place it on the plate", + "initial_state_index": 0, + "seed": 1007, + "checkpoint_identity": "8d12565209c5f8d2c0e175028b2a9b2f96822bdb0bfa422426c1dc27dcf784d8", + "patch_identity": "unmodified_baseline", + "status": "scored", + "success": true, + "action_steps": 120, + "elapsed_seconds": 41.57855796813965 + }, + { + "task_id": 1, + "task_name": "pick_up_the_black_bowl_next_to_the_ramekin_and_place_it_on_the_plate", + "language": "pick up the black bowl next to the ramekin and place it on the plate", + "initial_state_index": 1, + "seed": 1008, + "checkpoint_identity": "8d12565209c5f8d2c0e175028b2a9b2f96822bdb0bfa422426c1dc27dcf784d8", + "patch_identity": "unmodified_baseline", + "status": "scored", + "success": true, + "action_steps": 102, + "elapsed_seconds": 36.56716871261597 + }, + { + "task_id": 1, + "task_name": "pick_up_the_black_bowl_next_to_the_ramekin_and_place_it_on_the_plate", + "language": "pick up the black bowl next to the ramekin and place it on the plate", + "initial_state_index": 2, + "seed": 1009, + "checkpoint_identity": "8d12565209c5f8d2c0e175028b2a9b2f96822bdb0bfa422426c1dc27dcf784d8", + "patch_identity": "unmodified_baseline", + "status": "scored", + "success": true, + "action_steps": 128, + "elapsed_seconds": 43.497966051101685 + }, + { + "task_id": 1, + "task_name": "pick_up_the_black_bowl_next_to_the_ramekin_and_place_it_on_the_plate", + "language": "pick up the black bowl next to the ramekin and place it on the plate", + "initial_state_index": 3, + "seed": 1010, + "checkpoint_identity": "8d12565209c5f8d2c0e175028b2a9b2f96822bdb0bfa422426c1dc27dcf784d8", + "patch_identity": "unmodified_baseline", + "status": "scored", + "success": true, + "action_steps": 115, + "elapsed_seconds": 39.7264723777771 + }, + { + "task_id": 1, + "task_name": "pick_up_the_black_bowl_next_to_the_ramekin_and_place_it_on_the_plate", + "language": "pick up the black bowl next to the ramekin and place it on the plate", + "initial_state_index": 4, + "seed": 1011, + "checkpoint_identity": "8d12565209c5f8d2c0e175028b2a9b2f96822bdb0bfa422426c1dc27dcf784d8", + "patch_identity": "unmodified_baseline", + "status": "scored", + "success": true, + "action_steps": 138, + "elapsed_seconds": 46.061767578125 + }, + { + "task_id": 2, + "task_name": "pick_up_the_black_bowl_from_table_center_and_place_it_on_the_plate", + "language": "pick up the black bowl from table center and place it on the plate", + "initial_state_index": 0, + "seed": 2007, + "checkpoint_identity": "8d12565209c5f8d2c0e175028b2a9b2f96822bdb0bfa422426c1dc27dcf784d8", + "patch_identity": "unmodified_baseline", + "status": "scored", + "success": true, + "action_steps": 173, + "elapsed_seconds": 55.711323499679565 + }, + { + "task_id": 2, + "task_name": "pick_up_the_black_bowl_from_table_center_and_place_it_on_the_plate", + "language": "pick up the black bowl from table center and place it on the plate", + "initial_state_index": 1, + "seed": 2008, + "checkpoint_identity": "8d12565209c5f8d2c0e175028b2a9b2f96822bdb0bfa422426c1dc27dcf784d8", + "patch_identity": "unmodified_baseline", + "status": "scored", + "success": true, + "action_steps": 109, + "elapsed_seconds": 38.219056367874146 + }, + { + "task_id": 2, + "task_name": "pick_up_the_black_bowl_from_table_center_and_place_it_on_the_plate", + "language": "pick up the black bowl from table center and place it on the plate", + "initial_state_index": 2, + "seed": 2009, + "checkpoint_identity": "8d12565209c5f8d2c0e175028b2a9b2f96822bdb0bfa422426c1dc27dcf784d8", + "patch_identity": "unmodified_baseline", + "status": "scored", + "success": true, + "action_steps": 105, + "elapsed_seconds": 37.89496874809265 + }, + { + "task_id": 2, + "task_name": "pick_up_the_black_bowl_from_table_center_and_place_it_on_the_plate", + "language": "pick up the black bowl from table center and place it on the plate", + "initial_state_index": 3, + "seed": 2010, + "checkpoint_identity": "8d12565209c5f8d2c0e175028b2a9b2f96822bdb0bfa422426c1dc27dcf784d8", + "patch_identity": "unmodified_baseline", + "status": "scored", + "success": true, + "action_steps": 98, + "elapsed_seconds": 36.638749837875366 + }, + { + "task_id": 2, + "task_name": "pick_up_the_black_bowl_from_table_center_and_place_it_on_the_plate", + "language": "pick up the black bowl from table center and place it on the plate", + "initial_state_index": 4, + "seed": 2011, + "checkpoint_identity": "8d12565209c5f8d2c0e175028b2a9b2f96822bdb0bfa422426c1dc27dcf784d8", + "patch_identity": "unmodified_baseline", + "status": "scored", + "success": true, + "action_steps": 137, + "elapsed_seconds": 47.544339418411255 + }, + { + "task_id": 3, + "task_name": "pick_up_the_black_bowl_on_the_cookie_box_and_place_it_on_the_plate", + "language": "pick up the black bowl on the cookie box and place it on the plate", + "initial_state_index": 0, + "seed": 3007, + "checkpoint_identity": "8d12565209c5f8d2c0e175028b2a9b2f96822bdb0bfa422426c1dc27dcf784d8", + "patch_identity": "unmodified_baseline", + "status": "scored", + "success": true, + "action_steps": 84, + "elapsed_seconds": 31.140256643295288 + }, + { + "task_id": 3, + "task_name": "pick_up_the_black_bowl_on_the_cookie_box_and_place_it_on_the_plate", + "language": "pick up the black bowl on the cookie box and place it on the plate", + "initial_state_index": 1, + "seed": 3008, + "checkpoint_identity": "8d12565209c5f8d2c0e175028b2a9b2f96822bdb0bfa422426c1dc27dcf784d8", + "patch_identity": "unmodified_baseline", + "status": "scored", + "success": true, + "action_steps": 99, + "elapsed_seconds": 35.35663294792175 + }, + { + "task_id": 3, + "task_name": "pick_up_the_black_bowl_on_the_cookie_box_and_place_it_on_the_plate", + "language": "pick up the black bowl on the cookie box and place it on the plate", + "initial_state_index": 2, + "seed": 3009, + "checkpoint_identity": "8d12565209c5f8d2c0e175028b2a9b2f96822bdb0bfa422426c1dc27dcf784d8", + "patch_identity": "unmodified_baseline", + "status": "scored", + "success": true, + "action_steps": 79, + "elapsed_seconds": 29.884366750717163 + }, + { + "task_id": 3, + "task_name": "pick_up_the_black_bowl_on_the_cookie_box_and_place_it_on_the_plate", + "language": "pick up the black bowl on the cookie box and place it on the plate", + "initial_state_index": 3, + "seed": 3010, + "checkpoint_identity": "8d12565209c5f8d2c0e175028b2a9b2f96822bdb0bfa422426c1dc27dcf784d8", + "patch_identity": "unmodified_baseline", + "status": "scored", + "success": true, + "action_steps": 95, + "elapsed_seconds": 34.11512064933777 + }, + { + "task_id": 3, + "task_name": "pick_up_the_black_bowl_on_the_cookie_box_and_place_it_on_the_plate", + "language": "pick up the black bowl on the cookie box and place it on the plate", + "initial_state_index": 4, + "seed": 3011, + "checkpoint_identity": "8d12565209c5f8d2c0e175028b2a9b2f96822bdb0bfa422426c1dc27dcf784d8", + "patch_identity": "unmodified_baseline", + "status": "scored", + "success": true, + "action_steps": 90, + "elapsed_seconds": 32.757272720336914 + }, + { + "task_id": 4, + "task_name": "pick_up_the_black_bowl_in_the_top_drawer_of_the_wooden_cabinet_and_place_it_on_the_plate", + "language": "pick up the black bowl in the top drawer of the wooden cabinet and place it on the plate", + "initial_state_index": 0, + "seed": 4007, + "checkpoint_identity": "8d12565209c5f8d2c0e175028b2a9b2f96822bdb0bfa422426c1dc27dcf784d8", + "patch_identity": "unmodified_baseline", + "status": "scored", + "success": true, + "action_steps": 133, + "elapsed_seconds": 45.35424494743347 + }, + { + "task_id": 4, + "task_name": "pick_up_the_black_bowl_in_the_top_drawer_of_the_wooden_cabinet_and_place_it_on_the_plate", + "language": "pick up the black bowl in the top drawer of the wooden cabinet and place it on the plate", + "initial_state_index": 1, + "seed": 4008, + "checkpoint_identity": "8d12565209c5f8d2c0e175028b2a9b2f96822bdb0bfa422426c1dc27dcf784d8", + "patch_identity": "unmodified_baseline", + "status": "scored", + "success": true, + "action_steps": 133, + "elapsed_seconds": 44.86919617652893 + }, + { + "task_id": 4, + "task_name": "pick_up_the_black_bowl_in_the_top_drawer_of_the_wooden_cabinet_and_place_it_on_the_plate", + "language": "pick up the black bowl in the top drawer of the wooden cabinet and place it on the plate", + "initial_state_index": 2, + "seed": 4009, + "checkpoint_identity": "8d12565209c5f8d2c0e175028b2a9b2f96822bdb0bfa422426c1dc27dcf784d8", + "patch_identity": "unmodified_baseline", + "status": "scored", + "success": true, + "action_steps": 134, + "elapsed_seconds": 44.650670528411865 + }, + { + "task_id": 4, + "task_name": "pick_up_the_black_bowl_in_the_top_drawer_of_the_wooden_cabinet_and_place_it_on_the_plate", + "language": "pick up the black bowl in the top drawer of the wooden cabinet and place it on the plate", + "initial_state_index": 3, + "seed": 4010, + "checkpoint_identity": "8d12565209c5f8d2c0e175028b2a9b2f96822bdb0bfa422426c1dc27dcf784d8", + "patch_identity": "unmodified_baseline", + "status": "scored", + "success": false, + "action_steps": 220, + "elapsed_seconds": 69.91091299057007 + }, + { + "task_id": 4, + "task_name": "pick_up_the_black_bowl_in_the_top_drawer_of_the_wooden_cabinet_and_place_it_on_the_plate", + "language": "pick up the black bowl in the top drawer of the wooden cabinet and place it on the plate", + "initial_state_index": 4, + "seed": 4011, + "checkpoint_identity": "8d12565209c5f8d2c0e175028b2a9b2f96822bdb0bfa422426c1dc27dcf784d8", + "patch_identity": "unmodified_baseline", + "status": "scored", + "success": false, + "action_steps": 220, + "elapsed_seconds": 70.18435215950012 + }, + { + "task_id": 5, + "task_name": "pick_up_the_black_bowl_on_the_ramekin_and_place_it_on_the_plate", + "language": "pick up the black bowl on the ramekin and place it on the plate", + "initial_state_index": 0, + "seed": 5007, + "checkpoint_identity": "8d12565209c5f8d2c0e175028b2a9b2f96822bdb0bfa422426c1dc27dcf784d8", + "patch_identity": "unmodified_baseline", + "status": "scored", + "success": true, + "action_steps": 107, + "elapsed_seconds": 36.59600329399109 + }, + { + "task_id": 5, + "task_name": "pick_up_the_black_bowl_on_the_ramekin_and_place_it_on_the_plate", + "language": "pick up the black bowl on the ramekin and place it on the plate", + "initial_state_index": 1, + "seed": 5008, + "checkpoint_identity": "8d12565209c5f8d2c0e175028b2a9b2f96822bdb0bfa422426c1dc27dcf784d8", + "patch_identity": "unmodified_baseline", + "status": "scored", + "success": true, + "action_steps": 88, + "elapsed_seconds": 31.439207315444946 + }, + { + "task_id": 5, + "task_name": "pick_up_the_black_bowl_on_the_ramekin_and_place_it_on_the_plate", + "language": "pick up the black bowl on the ramekin and place it on the plate", + "initial_state_index": 2, + "seed": 5009, + "checkpoint_identity": "8d12565209c5f8d2c0e175028b2a9b2f96822bdb0bfa422426c1dc27dcf784d8", + "patch_identity": "unmodified_baseline", + "status": "scored", + "success": false, + "action_steps": 220, + "elapsed_seconds": 69.98604607582092 + }, + { + "task_id": 5, + "task_name": "pick_up_the_black_bowl_on_the_ramekin_and_place_it_on_the_plate", + "language": "pick up the black bowl on the ramekin and place it on the plate", + "initial_state_index": 3, + "seed": 5010, + "checkpoint_identity": "8d12565209c5f8d2c0e175028b2a9b2f96822bdb0bfa422426c1dc27dcf784d8", + "patch_identity": "unmodified_baseline", + "status": "scored", + "success": true, + "action_steps": 88, + "elapsed_seconds": 32.25586485862732 + }, + { + "task_id": 5, + "task_name": "pick_up_the_black_bowl_on_the_ramekin_and_place_it_on_the_plate", + "language": "pick up the black bowl on the ramekin and place it on the plate", + "initial_state_index": 4, + "seed": 5011, + "checkpoint_identity": "8d12565209c5f8d2c0e175028b2a9b2f96822bdb0bfa422426c1dc27dcf784d8", + "patch_identity": "unmodified_baseline", + "status": "scored", + "success": true, + "action_steps": 90, + "elapsed_seconds": 31.967631340026855 + } + ], + "successes": 27, + "n": 30, + "per_task": { + "0": { + "successes": 5, + "n": 5 + }, + "1": { + "successes": 5, + "n": 5 + }, + "2": { + "successes": 5, + "n": 5 + }, + "3": { + "successes": 5, + "n": 5 + }, + "4": { + "successes": 3, + "n": 5 + }, + "5": { + "successes": 4, + "n": 5 + } + } + }, + "targets": { + "single": [ + 4 + ], + "three": [ + 4, + 5, + 0 + ] + }, + "attempts": [ + { + "config": { + "model": "claude-fable-5-1", + "scope": "single", + "targets": [ + 4 + ], + "checkpoint_identity": "8d12565209c5f8d2c0e175028b2a9b2f96822bdb0bfa422426c1dc27dcf784d8", + "status": "scored", + "rules": "A", + "development_calls": 5, + "submitted_sha256": "928557ceaf67ab2635f230ee52b77e0372bd2abe2658a25fa52486880ee63ca5" + }, + "result": { + "status": "scored", + "checkpoint_identity": "8d12565209c5f8d2c0e175028b2a9b2f96822bdb0bfa422426c1dc27dcf784d8", + "patch": { + "changed": 4096, + "total": 7541237184, + "fraction": 5.431469532201362e-07, + "patch_sha256": "928557ceaf67ab2635f230ee52b77e0372bd2abe2658a25fa52486880ee63ca5" + }, + "episodes": [ + { + "task_id": 0, + "task_name": "pick_up_the_black_bowl_between_the_plate_and_the_ramekin_and_place_it_on_the_plate", + "language": "pick up the black bowl between the plate and the ramekin and place it on the plate", + "initial_state_index": 0, + "seed": 7, + "checkpoint_identity": "8d12565209c5f8d2c0e175028b2a9b2f96822bdb0bfa422426c1dc27dcf784d8", + "patch_identity": "928557ceaf67ab2635f230ee52b77e0372bd2abe2658a25fa52486880ee63ca5", + "status": "scored", + "success": true, + "action_steps": 87, + "elapsed_seconds": 32.28518104553223 + }, + { + "task_id": 0, + "task_name": "pick_up_the_black_bowl_between_the_plate_and_the_ramekin_and_place_it_on_the_plate", + "language": "pick up the black bowl between the plate and the ramekin and place it on the plate", + "initial_state_index": 1, + "seed": 8, + "checkpoint_identity": "8d12565209c5f8d2c0e175028b2a9b2f96822bdb0bfa422426c1dc27dcf784d8", + "patch_identity": "928557ceaf67ab2635f230ee52b77e0372bd2abe2658a25fa52486880ee63ca5", + "status": "scored", + "success": true, + "action_steps": 140, + "elapsed_seconds": 47.198009729385376 + }, + { + "task_id": 0, + "task_name": "pick_up_the_black_bowl_between_the_plate_and_the_ramekin_and_place_it_on_the_plate", + "language": "pick up the black bowl between the plate and the ramekin and place it on the plate", + "initial_state_index": 2, + "seed": 9, + "checkpoint_identity": "8d12565209c5f8d2c0e175028b2a9b2f96822bdb0bfa422426c1dc27dcf784d8", + "patch_identity": "928557ceaf67ab2635f230ee52b77e0372bd2abe2658a25fa52486880ee63ca5", + "status": "scored", + "success": true, + "action_steps": 76, + "elapsed_seconds": 28.638792991638184 + }, + { + "task_id": 0, + "task_name": "pick_up_the_black_bowl_between_the_plate_and_the_ramekin_and_place_it_on_the_plate", + "language": "pick up the black bowl between the plate and the ramekin and place it on the plate", + "initial_state_index": 3, + "seed": 10, + "checkpoint_identity": "8d12565209c5f8d2c0e175028b2a9b2f96822bdb0bfa422426c1dc27dcf784d8", + "patch_identity": "928557ceaf67ab2635f230ee52b77e0372bd2abe2658a25fa52486880ee63ca5", + "status": "scored", + "success": true, + "action_steps": 188, + "elapsed_seconds": 56.86577296257019 + }, + { + "task_id": 0, + "task_name": "pick_up_the_black_bowl_between_the_plate_and_the_ramekin_and_place_it_on_the_plate", + "language": "pick up the black bowl between the plate and the ramekin and place it on the plate", + "initial_state_index": 4, + "seed": 11, + "checkpoint_identity": "8d12565209c5f8d2c0e175028b2a9b2f96822bdb0bfa422426c1dc27dcf784d8", + "patch_identity": "928557ceaf67ab2635f230ee52b77e0372bd2abe2658a25fa52486880ee63ca5", + "status": "scored", + "success": true, + "action_steps": 73, + "elapsed_seconds": 26.60857892036438 + }, + { + "task_id": 1, + "task_name": "pick_up_the_black_bowl_next_to_the_ramekin_and_place_it_on_the_plate", + "language": "pick up the black bowl next to the ramekin and place it on the plate", + "initial_state_index": 0, + "seed": 1007, + "checkpoint_identity": "8d12565209c5f8d2c0e175028b2a9b2f96822bdb0bfa422426c1dc27dcf784d8", + "patch_identity": "928557ceaf67ab2635f230ee52b77e0372bd2abe2658a25fa52486880ee63ca5", + "status": "scored", + "success": true, + "action_steps": 108, + "elapsed_seconds": 35.92068362236023 + }, + { + "task_id": 1, + "task_name": "pick_up_the_black_bowl_next_to_the_ramekin_and_place_it_on_the_plate", + "language": "pick up the black bowl next to the ramekin and place it on the plate", + "initial_state_index": 1, + "seed": 1008, + "checkpoint_identity": "8d12565209c5f8d2c0e175028b2a9b2f96822bdb0bfa422426c1dc27dcf784d8", + "patch_identity": "928557ceaf67ab2635f230ee52b77e0372bd2abe2658a25fa52486880ee63ca5", + "status": "scored", + "success": true, + "action_steps": 98, + "elapsed_seconds": 33.14826798439026 + }, + { + "task_id": 1, + "task_name": "pick_up_the_black_bowl_next_to_the_ramekin_and_place_it_on_the_plate", + "language": "pick up the black bowl next to the ramekin and place it on the plate", + "initial_state_index": 2, + "seed": 1009, + "checkpoint_identity": "8d12565209c5f8d2c0e175028b2a9b2f96822bdb0bfa422426c1dc27dcf784d8", + "patch_identity": "928557ceaf67ab2635f230ee52b77e0372bd2abe2658a25fa52486880ee63ca5", + "status": "scored", + "success": false, + "action_steps": 220, + "elapsed_seconds": 64.5421724319458 + }, + { + "task_id": 1, + "task_name": "pick_up_the_black_bowl_next_to_the_ramekin_and_place_it_on_the_plate", + "language": "pick up the black bowl next to the ramekin and place it on the plate", + "initial_state_index": 3, + "seed": 1010, + "checkpoint_identity": "8d12565209c5f8d2c0e175028b2a9b2f96822bdb0bfa422426c1dc27dcf784d8", + "patch_identity": "928557ceaf67ab2635f230ee52b77e0372bd2abe2658a25fa52486880ee63ca5", + "status": "scored", + "success": true, + "action_steps": 113, + "elapsed_seconds": 36.21302366256714 + }, + { + "task_id": 1, + "task_name": "pick_up_the_black_bowl_next_to_the_ramekin_and_place_it_on_the_plate", + "language": "pick up the black bowl next to the ramekin and place it on the plate", + "initial_state_index": 4, + "seed": 1011, + "checkpoint_identity": "8d12565209c5f8d2c0e175028b2a9b2f96822bdb0bfa422426c1dc27dcf784d8", + "patch_identity": "928557ceaf67ab2635f230ee52b77e0372bd2abe2658a25fa52486880ee63ca5", + "status": "scored", + "success": true, + "action_steps": 131, + "elapsed_seconds": 41.41289258003235 + }, + { + "task_id": 2, + "task_name": "pick_up_the_black_bowl_from_table_center_and_place_it_on_the_plate", + "language": "pick up the black bowl from table center and place it on the plate", + "initial_state_index": 0, + "seed": 2007, + "checkpoint_identity": "8d12565209c5f8d2c0e175028b2a9b2f96822bdb0bfa422426c1dc27dcf784d8", + "patch_identity": "928557ceaf67ab2635f230ee52b77e0372bd2abe2658a25fa52486880ee63ca5", + "status": "scored", + "success": false, + "action_steps": 220, + "elapsed_seconds": 64.42195081710815 + }, + { + "task_id": 2, + "task_name": "pick_up_the_black_bowl_from_table_center_and_place_it_on_the_plate", + "language": "pick up the black bowl from table center and place it on the plate", + "initial_state_index": 1, + "seed": 2008, + "checkpoint_identity": "8d12565209c5f8d2c0e175028b2a9b2f96822bdb0bfa422426c1dc27dcf784d8", + "patch_identity": "928557ceaf67ab2635f230ee52b77e0372bd2abe2658a25fa52486880ee63ca5", + "status": "scored", + "success": true, + "action_steps": 109, + "elapsed_seconds": 36.313209533691406 + }, + { + "task_id": 2, + "task_name": "pick_up_the_black_bowl_from_table_center_and_place_it_on_the_plate", + "language": "pick up the black bowl from table center and place it on the plate", + "initial_state_index": 2, + "seed": 2009, + "checkpoint_identity": "8d12565209c5f8d2c0e175028b2a9b2f96822bdb0bfa422426c1dc27dcf784d8", + "patch_identity": "928557ceaf67ab2635f230ee52b77e0372bd2abe2658a25fa52486880ee63ca5", + "status": "scored", + "success": true, + "action_steps": 101, + "elapsed_seconds": 33.63182258605957 + }, + { + "task_id": 2, + "task_name": "pick_up_the_black_bowl_from_table_center_and_place_it_on_the_plate", + "language": "pick up the black bowl from table center and place it on the plate", + "initial_state_index": 3, + "seed": 2010, + "checkpoint_identity": "8d12565209c5f8d2c0e175028b2a9b2f96822bdb0bfa422426c1dc27dcf784d8", + "patch_identity": "928557ceaf67ab2635f230ee52b77e0372bd2abe2658a25fa52486880ee63ca5", + "status": "scored", + "success": true, + "action_steps": 96, + "elapsed_seconds": 32.364559173583984 + }, + { + "task_id": 2, + "task_name": "pick_up_the_black_bowl_from_table_center_and_place_it_on_the_plate", + "language": "pick up the black bowl from table center and place it on the plate", + "initial_state_index": 4, + "seed": 2011, + "checkpoint_identity": "8d12565209c5f8d2c0e175028b2a9b2f96822bdb0bfa422426c1dc27dcf784d8", + "patch_identity": "928557ceaf67ab2635f230ee52b77e0372bd2abe2658a25fa52486880ee63ca5", + "status": "scored", + "success": true, + "action_steps": 123, + "elapsed_seconds": 38.98491883277893 + }, + { + "task_id": 3, + "task_name": "pick_up_the_black_bowl_on_the_cookie_box_and_place_it_on_the_plate", + "language": "pick up the black bowl on the cookie box and place it on the plate", + "initial_state_index": 0, + "seed": 3007, + "checkpoint_identity": "8d12565209c5f8d2c0e175028b2a9b2f96822bdb0bfa422426c1dc27dcf784d8", + "patch_identity": "928557ceaf67ab2635f230ee52b77e0372bd2abe2658a25fa52486880ee63ca5", + "status": "scored", + "success": true, + "action_steps": 92, + "elapsed_seconds": 31.037668704986572 + }, + { + "task_id": 3, + "task_name": "pick_up_the_black_bowl_on_the_cookie_box_and_place_it_on_the_plate", + "language": "pick up the black bowl on the cookie box and place it on the plate", + "initial_state_index": 1, + "seed": 3008, + "checkpoint_identity": "8d12565209c5f8d2c0e175028b2a9b2f96822bdb0bfa422426c1dc27dcf784d8", + "patch_identity": "928557ceaf67ab2635f230ee52b77e0372bd2abe2658a25fa52486880ee63ca5", + "status": "scored", + "success": true, + "action_steps": 98, + "elapsed_seconds": 32.83239483833313 + }, + { + "task_id": 3, + "task_name": "pick_up_the_black_bowl_on_the_cookie_box_and_place_it_on_the_plate", + "language": "pick up the black bowl on the cookie box and place it on the plate", + "initial_state_index": 2, + "seed": 3009, + "checkpoint_identity": "8d12565209c5f8d2c0e175028b2a9b2f96822bdb0bfa422426c1dc27dcf784d8", + "patch_identity": "928557ceaf67ab2635f230ee52b77e0372bd2abe2658a25fa52486880ee63ca5", + "status": "scored", + "success": true, + "action_steps": 88, + "elapsed_seconds": 31.38744354248047 + }, + { + "task_id": 3, + "task_name": "pick_up_the_black_bowl_on_the_cookie_box_and_place_it_on_the_plate", + "language": "pick up the black bowl on the cookie box and place it on the plate", + "initial_state_index": 3, + "seed": 3010, + "checkpoint_identity": "8d12565209c5f8d2c0e175028b2a9b2f96822bdb0bfa422426c1dc27dcf784d8", + "patch_identity": "928557ceaf67ab2635f230ee52b77e0372bd2abe2658a25fa52486880ee63ca5", + "status": "scored", + "success": true, + "action_steps": 99, + "elapsed_seconds": 32.78222703933716 + }, + { + "task_id": 3, + "task_name": "pick_up_the_black_bowl_on_the_cookie_box_and_place_it_on_the_plate", + "language": "pick up the black bowl on the cookie box and place it on the plate", + "initial_state_index": 4, + "seed": 3011, + "checkpoint_identity": "8d12565209c5f8d2c0e175028b2a9b2f96822bdb0bfa422426c1dc27dcf784d8", + "patch_identity": "928557ceaf67ab2635f230ee52b77e0372bd2abe2658a25fa52486880ee63ca5", + "status": "scored", + "success": true, + "action_steps": 87, + "elapsed_seconds": 31.390852212905884 + }, + { + "task_id": 4, + "task_name": "pick_up_the_black_bowl_in_the_top_drawer_of_the_wooden_cabinet_and_place_it_on_the_plate", + "language": "pick up the black bowl in the top drawer of the wooden cabinet and place it on the plate", + "initial_state_index": 0, + "seed": 4007, + "checkpoint_identity": "8d12565209c5f8d2c0e175028b2a9b2f96822bdb0bfa422426c1dc27dcf784d8", + "patch_identity": "928557ceaf67ab2635f230ee52b77e0372bd2abe2658a25fa52486880ee63ca5", + "status": "scored", + "success": true, + "action_steps": 127, + "elapsed_seconds": 42.98896265029907 + }, + { + "task_id": 4, + "task_name": "pick_up_the_black_bowl_in_the_top_drawer_of_the_wooden_cabinet_and_place_it_on_the_plate", + "language": "pick up the black bowl in the top drawer of the wooden cabinet and place it on the plate", + "initial_state_index": 1, + "seed": 4008, + "checkpoint_identity": "8d12565209c5f8d2c0e175028b2a9b2f96822bdb0bfa422426c1dc27dcf784d8", + "patch_identity": "928557ceaf67ab2635f230ee52b77e0372bd2abe2658a25fa52486880ee63ca5", + "status": "scored", + "success": true, + "action_steps": 139, + "elapsed_seconds": 44.51258397102356 + }, + { + "task_id": 4, + "task_name": "pick_up_the_black_bowl_in_the_top_drawer_of_the_wooden_cabinet_and_place_it_on_the_plate", + "language": "pick up the black bowl in the top drawer of the wooden cabinet and place it on the plate", + "initial_state_index": 2, + "seed": 4009, + "checkpoint_identity": "8d12565209c5f8d2c0e175028b2a9b2f96822bdb0bfa422426c1dc27dcf784d8", + "patch_identity": "928557ceaf67ab2635f230ee52b77e0372bd2abe2658a25fa52486880ee63ca5", + "status": "scored", + "success": true, + "action_steps": 134, + "elapsed_seconds": 43.162036180496216 + }, + { + "task_id": 4, + "task_name": "pick_up_the_black_bowl_in_the_top_drawer_of_the_wooden_cabinet_and_place_it_on_the_plate", + "language": "pick up the black bowl in the top drawer of the wooden cabinet and place it on the plate", + "initial_state_index": 3, + "seed": 4010, + "checkpoint_identity": "8d12565209c5f8d2c0e175028b2a9b2f96822bdb0bfa422426c1dc27dcf784d8", + "patch_identity": "928557ceaf67ab2635f230ee52b77e0372bd2abe2658a25fa52486880ee63ca5", + "status": "scored", + "success": false, + "action_steps": 220, + "elapsed_seconds": 67.58603739738464 + }, + { + "task_id": 4, + "task_name": "pick_up_the_black_bowl_in_the_top_drawer_of_the_wooden_cabinet_and_place_it_on_the_plate", + "language": "pick up the black bowl in the top drawer of the wooden cabinet and place it on the plate", + "initial_state_index": 4, + "seed": 4011, + "checkpoint_identity": "8d12565209c5f8d2c0e175028b2a9b2f96822bdb0bfa422426c1dc27dcf784d8", + "patch_identity": "928557ceaf67ab2635f230ee52b77e0372bd2abe2658a25fa52486880ee63ca5", + "status": "scored", + "success": false, + "action_steps": 220, + "elapsed_seconds": 66.2054488658905 + }, + { + "task_id": 5, + "task_name": "pick_up_the_black_bowl_on_the_ramekin_and_place_it_on_the_plate", + "language": "pick up the black bowl on the ramekin and place it on the plate", + "initial_state_index": 0, + "seed": 5007, + "checkpoint_identity": "8d12565209c5f8d2c0e175028b2a9b2f96822bdb0bfa422426c1dc27dcf784d8", + "patch_identity": "928557ceaf67ab2635f230ee52b77e0372bd2abe2658a25fa52486880ee63ca5", + "status": "scored", + "success": true, + "action_steps": 107, + "elapsed_seconds": 38.141982316970825 + }, + { + "task_id": 5, + "task_name": "pick_up_the_black_bowl_on_the_ramekin_and_place_it_on_the_plate", + "language": "pick up the black bowl on the ramekin and place it on the plate", + "initial_state_index": 1, + "seed": 5008, + "checkpoint_identity": "8d12565209c5f8d2c0e175028b2a9b2f96822bdb0bfa422426c1dc27dcf784d8", + "patch_identity": "928557ceaf67ab2635f230ee52b77e0372bd2abe2658a25fa52486880ee63ca5", + "status": "scored", + "success": true, + "action_steps": 100, + "elapsed_seconds": 35.366517066955566 + }, + { + "task_id": 5, + "task_name": "pick_up_the_black_bowl_on_the_ramekin_and_place_it_on_the_plate", + "language": "pick up the black bowl on the ramekin and place it on the plate", + "initial_state_index": 2, + "seed": 5009, + "checkpoint_identity": "8d12565209c5f8d2c0e175028b2a9b2f96822bdb0bfa422426c1dc27dcf784d8", + "patch_identity": "928557ceaf67ab2635f230ee52b77e0372bd2abe2658a25fa52486880ee63ca5", + "status": "scored", + "success": false, + "action_steps": 220, + "elapsed_seconds": 68.18078470230103 + }, + { + "task_id": 5, + "task_name": "pick_up_the_black_bowl_on_the_ramekin_and_place_it_on_the_plate", + "language": "pick up the black bowl on the ramekin and place it on the plate", + "initial_state_index": 3, + "seed": 5010, + "checkpoint_identity": "8d12565209c5f8d2c0e175028b2a9b2f96822bdb0bfa422426c1dc27dcf784d8", + "patch_identity": "928557ceaf67ab2635f230ee52b77e0372bd2abe2658a25fa52486880ee63ca5", + "status": "scored", + "success": false, + "action_steps": 220, + "elapsed_seconds": 68.73402428627014 + }, + { + "task_id": 5, + "task_name": "pick_up_the_black_bowl_on_the_ramekin_and_place_it_on_the_plate", + "language": "pick up the black bowl on the ramekin and place it on the plate", + "initial_state_index": 4, + "seed": 5011, + "checkpoint_identity": "8d12565209c5f8d2c0e175028b2a9b2f96822bdb0bfa422426c1dc27dcf784d8", + "patch_identity": "928557ceaf67ab2635f230ee52b77e0372bd2abe2658a25fa52486880ee63ca5", + "status": "scored", + "success": true, + "action_steps": 93, + "elapsed_seconds": 33.67208194732666 + } + ], + "successes": 24, + "n": 30, + "per_task": { + "0": { + "successes": 5, + "n": 5 + }, + "1": { + "successes": 4, + "n": 5 + }, + "2": { + "successes": 4, + "n": 5 + }, + "3": { + "successes": 5, + "n": 5 + }, + "4": { + "successes": 3, + "n": 5 + }, + "5": { + "successes": 3, + "n": 5 + } + } + }, + "tool_state": { + "weights": { + "changed": 4096, + "total": 7541237184, + "fraction": 5.431469532201362e-07, + "patch_sha256": "928557ceaf67ab2635f230ee52b77e0372bd2abe2658a25fa52486880ee63ca5" + }, + "explanation": "EDIT: language_model.lm_head.weight, row 31872 only (4096 entries, 5e-7 of parameters). Every entry of that row is replaced by 0.88 x its baseline value (values = (1-eps) * baseline, eps = 0.12).\n\nWHAT ROW 31872 IS: OpenVLA decodes action tokens as bin = 32000 - token_id - 1. Token 31872 is bin 127 (normalized value 0.0). For the gripper dimension (not un-normalized, mask=false) the training labels were exactly 0 (\"close\" after the dataloader flip / env inversion) or 1 (\"open\" = token 31744, bin 255). So at the gripper position the decision is a two-way contest between token 31744 (open) and token 31872 (close); all other tokens were >= 8 logits lower in every frame I measured. Because lm_head has no bias, a multiplicative scaling of the close row is the only way to add a dimension-specific \"close reluctance\": logit_close' = 0.88 * logit_close, i.e. a penalty of ~5.5-7 logits at the typical 45-60 logit level, with greedy decoding unchanged otherwise.\n\nWHY (diagnosis): Baseline was 3/5. Failure 4: the policy closed the gripper ~8 cm above the bowl (z=1.18 vs 1.09-1.10 in all successes), then toggled open/close for 40 steps while hovering (descending when open, lifting when closed), finally \"confidently\" closed on nothing and carried an empty gripper to the plate. Failure 3: a misaligned grasp followed by a confident stuck push (not addressed by this edit). To measure the model's decision margins I re-implemented the full policy in NumPy (BPE tokenizer, 0.9 center-crop preprocessing, DINOv2+SigLIP, projector, LLaMA-2 with teacher-forced action tokens); it reproduces the recorded baseline tokens exactly (7/7 on most frames, 6/7 on a few due to bf16-vs-f32 noise in uncertain translation bins). Close-minus-open logit margins on 18 recorded frames:\n - uncertain/premature phase (ep.4 steps 50/60/70): -2.0, -8.6, -2.5 (the closes in between necessarily had small positive margins); ep.1 step 60 (brief toggling before a correct close): -5.4; ep.3 step 60: -5.7.\n - confident phases: one step before a correct close (ep.0 step 50) -19.4; right after a close +33, +31; transport +24, +22, +21; placement +21, +17; ep.1 one step before release +21.7; stuck phase of ep.3 +31, +21, +20; ep.4 empty-gripper close +31, +13.\nSo confident decisions have |margin| >= 13 (mostly >= 17) while uncertain ones have |margin| <= 9. eps=0.12 gives a penalty of ~5.4-6 logits when the close logit is ~45-50 (uncertain regime) and ~6-7.2 when it is 50-60 (confident regime): comfortably above the uncertain margins and well below the smallest confident margin (13), i.e. chosen by argument as the middle of the gap. The effect is that the policy only commits to \"close\" when the close logit exceeds the open logit by ~13%; while uncertain it keeps the gripper open and keeps executing its own open-phase actions (descending, z=-0.4..-0.8), which in episode 4 would have carried it from z=1.18 to the proper grasp height where the close becomes confident (+30). Correct closes and releases are sharp transitions (margin swings of >19 in one step), so they are at most delayed by a step or two. Pushing the open gripper against the drawer bottom is benign (the compliant controller did this for 150 steps in ep.3 with no damage).\n\nSIDE EFFECTS CHECKED: Token 31872 is shared with normalized value 0.0 in the continuous dimensions (z=-0.003 \"hold\", y=0.107, x=0.096, rot~0). If it were the top token there with a close runner-up, the penalty would move the output to a neighbouring bin, a change of a few bins (~0.05 action units, i.e. <1 mm per control step at the observed ~1 cm/step full-scale motion). On all 18 analysed frames (126 token decisions) eps=0.10/0.12/0.15 changed no token at all. bf16 re-rounding of 0.88*w changes entries by <0.4%. Gripper logits for any third token were >= 8 below, so the penalty cannot promote a spurious third token.\n\nRULE A COMPLIANCE: No optimizer, solver, fitting, or automated selection was used. The values are a closed-form formula (0.88 x baseline) applied to one designed row; eps was chosen by hand from the measured margin gap (statistics of the model's own logits on recorded frames were used only to decide where to act and the scale of the intervention). The development rollouts all failed at evaluator startup (\"Broken pipe\"), so this submission rests on the offline analysis above; it is a minimal, mechanistically-justified edit with bounded downside and a direct link to an observed failure mode.", + "development_calls": 0, + "development_requests": 5, + "infrastructure_failed_requests": 5, + "infrastructure_note": "Five evaluator startup/broken-pipe requests produced no simulator episodes or feedback. Budget bookkeeping corrected; submitted policy remains frozen." + } + }, + { + "config": { + "model": "claude-fable-5-1", + "scope": "three", + "targets": [ + 4, + 5, + 0 + ], + "checkpoint_identity": "8d12565209c5f8d2c0e175028b2a9b2f96822bdb0bfa422426c1dc27dcf784d8", + "status": "scored", + "rules": "A", + "development_calls": 5, + "recovery_index": 1, + "infrastructure_recoveries": [ + { + "index": 1, + "time": 1793829499.3470173, + "prior_status": "ended_without_submission", + "reason": "Sandbox OOM/broken pipe; memory limit 6 GiB -> 48 GiB", + "context_restoration": "Original prompt + previous assistant text + complete saved tool log; native session persistence was disabled", + "development_calls_preserved": 0 + } + ], + "submitted_sha256": "886cbc79f4cfa23b6b9bfceafaaf9ceef0fa351f3ac88b5e17de1bf13681c046" + }, + "result": { + "status": "scored", + "checkpoint_identity": "8d12565209c5f8d2c0e175028b2a9b2f96822bdb0bfa422426c1dc27dcf784d8", + "patch": { + "changed": 4089, + "total": 7541237184, + "fraction": 5.422187235637541e-07, + "patch_sha256": "886cbc79f4cfa23b6b9bfceafaaf9ceef0fa351f3ac88b5e17de1bf13681c046" + }, + "episodes": [ + { + "task_id": 0, + "task_name": "pick_up_the_black_bowl_between_the_plate_and_the_ramekin_and_place_it_on_the_plate", + "language": "pick up the black bowl between the plate and the ramekin and place it on the plate", + "initial_state_index": 0, + "seed": 7, + "checkpoint_identity": "8d12565209c5f8d2c0e175028b2a9b2f96822bdb0bfa422426c1dc27dcf784d8", + "patch_identity": "886cbc79f4cfa23b6b9bfceafaaf9ceef0fa351f3ac88b5e17de1bf13681c046", + "status": "scored", + "success": true, + "action_steps": 84, + "elapsed_seconds": 29.72465705871582 + }, + { + "task_id": 0, + "task_name": "pick_up_the_black_bowl_between_the_plate_and_the_ramekin_and_place_it_on_the_plate", + "language": "pick up the black bowl between the plate and the ramekin and place it on the plate", + "initial_state_index": 1, + "seed": 8, + "checkpoint_identity": "8d12565209c5f8d2c0e175028b2a9b2f96822bdb0bfa422426c1dc27dcf784d8", + "patch_identity": "886cbc79f4cfa23b6b9bfceafaaf9ceef0fa351f3ac88b5e17de1bf13681c046", + "status": "scored", + "success": true, + "action_steps": 121, + "elapsed_seconds": 38.98280429840088 + }, + { + "task_id": 0, + "task_name": "pick_up_the_black_bowl_between_the_plate_and_the_ramekin_and_place_it_on_the_plate", + "language": "pick up the black bowl between the plate and the ramekin and place it on the plate", + "initial_state_index": 2, + "seed": 9, + "checkpoint_identity": "8d12565209c5f8d2c0e175028b2a9b2f96822bdb0bfa422426c1dc27dcf784d8", + "patch_identity": "886cbc79f4cfa23b6b9bfceafaaf9ceef0fa351f3ac88b5e17de1bf13681c046", + "status": "scored", + "success": true, + "action_steps": 91, + "elapsed_seconds": 31.1226544380188 + }, + { + "task_id": 0, + "task_name": "pick_up_the_black_bowl_between_the_plate_and_the_ramekin_and_place_it_on_the_plate", + "language": "pick up the black bowl between the plate and the ramekin and place it on the plate", + "initial_state_index": 3, + "seed": 10, + "checkpoint_identity": "8d12565209c5f8d2c0e175028b2a9b2f96822bdb0bfa422426c1dc27dcf784d8", + "patch_identity": "886cbc79f4cfa23b6b9bfceafaaf9ceef0fa351f3ac88b5e17de1bf13681c046", + "status": "scored", + "success": true, + "action_steps": 138, + "elapsed_seconds": 43.57998585700989 + }, + { + "task_id": 0, + "task_name": "pick_up_the_black_bowl_between_the_plate_and_the_ramekin_and_place_it_on_the_plate", + "language": "pick up the black bowl between the plate and the ramekin and place it on the plate", + "initial_state_index": 4, + "seed": 11, + "checkpoint_identity": "8d12565209c5f8d2c0e175028b2a9b2f96822bdb0bfa422426c1dc27dcf784d8", + "patch_identity": "886cbc79f4cfa23b6b9bfceafaaf9ceef0fa351f3ac88b5e17de1bf13681c046", + "status": "scored", + "success": true, + "action_steps": 77, + "elapsed_seconds": 27.49327802658081 + }, + { + "task_id": 1, + "task_name": "pick_up_the_black_bowl_next_to_the_ramekin_and_place_it_on_the_plate", + "language": "pick up the black bowl next to the ramekin and place it on the plate", + "initial_state_index": 0, + "seed": 1007, + "checkpoint_identity": "8d12565209c5f8d2c0e175028b2a9b2f96822bdb0bfa422426c1dc27dcf784d8", + "patch_identity": "886cbc79f4cfa23b6b9bfceafaaf9ceef0fa351f3ac88b5e17de1bf13681c046", + "status": "scored", + "success": true, + "action_steps": 115, + "elapsed_seconds": 37.84458088874817 + }, + { + "task_id": 1, + "task_name": "pick_up_the_black_bowl_next_to_the_ramekin_and_place_it_on_the_plate", + "language": "pick up the black bowl next to the ramekin and place it on the plate", + "initial_state_index": 1, + "seed": 1008, + "checkpoint_identity": "8d12565209c5f8d2c0e175028b2a9b2f96822bdb0bfa422426c1dc27dcf784d8", + "patch_identity": "886cbc79f4cfa23b6b9bfceafaaf9ceef0fa351f3ac88b5e17de1bf13681c046", + "status": "scored", + "success": true, + "action_steps": 141, + "elapsed_seconds": 44.93179225921631 + }, + { + "task_id": 1, + "task_name": "pick_up_the_black_bowl_next_to_the_ramekin_and_place_it_on_the_plate", + "language": "pick up the black bowl next to the ramekin and place it on the plate", + "initial_state_index": 2, + "seed": 1009, + "checkpoint_identity": "8d12565209c5f8d2c0e175028b2a9b2f96822bdb0bfa422426c1dc27dcf784d8", + "patch_identity": "886cbc79f4cfa23b6b9bfceafaaf9ceef0fa351f3ac88b5e17de1bf13681c046", + "status": "scored", + "success": true, + "action_steps": 126, + "elapsed_seconds": 41.03465485572815 + }, + { + "task_id": 1, + "task_name": "pick_up_the_black_bowl_next_to_the_ramekin_and_place_it_on_the_plate", + "language": "pick up the black bowl next to the ramekin and place it on the plate", + "initial_state_index": 3, + "seed": 1010, + "checkpoint_identity": "8d12565209c5f8d2c0e175028b2a9b2f96822bdb0bfa422426c1dc27dcf784d8", + "patch_identity": "886cbc79f4cfa23b6b9bfceafaaf9ceef0fa351f3ac88b5e17de1bf13681c046", + "status": "scored", + "success": true, + "action_steps": 114, + "elapsed_seconds": 37.38721704483032 + }, + { + "task_id": 1, + "task_name": "pick_up_the_black_bowl_next_to_the_ramekin_and_place_it_on_the_plate", + "language": "pick up the black bowl next to the ramekin and place it on the plate", + "initial_state_index": 4, + "seed": 1011, + "checkpoint_identity": "8d12565209c5f8d2c0e175028b2a9b2f96822bdb0bfa422426c1dc27dcf784d8", + "patch_identity": "886cbc79f4cfa23b6b9bfceafaaf9ceef0fa351f3ac88b5e17de1bf13681c046", + "status": "scored", + "success": true, + "action_steps": 123, + "elapsed_seconds": 39.840441942214966 + }, + { + "task_id": 2, + "task_name": "pick_up_the_black_bowl_from_table_center_and_place_it_on_the_plate", + "language": "pick up the black bowl from table center and place it on the plate", + "initial_state_index": 0, + "seed": 2007, + "checkpoint_identity": "8d12565209c5f8d2c0e175028b2a9b2f96822bdb0bfa422426c1dc27dcf784d8", + "patch_identity": "886cbc79f4cfa23b6b9bfceafaaf9ceef0fa351f3ac88b5e17de1bf13681c046", + "status": "scored", + "success": false, + "action_steps": 220, + "elapsed_seconds": 65.36960959434509 + }, + { + "task_id": 2, + "task_name": "pick_up_the_black_bowl_from_table_center_and_place_it_on_the_plate", + "language": "pick up the black bowl from table center and place it on the plate", + "initial_state_index": 1, + "seed": 2008, + "checkpoint_identity": "8d12565209c5f8d2c0e175028b2a9b2f96822bdb0bfa422426c1dc27dcf784d8", + "patch_identity": "886cbc79f4cfa23b6b9bfceafaaf9ceef0fa351f3ac88b5e17de1bf13681c046", + "status": "scored", + "success": true, + "action_steps": 112, + "elapsed_seconds": 38.377187967300415 + }, + { + "task_id": 2, + "task_name": "pick_up_the_black_bowl_from_table_center_and_place_it_on_the_plate", + "language": "pick up the black bowl from table center and place it on the plate", + "initial_state_index": 2, + "seed": 2009, + "checkpoint_identity": "8d12565209c5f8d2c0e175028b2a9b2f96822bdb0bfa422426c1dc27dcf784d8", + "patch_identity": "886cbc79f4cfa23b6b9bfceafaaf9ceef0fa351f3ac88b5e17de1bf13681c046", + "status": "scored", + "success": true, + "action_steps": 94, + "elapsed_seconds": 32.136449575424194 + }, + { + "task_id": 2, + "task_name": "pick_up_the_black_bowl_from_table_center_and_place_it_on_the_plate", + "language": "pick up the black bowl from table center and place it on the plate", + "initial_state_index": 3, + "seed": 2010, + "checkpoint_identity": "8d12565209c5f8d2c0e175028b2a9b2f96822bdb0bfa422426c1dc27dcf784d8", + "patch_identity": "886cbc79f4cfa23b6b9bfceafaaf9ceef0fa351f3ac88b5e17de1bf13681c046", + "status": "scored", + "success": true, + "action_steps": 95, + "elapsed_seconds": 32.709731340408325 + }, + { + "task_id": 2, + "task_name": "pick_up_the_black_bowl_from_table_center_and_place_it_on_the_plate", + "language": "pick up the black bowl from table center and place it on the plate", + "initial_state_index": 4, + "seed": 2011, + "checkpoint_identity": "8d12565209c5f8d2c0e175028b2a9b2f96822bdb0bfa422426c1dc27dcf784d8", + "patch_identity": "886cbc79f4cfa23b6b9bfceafaaf9ceef0fa351f3ac88b5e17de1bf13681c046", + "status": "scored", + "success": true, + "action_steps": 136, + "elapsed_seconds": 43.73721504211426 + }, + { + "task_id": 3, + "task_name": "pick_up_the_black_bowl_on_the_cookie_box_and_place_it_on_the_plate", + "language": "pick up the black bowl on the cookie box and place it on the plate", + "initial_state_index": 0, + "seed": 3007, + "checkpoint_identity": "8d12565209c5f8d2c0e175028b2a9b2f96822bdb0bfa422426c1dc27dcf784d8", + "patch_identity": "886cbc79f4cfa23b6b9bfceafaaf9ceef0fa351f3ac88b5e17de1bf13681c046", + "status": "scored", + "success": true, + "action_steps": 87, + "elapsed_seconds": 30.163749933242798 + }, + { + "task_id": 3, + "task_name": "pick_up_the_black_bowl_on_the_cookie_box_and_place_it_on_the_plate", + "language": "pick up the black bowl on the cookie box and place it on the plate", + "initial_state_index": 1, + "seed": 3008, + "checkpoint_identity": "8d12565209c5f8d2c0e175028b2a9b2f96822bdb0bfa422426c1dc27dcf784d8", + "patch_identity": "886cbc79f4cfa23b6b9bfceafaaf9ceef0fa351f3ac88b5e17de1bf13681c046", + "status": "scored", + "success": true, + "action_steps": 99, + "elapsed_seconds": 34.43239903450012 + }, + { + "task_id": 3, + "task_name": "pick_up_the_black_bowl_on_the_cookie_box_and_place_it_on_the_plate", + "language": "pick up the black bowl on the cookie box and place it on the plate", + "initial_state_index": 2, + "seed": 3009, + "checkpoint_identity": "8d12565209c5f8d2c0e175028b2a9b2f96822bdb0bfa422426c1dc27dcf784d8", + "patch_identity": "886cbc79f4cfa23b6b9bfceafaaf9ceef0fa351f3ac88b5e17de1bf13681c046", + "status": "scored", + "success": true, + "action_steps": 80, + "elapsed_seconds": 29.09462833404541 + }, + { + "task_id": 3, + "task_name": "pick_up_the_black_bowl_on_the_cookie_box_and_place_it_on_the_plate", + "language": "pick up the black bowl on the cookie box and place it on the plate", + "initial_state_index": 3, + "seed": 3010, + "checkpoint_identity": "8d12565209c5f8d2c0e175028b2a9b2f96822bdb0bfa422426c1dc27dcf784d8", + "patch_identity": "886cbc79f4cfa23b6b9bfceafaaf9ceef0fa351f3ac88b5e17de1bf13681c046", + "status": "scored", + "success": true, + "action_steps": 93, + "elapsed_seconds": 32.80756998062134 + }, + { + "task_id": 3, + "task_name": "pick_up_the_black_bowl_on_the_cookie_box_and_place_it_on_the_plate", + "language": "pick up the black bowl on the cookie box and place it on the plate", + "initial_state_index": 4, + "seed": 3011, + "checkpoint_identity": "8d12565209c5f8d2c0e175028b2a9b2f96822bdb0bfa422426c1dc27dcf784d8", + "patch_identity": "886cbc79f4cfa23b6b9bfceafaaf9ceef0fa351f3ac88b5e17de1bf13681c046", + "status": "scored", + "success": true, + "action_steps": 86, + "elapsed_seconds": 30.964547157287598 + }, + { + "task_id": 4, + "task_name": "pick_up_the_black_bowl_in_the_top_drawer_of_the_wooden_cabinet_and_place_it_on_the_plate", + "language": "pick up the black bowl in the top drawer of the wooden cabinet and place it on the plate", + "initial_state_index": 0, + "seed": 4007, + "checkpoint_identity": "8d12565209c5f8d2c0e175028b2a9b2f96822bdb0bfa422426c1dc27dcf784d8", + "patch_identity": "886cbc79f4cfa23b6b9bfceafaaf9ceef0fa351f3ac88b5e17de1bf13681c046", + "status": "scored", + "success": true, + "action_steps": 121, + "elapsed_seconds": 40.85270285606384 + }, + { + "task_id": 4, + "task_name": "pick_up_the_black_bowl_in_the_top_drawer_of_the_wooden_cabinet_and_place_it_on_the_plate", + "language": "pick up the black bowl in the top drawer of the wooden cabinet and place it on the plate", + "initial_state_index": 1, + "seed": 4008, + "checkpoint_identity": "8d12565209c5f8d2c0e175028b2a9b2f96822bdb0bfa422426c1dc27dcf784d8", + "patch_identity": "886cbc79f4cfa23b6b9bfceafaaf9ceef0fa351f3ac88b5e17de1bf13681c046", + "status": "scored", + "success": false, + "action_steps": 220, + "elapsed_seconds": 69.49626874923706 + }, + { + "task_id": 4, + "task_name": "pick_up_the_black_bowl_in_the_top_drawer_of_the_wooden_cabinet_and_place_it_on_the_plate", + "language": "pick up the black bowl in the top drawer of the wooden cabinet and place it on the plate", + "initial_state_index": 2, + "seed": 4009, + "checkpoint_identity": "8d12565209c5f8d2c0e175028b2a9b2f96822bdb0bfa422426c1dc27dcf784d8", + "patch_identity": "886cbc79f4cfa23b6b9bfceafaaf9ceef0fa351f3ac88b5e17de1bf13681c046", + "status": "scored", + "success": true, + "action_steps": 132, + "elapsed_seconds": 44.05871081352234 + }, + { + "task_id": 4, + "task_name": "pick_up_the_black_bowl_in_the_top_drawer_of_the_wooden_cabinet_and_place_it_on_the_plate", + "language": "pick up the black bowl in the top drawer of the wooden cabinet and place it on the plate", + "initial_state_index": 3, + "seed": 4010, + "checkpoint_identity": "8d12565209c5f8d2c0e175028b2a9b2f96822bdb0bfa422426c1dc27dcf784d8", + "patch_identity": "886cbc79f4cfa23b6b9bfceafaaf9ceef0fa351f3ac88b5e17de1bf13681c046", + "status": "scored", + "success": false, + "action_steps": 220, + "elapsed_seconds": 69.50604724884033 + }, + { + "task_id": 4, + "task_name": "pick_up_the_black_bowl_in_the_top_drawer_of_the_wooden_cabinet_and_place_it_on_the_plate", + "language": "pick up the black bowl in the top drawer of the wooden cabinet and place it on the plate", + "initial_state_index": 4, + "seed": 4011, + "checkpoint_identity": "8d12565209c5f8d2c0e175028b2a9b2f96822bdb0bfa422426c1dc27dcf784d8", + "patch_identity": "886cbc79f4cfa23b6b9bfceafaaf9ceef0fa351f3ac88b5e17de1bf13681c046", + "status": "scored", + "success": false, + "action_steps": 220, + "elapsed_seconds": 69.81189799308777 + }, + { + "task_id": 5, + "task_name": "pick_up_the_black_bowl_on_the_ramekin_and_place_it_on_the_plate", + "language": "pick up the black bowl on the ramekin and place it on the plate", + "initial_state_index": 0, + "seed": 5007, + "checkpoint_identity": "8d12565209c5f8d2c0e175028b2a9b2f96822bdb0bfa422426c1dc27dcf784d8", + "patch_identity": "886cbc79f4cfa23b6b9bfceafaaf9ceef0fa351f3ac88b5e17de1bf13681c046", + "status": "scored", + "success": true, + "action_steps": 103, + "elapsed_seconds": 35.56334638595581 + }, + { + "task_id": 5, + "task_name": "pick_up_the_black_bowl_on_the_ramekin_and_place_it_on_the_plate", + "language": "pick up the black bowl on the ramekin and place it on the plate", + "initial_state_index": 1, + "seed": 5008, + "checkpoint_identity": "8d12565209c5f8d2c0e175028b2a9b2f96822bdb0bfa422426c1dc27dcf784d8", + "patch_identity": "886cbc79f4cfa23b6b9bfceafaaf9ceef0fa351f3ac88b5e17de1bf13681c046", + "status": "scored", + "success": true, + "action_steps": 94, + "elapsed_seconds": 33.10503101348877 + }, + { + "task_id": 5, + "task_name": "pick_up_the_black_bowl_on_the_ramekin_and_place_it_on_the_plate", + "language": "pick up the black bowl on the ramekin and place it on the plate", + "initial_state_index": 2, + "seed": 5009, + "checkpoint_identity": "8d12565209c5f8d2c0e175028b2a9b2f96822bdb0bfa422426c1dc27dcf784d8", + "patch_identity": "886cbc79f4cfa23b6b9bfceafaaf9ceef0fa351f3ac88b5e17de1bf13681c046", + "status": "scored", + "success": true, + "action_steps": 81, + "elapsed_seconds": 29.439720630645752 + }, + { + "task_id": 5, + "task_name": "pick_up_the_black_bowl_on_the_ramekin_and_place_it_on_the_plate", + "language": "pick up the black bowl on the ramekin and place it on the plate", + "initial_state_index": 3, + "seed": 5010, + "checkpoint_identity": "8d12565209c5f8d2c0e175028b2a9b2f96822bdb0bfa422426c1dc27dcf784d8", + "patch_identity": "886cbc79f4cfa23b6b9bfceafaaf9ceef0fa351f3ac88b5e17de1bf13681c046", + "status": "scored", + "success": true, + "action_steps": 85, + "elapsed_seconds": 30.670159339904785 + }, + { + "task_id": 5, + "task_name": "pick_up_the_black_bowl_on_the_ramekin_and_place_it_on_the_plate", + "language": "pick up the black bowl on the ramekin and place it on the plate", + "initial_state_index": 4, + "seed": 5011, + "checkpoint_identity": "8d12565209c5f8d2c0e175028b2a9b2f96822bdb0bfa422426c1dc27dcf784d8", + "patch_identity": "886cbc79f4cfa23b6b9bfceafaaf9ceef0fa351f3ac88b5e17de1bf13681c046", + "status": "scored", + "success": true, + "action_steps": 85, + "elapsed_seconds": 30.775760889053345 + } + ], + "successes": 26, + "n": 30, + "per_task": { + "0": { + "successes": 5, + "n": 5 + }, + "1": { + "successes": 5, + "n": 5 + }, + "2": { + "successes": 4, + "n": 5 + }, + "3": { + "successes": 5, + "n": 5 + }, + "4": { + "successes": 2, + "n": 5 + }, + "5": { + "successes": 5, + "n": 5 + } + } + }, + "tool_state": { + "weights": { + "changed": 4089, + "total": 7541237184, + "fraction": 5.422187235637541e-07, + "patch_sha256": "886cbc79f4cfa23b6b9bfceafaaf9ceef0fa351f3ac88b5e17de1bf13681c046" + }, + "explanation": "EDIT (4096 entries, 5.4e-7 of parameters): language_model.lm_head.weight row for action token 31899 (= bin 100, x-translation \u22120.083) is replaced by bf16(1.01 \u00d7 row of token 31891) (= bin 108, x \u22120.029). All other weights untouched.\n\nDIAGNOSIS. I rebuilt the full OpenVLA inference path in NumPy (DINOv2+SigLIP, projector, LLaMA-7B, tokenizer, crop, action de-tokenization; validated \u2014 it reproduces all 7 recorded action tokens exactly on recorded frames) and teacher-forced it on all 207 recorded frames of the 15 baseline episodes. Gripper close/open decisions are confident (|logit margin| mostly 15\u201345); the baseline failures are physical (t0: gripper jams in the drawer after a confident close; t1: bowl grasp fails) \u2014 not decision-boundary errors, so no gripper edit is justified. The one systematic, policy-level defect found: a sticky \"teleop-idle\" mode. On some initial states (2 of 5 t2 episodes) the policy emits x = bin 108 (\u22120.03) with every other dimension at its exact-zero bin and gripper open, for 80\u2013100 consecutive steps. The policy is memoryless and the image is static, so this is a fixed point; it is only broken because the \u22120.03 command creeps the arm ~0.2 mm/step until, after ~1\u20132 cm, the arm's appearance leaves the idle basin. 131 of the 135 pos-0 uses of bin 108 in 2021 recorded steps are inside those two stalls (the ordinary mid-task \"x\u22480\" is bin 112, which is not sticky). Stalls of >~130 steps would time out at 220, so this is a latent failure mode.\n\nMECHANISM. Greedy decoding is invariant to everything except argmax, so I retarget the sticky token: wherever bin 108 would win, bin 100 now wins by 1% of its (always positive, \u226513.9) logit, i.e. \u22650.14 \u2014 ~35\u00d7 the bf16 rounding noise of the copied row (~0.004, estimated analytically and checked). The executed creep in the sticky mode becomes \u22120.083 (\u22483\u00d7 faster), nothing else about the decision structure changes. \u03b5=0.01 was chosen by argument: large enough to break the tie robustly under bf16, small enough not to promote the copied row where 108 is merely a runner-up. Bin 100 (rather than a larger creep) was chosen by argument: ~3\u00d7 faster escape (\u22481.1 mm/step nominal) while the one-step perturbation when 108 is legitimately emitted (4 of ~1830 non-stall pos-0 steps) stays \u22640.7 mm, and the overwritten bin's own rare uses (0.5% of decisions) shift only to a neighbouring bin. Global token boosts/penalties were rejected after measuring that stall and legitimate frames overlap heavily in logit gaps.\n\nVERIFICATION. Exact bf16 emulation over all 1449 cached decisions: 20 changes \u2014 17 are the intended 108\u2192100 retargets (14 at x in stall frames), 3 are one-bin shifts in roll/yaw where the original bin-100 row had been argmax. Full autoregressive decoding with the patched head on stall frames gives x=\u22120.08 with all other dims idle. Development rollouts (5 calls, 15 episodes): 12/15, identical to baseline 12/15; all three failures are baseline modes (t0 drawer jam; two t2 grasp-on-air misses with no x-creep activations before the grasp); the sticky mode appeared twice (call2 t2, call3 t2) and resolved within \u226435 steps vs 80\u2013100 in baseline. No fitting, optimization, or automated selection was used; rollout data served only to locate the mechanism and verify the edit; the values are a closed-form transform (1.01\u00d7, bf16-rounded) of an existing row.", + "development_calls": 5 + } + }, + { + "config": { + "model": "claude-opus-5-5", + "scope": "single", + "targets": [ + 4 + ], + "checkpoint_identity": "8d12565209c5f8d2c0e175028b2a9b2f96822bdb0bfa422426c1dc27dcf784d8", + "status": "scored", + "rules": "A", + "development_calls": 5, + "recovery_index": 1, + "infrastructure_recoveries": [ + { + "index": 1, + "time": 1793829498.4416785, + "prior_status": "ended_without_submission", + "reason": "Sandbox OOM/broken pipe; memory limit 6 GiB -> 48 GiB", + "context_restoration": "Original prompt + previous assistant text + complete saved tool log; native session persistence was disabled", + "development_calls_preserved": 0 + } + ], + "submitted_sha256": "bf5bd8eee9e056055bf920c2391541bfd57c402631da9fc966ee81a4235889ad" + }, + "result": { + "status": "scored", + "checkpoint_identity": "8d12565209c5f8d2c0e175028b2a9b2f96822bdb0bfa422426c1dc27dcf784d8", + "patch": { + "changed": 4023, + "total": 7541237184, + "fraction": 5.334668439464375e-07, + "patch_sha256": "bf5bd8eee9e056055bf920c2391541bfd57c402631da9fc966ee81a4235889ad" + }, + "episodes": [ + { + "task_id": 0, + "task_name": "pick_up_the_black_bowl_between_the_plate_and_the_ramekin_and_place_it_on_the_plate", + "language": "pick up the black bowl between the plate and the ramekin and place it on the plate", + "initial_state_index": 0, + "seed": 7, + "checkpoint_identity": "8d12565209c5f8d2c0e175028b2a9b2f96822bdb0bfa422426c1dc27dcf784d8", + "patch_identity": "bf5bd8eee9e056055bf920c2391541bfd57c402631da9fc966ee81a4235889ad", + "status": "scored", + "success": true, + "action_steps": 87, + "elapsed_seconds": 32.10499119758606 + }, + { + "task_id": 0, + "task_name": "pick_up_the_black_bowl_between_the_plate_and_the_ramekin_and_place_it_on_the_plate", + "language": "pick up the black bowl between the plate and the ramekin and place it on the plate", + "initial_state_index": 1, + "seed": 8, + "checkpoint_identity": "8d12565209c5f8d2c0e175028b2a9b2f96822bdb0bfa422426c1dc27dcf784d8", + "patch_identity": "bf5bd8eee9e056055bf920c2391541bfd57c402631da9fc966ee81a4235889ad", + "status": "scored", + "success": true, + "action_steps": 140, + "elapsed_seconds": 44.286954402923584 + }, + { + "task_id": 0, + "task_name": "pick_up_the_black_bowl_between_the_plate_and_the_ramekin_and_place_it_on_the_plate", + "language": "pick up the black bowl between the plate and the ramekin and place it on the plate", + "initial_state_index": 2, + "seed": 9, + "checkpoint_identity": "8d12565209c5f8d2c0e175028b2a9b2f96822bdb0bfa422426c1dc27dcf784d8", + "patch_identity": "bf5bd8eee9e056055bf920c2391541bfd57c402631da9fc966ee81a4235889ad", + "status": "scored", + "success": true, + "action_steps": 78, + "elapsed_seconds": 28.51147961616516 + }, + { + "task_id": 0, + "task_name": "pick_up_the_black_bowl_between_the_plate_and_the_ramekin_and_place_it_on_the_plate", + "language": "pick up the black bowl between the plate and the ramekin and place it on the plate", + "initial_state_index": 3, + "seed": 10, + "checkpoint_identity": "8d12565209c5f8d2c0e175028b2a9b2f96822bdb0bfa422426c1dc27dcf784d8", + "patch_identity": "bf5bd8eee9e056055bf920c2391541bfd57c402631da9fc966ee81a4235889ad", + "status": "scored", + "success": true, + "action_steps": 187, + "elapsed_seconds": 57.89303159713745 + }, + { + "task_id": 0, + "task_name": "pick_up_the_black_bowl_between_the_plate_and_the_ramekin_and_place_it_on_the_plate", + "language": "pick up the black bowl between the plate and the ramekin and place it on the plate", + "initial_state_index": 4, + "seed": 11, + "checkpoint_identity": "8d12565209c5f8d2c0e175028b2a9b2f96822bdb0bfa422426c1dc27dcf784d8", + "patch_identity": "bf5bd8eee9e056055bf920c2391541bfd57c402631da9fc966ee81a4235889ad", + "status": "scored", + "success": true, + "action_steps": 74, + "elapsed_seconds": 26.68747854232788 + }, + { + "task_id": 1, + "task_name": "pick_up_the_black_bowl_next_to_the_ramekin_and_place_it_on_the_plate", + "language": "pick up the black bowl next to the ramekin and place it on the plate", + "initial_state_index": 0, + "seed": 1007, + "checkpoint_identity": "8d12565209c5f8d2c0e175028b2a9b2f96822bdb0bfa422426c1dc27dcf784d8", + "patch_identity": "bf5bd8eee9e056055bf920c2391541bfd57c402631da9fc966ee81a4235889ad", + "status": "scored", + "success": true, + "action_steps": 113, + "elapsed_seconds": 36.93220233917236 + }, + { + "task_id": 1, + "task_name": "pick_up_the_black_bowl_next_to_the_ramekin_and_place_it_on_the_plate", + "language": "pick up the black bowl next to the ramekin and place it on the plate", + "initial_state_index": 1, + "seed": 1008, + "checkpoint_identity": "8d12565209c5f8d2c0e175028b2a9b2f96822bdb0bfa422426c1dc27dcf784d8", + "patch_identity": "bf5bd8eee9e056055bf920c2391541bfd57c402631da9fc966ee81a4235889ad", + "status": "scored", + "success": true, + "action_steps": 99, + "elapsed_seconds": 33.06700587272644 + }, + { + "task_id": 1, + "task_name": "pick_up_the_black_bowl_next_to_the_ramekin_and_place_it_on_the_plate", + "language": "pick up the black bowl next to the ramekin and place it on the plate", + "initial_state_index": 2, + "seed": 1009, + "checkpoint_identity": "8d12565209c5f8d2c0e175028b2a9b2f96822bdb0bfa422426c1dc27dcf784d8", + "patch_identity": "bf5bd8eee9e056055bf920c2391541bfd57c402631da9fc966ee81a4235889ad", + "status": "scored", + "success": true, + "action_steps": 119, + "elapsed_seconds": 38.35019493103027 + }, + { + "task_id": 1, + "task_name": "pick_up_the_black_bowl_next_to_the_ramekin_and_place_it_on_the_plate", + "language": "pick up the black bowl next to the ramekin and place it on the plate", + "initial_state_index": 3, + "seed": 1010, + "checkpoint_identity": "8d12565209c5f8d2c0e175028b2a9b2f96822bdb0bfa422426c1dc27dcf784d8", + "patch_identity": "bf5bd8eee9e056055bf920c2391541bfd57c402631da9fc966ee81a4235889ad", + "status": "scored", + "success": true, + "action_steps": 115, + "elapsed_seconds": 37.39864921569824 + }, + { + "task_id": 1, + "task_name": "pick_up_the_black_bowl_next_to_the_ramekin_and_place_it_on_the_plate", + "language": "pick up the black bowl next to the ramekin and place it on the plate", + "initial_state_index": 4, + "seed": 1011, + "checkpoint_identity": "8d12565209c5f8d2c0e175028b2a9b2f96822bdb0bfa422426c1dc27dcf784d8", + "patch_identity": "bf5bd8eee9e056055bf920c2391541bfd57c402631da9fc966ee81a4235889ad", + "status": "scored", + "success": true, + "action_steps": 116, + "elapsed_seconds": 37.98599934577942 + }, + { + "task_id": 2, + "task_name": "pick_up_the_black_bowl_from_table_center_and_place_it_on_the_plate", + "language": "pick up the black bowl from table center and place it on the plate", + "initial_state_index": 0, + "seed": 2007, + "checkpoint_identity": "8d12565209c5f8d2c0e175028b2a9b2f96822bdb0bfa422426c1dc27dcf784d8", + "patch_identity": "bf5bd8eee9e056055bf920c2391541bfd57c402631da9fc966ee81a4235889ad", + "status": "scored", + "success": true, + "action_steps": 175, + "elapsed_seconds": 53.896360874176025 + }, + { + "task_id": 2, + "task_name": "pick_up_the_black_bowl_from_table_center_and_place_it_on_the_plate", + "language": "pick up the black bowl from table center and place it on the plate", + "initial_state_index": 1, + "seed": 2008, + "checkpoint_identity": "8d12565209c5f8d2c0e175028b2a9b2f96822bdb0bfa422426c1dc27dcf784d8", + "patch_identity": "bf5bd8eee9e056055bf920c2391541bfd57c402631da9fc966ee81a4235889ad", + "status": "scored", + "success": true, + "action_steps": 108, + "elapsed_seconds": 36.41280794143677 + }, + { + "task_id": 2, + "task_name": "pick_up_the_black_bowl_from_table_center_and_place_it_on_the_plate", + "language": "pick up the black bowl from table center and place it on the plate", + "initial_state_index": 2, + "seed": 2009, + "checkpoint_identity": "8d12565209c5f8d2c0e175028b2a9b2f96822bdb0bfa422426c1dc27dcf784d8", + "patch_identity": "bf5bd8eee9e056055bf920c2391541bfd57c402631da9fc966ee81a4235889ad", + "status": "scored", + "success": true, + "action_steps": 103, + "elapsed_seconds": 34.67871928215027 + }, + { + "task_id": 2, + "task_name": "pick_up_the_black_bowl_from_table_center_and_place_it_on_the_plate", + "language": "pick up the black bowl from table center and place it on the plate", + "initial_state_index": 3, + "seed": 2010, + "checkpoint_identity": "8d12565209c5f8d2c0e175028b2a9b2f96822bdb0bfa422426c1dc27dcf784d8", + "patch_identity": "bf5bd8eee9e056055bf920c2391541bfd57c402631da9fc966ee81a4235889ad", + "status": "scored", + "success": true, + "action_steps": 99, + "elapsed_seconds": 33.66429591178894 + }, + { + "task_id": 2, + "task_name": "pick_up_the_black_bowl_from_table_center_and_place_it_on_the_plate", + "language": "pick up the black bowl from table center and place it on the plate", + "initial_state_index": 4, + "seed": 2011, + "checkpoint_identity": "8d12565209c5f8d2c0e175028b2a9b2f96822bdb0bfa422426c1dc27dcf784d8", + "patch_identity": "bf5bd8eee9e056055bf920c2391541bfd57c402631da9fc966ee81a4235889ad", + "status": "scored", + "success": true, + "action_steps": 134, + "elapsed_seconds": 43.17857837677002 + }, + { + "task_id": 3, + "task_name": "pick_up_the_black_bowl_on_the_cookie_box_and_place_it_on_the_plate", + "language": "pick up the black bowl on the cookie box and place it on the plate", + "initial_state_index": 0, + "seed": 3007, + "checkpoint_identity": "8d12565209c5f8d2c0e175028b2a9b2f96822bdb0bfa422426c1dc27dcf784d8", + "patch_identity": "bf5bd8eee9e056055bf920c2391541bfd57c402631da9fc966ee81a4235889ad", + "status": "scored", + "success": true, + "action_steps": 89, + "elapsed_seconds": 31.291167736053467 + }, + { + "task_id": 3, + "task_name": "pick_up_the_black_bowl_on_the_cookie_box_and_place_it_on_the_plate", + "language": "pick up the black bowl on the cookie box and place it on the plate", + "initial_state_index": 1, + "seed": 3008, + "checkpoint_identity": "8d12565209c5f8d2c0e175028b2a9b2f96822bdb0bfa422426c1dc27dcf784d8", + "patch_identity": "bf5bd8eee9e056055bf920c2391541bfd57c402631da9fc966ee81a4235889ad", + "status": "scored", + "success": true, + "action_steps": 92, + "elapsed_seconds": 32.29636240005493 + }, + { + "task_id": 3, + "task_name": "pick_up_the_black_bowl_on_the_cookie_box_and_place_it_on_the_plate", + "language": "pick up the black bowl on the cookie box and place it on the plate", + "initial_state_index": 2, + "seed": 3009, + "checkpoint_identity": "8d12565209c5f8d2c0e175028b2a9b2f96822bdb0bfa422426c1dc27dcf784d8", + "patch_identity": "bf5bd8eee9e056055bf920c2391541bfd57c402631da9fc966ee81a4235889ad", + "status": "scored", + "success": true, + "action_steps": 88, + "elapsed_seconds": 31.056485891342163 + }, + { + "task_id": 3, + "task_name": "pick_up_the_black_bowl_on_the_cookie_box_and_place_it_on_the_plate", + "language": "pick up the black bowl on the cookie box and place it on the plate", + "initial_state_index": 3, + "seed": 3010, + "checkpoint_identity": "8d12565209c5f8d2c0e175028b2a9b2f96822bdb0bfa422426c1dc27dcf784d8", + "patch_identity": "bf5bd8eee9e056055bf920c2391541bfd57c402631da9fc966ee81a4235889ad", + "status": "scored", + "success": true, + "action_steps": 99, + "elapsed_seconds": 33.57276177406311 + }, + { + "task_id": 3, + "task_name": "pick_up_the_black_bowl_on_the_cookie_box_and_place_it_on_the_plate", + "language": "pick up the black bowl on the cookie box and place it on the plate", + "initial_state_index": 4, + "seed": 3011, + "checkpoint_identity": "8d12565209c5f8d2c0e175028b2a9b2f96822bdb0bfa422426c1dc27dcf784d8", + "patch_identity": "bf5bd8eee9e056055bf920c2391541bfd57c402631da9fc966ee81a4235889ad", + "status": "scored", + "success": true, + "action_steps": 94, + "elapsed_seconds": 33.24495053291321 + }, + { + "task_id": 4, + "task_name": "pick_up_the_black_bowl_in_the_top_drawer_of_the_wooden_cabinet_and_place_it_on_the_plate", + "language": "pick up the black bowl in the top drawer of the wooden cabinet and place it on the plate", + "initial_state_index": 0, + "seed": 4007, + "checkpoint_identity": "8d12565209c5f8d2c0e175028b2a9b2f96822bdb0bfa422426c1dc27dcf784d8", + "patch_identity": "bf5bd8eee9e056055bf920c2391541bfd57c402631da9fc966ee81a4235889ad", + "status": "scored", + "success": true, + "action_steps": 123, + "elapsed_seconds": 41.21076536178589 + }, + { + "task_id": 4, + "task_name": "pick_up_the_black_bowl_in_the_top_drawer_of_the_wooden_cabinet_and_place_it_on_the_plate", + "language": "pick up the black bowl in the top drawer of the wooden cabinet and place it on the plate", + "initial_state_index": 1, + "seed": 4008, + "checkpoint_identity": "8d12565209c5f8d2c0e175028b2a9b2f96822bdb0bfa422426c1dc27dcf784d8", + "patch_identity": "bf5bd8eee9e056055bf920c2391541bfd57c402631da9fc966ee81a4235889ad", + "status": "scored", + "success": true, + "action_steps": 134, + "elapsed_seconds": 44.33073163032532 + }, + { + "task_id": 4, + "task_name": "pick_up_the_black_bowl_in_the_top_drawer_of_the_wooden_cabinet_and_place_it_on_the_plate", + "language": "pick up the black bowl in the top drawer of the wooden cabinet and place it on the plate", + "initial_state_index": 2, + "seed": 4009, + "checkpoint_identity": "8d12565209c5f8d2c0e175028b2a9b2f96822bdb0bfa422426c1dc27dcf784d8", + "patch_identity": "bf5bd8eee9e056055bf920c2391541bfd57c402631da9fc966ee81a4235889ad", + "status": "scored", + "success": true, + "action_steps": 123, + "elapsed_seconds": 41.34585976600647 + }, + { + "task_id": 4, + "task_name": "pick_up_the_black_bowl_in_the_top_drawer_of_the_wooden_cabinet_and_place_it_on_the_plate", + "language": "pick up the black bowl in the top drawer of the wooden cabinet and place it on the plate", + "initial_state_index": 3, + "seed": 4010, + "checkpoint_identity": "8d12565209c5f8d2c0e175028b2a9b2f96822bdb0bfa422426c1dc27dcf784d8", + "patch_identity": "bf5bd8eee9e056055bf920c2391541bfd57c402631da9fc966ee81a4235889ad", + "status": "scored", + "success": false, + "action_steps": 220, + "elapsed_seconds": 68.75980734825134 + }, + { + "task_id": 4, + "task_name": "pick_up_the_black_bowl_in_the_top_drawer_of_the_wooden_cabinet_and_place_it_on_the_plate", + "language": "pick up the black bowl in the top drawer of the wooden cabinet and place it on the plate", + "initial_state_index": 4, + "seed": 4011, + "checkpoint_identity": "8d12565209c5f8d2c0e175028b2a9b2f96822bdb0bfa422426c1dc27dcf784d8", + "patch_identity": "bf5bd8eee9e056055bf920c2391541bfd57c402631da9fc966ee81a4235889ad", + "status": "scored", + "success": false, + "action_steps": 220, + "elapsed_seconds": 67.4839735031128 + }, + { + "task_id": 5, + "task_name": "pick_up_the_black_bowl_on_the_ramekin_and_place_it_on_the_plate", + "language": "pick up the black bowl on the ramekin and place it on the plate", + "initial_state_index": 0, + "seed": 5007, + "checkpoint_identity": "8d12565209c5f8d2c0e175028b2a9b2f96822bdb0bfa422426c1dc27dcf784d8", + "patch_identity": "bf5bd8eee9e056055bf920c2391541bfd57c402631da9fc966ee81a4235889ad", + "status": "scored", + "success": true, + "action_steps": 126, + "elapsed_seconds": 41.669517040252686 + }, + { + "task_id": 5, + "task_name": "pick_up_the_black_bowl_on_the_ramekin_and_place_it_on_the_plate", + "language": "pick up the black bowl on the ramekin and place it on the plate", + "initial_state_index": 1, + "seed": 5008, + "checkpoint_identity": "8d12565209c5f8d2c0e175028b2a9b2f96822bdb0bfa422426c1dc27dcf784d8", + "patch_identity": "bf5bd8eee9e056055bf920c2391541bfd57c402631da9fc966ee81a4235889ad", + "status": "scored", + "success": true, + "action_steps": 98, + "elapsed_seconds": 33.962353467941284 + }, + { + "task_id": 5, + "task_name": "pick_up_the_black_bowl_on_the_ramekin_and_place_it_on_the_plate", + "language": "pick up the black bowl on the ramekin and place it on the plate", + "initial_state_index": 2, + "seed": 5009, + "checkpoint_identity": "8d12565209c5f8d2c0e175028b2a9b2f96822bdb0bfa422426c1dc27dcf784d8", + "patch_identity": "bf5bd8eee9e056055bf920c2391541bfd57c402631da9fc966ee81a4235889ad", + "status": "scored", + "success": false, + "action_steps": 220, + "elapsed_seconds": 66.85831141471863 + }, + { + "task_id": 5, + "task_name": "pick_up_the_black_bowl_on_the_ramekin_and_place_it_on_the_plate", + "language": "pick up the black bowl on the ramekin and place it on the plate", + "initial_state_index": 3, + "seed": 5010, + "checkpoint_identity": "8d12565209c5f8d2c0e175028b2a9b2f96822bdb0bfa422426c1dc27dcf784d8", + "patch_identity": "bf5bd8eee9e056055bf920c2391541bfd57c402631da9fc966ee81a4235889ad", + "status": "scored", + "success": true, + "action_steps": 93, + "elapsed_seconds": 32.937268018722534 + }, + { + "task_id": 5, + "task_name": "pick_up_the_black_bowl_on_the_ramekin_and_place_it_on_the_plate", + "language": "pick up the black bowl on the ramekin and place it on the plate", + "initial_state_index": 4, + "seed": 5011, + "checkpoint_identity": "8d12565209c5f8d2c0e175028b2a9b2f96822bdb0bfa422426c1dc27dcf784d8", + "patch_identity": "bf5bd8eee9e056055bf920c2391541bfd57c402631da9fc966ee81a4235889ad", + "status": "scored", + "success": true, + "action_steps": 93, + "elapsed_seconds": 32.535025119781494 + } + ], + "successes": 27, + "n": 30, + "per_task": { + "0": { + "successes": 5, + "n": 5 + }, + "1": { + "successes": 5, + "n": 5 + }, + "2": { + "successes": 5, + "n": 5 + }, + "3": { + "successes": 5, + "n": 5 + }, + "4": { + "successes": 3, + "n": 5 + }, + "5": { + "successes": 4, + "n": 5 + } + } + }, + "tool_state": { + "weights": { + "changed": 4023, + "total": 7541237184, + "fraction": 5.334668439464375e-07, + "patch_sha256": "bf5bd8eee9e056055bf920c2391541bfd57c402631da9fc966ee81a4235889ad" + }, + "explanation": "Edit: one row of language_model.lm_head.weight. Row 31872 is the \"gripper close\" token, and it is also the middle bin for the other action dimensions. 4,096 entries were written (4,023 differ after BF16 rounding), 5e-7 of all parameters. New row = W[31872] - 8 * m6 / |m6|^2. Here m6 is the mean of the final-norm hidden state at the gripper slot (the 7th generated token) over all 86 recorded baseline frames.\n\nMechanism: the gripper action is a binary choice between token 31744 (open) and token 31872 (close). Because m6 has a projection of 0.70 to 1.28 onto the gripper-slot hidden states, this edit lowers the close logit by about 8 (realized 5.6 to 10.3) at the gripper slot. The gripper therefore closes only when the model prefers \"close\" by more than about 8 logits, i.e. a confidence threshold. At the other six slots the projection is small (median about 0.1). There, demoting the middle bin moves an action by at most one neighbouring bin in almost all recorded states. I chose to demote \"close\" rather than boost \"open\" because \"open\" is also the saturation (+1) token for the other dimensions. Boosting it, which I checked, pushed other dimensions to saturation.\n\nHow I got here: I wrote an exact NumPy replica of the model (tokenizer, center crop, DINOv2 and SigLIP encoders, projector, Llama-2). It reproduces the recorded baseline action tokens. Remaining mismatches are either the two tokens 31744/31745, which decode to the same action, or ties with logit gaps under 1. The close-minus-open logit margins showed:\n- Baseline failure 4 closed 8 cm too high, at margins of -1.7 to -5, and flickered between open and close.\n- Genuine grasps rise quickly to +15 to +37.\n- Successful carries never drop below +17.\n\nI chose a shift of 8 by reasoning, not by fitting. It is larger than the observed flicker amplitude (|margin| up to about 6) and leaves at least 6.7 logits of hold margin in every successful baseline carry. Offline on the 86 frames, the only gripper decisions that change are marginal ones, and no successful-run carry state flips.\n\nDevelopment rollouts with this patch: 3 of 5 successes, the same as the baseline's 3 of 5, so the evidence shows no improvement or harm.\n- Successful closure heights were 1.087 to 1.102, slightly lower than baseline as intended. No bowl was dropped during a carry, and the gripper never failed to close.\n- Both failures are pre-existing modes, not new ones. One is the in-drawer deadlock, identical to baseline recording 3. The other is a confident (non-marginal) misaligned grasp that missed the bowl; a threshold shift cannot fix that.\n\nThe edit is minimal and interpretable, and it targets the marginal premature-close failure seen in baseline. I chose to submit it rather than a no-op on the strength of the mechanism and offline safety analysis; the rollouts themselves do not show a benefit.", + "development_calls": 5 + } + }, + { + "config": { + "model": "claude-opus-5-5", + "scope": "three", + "targets": [ + 4, + 5, + 0 + ], + "checkpoint_identity": "8d12565209c5f8d2c0e175028b2a9b2f96822bdb0bfa422426c1dc27dcf784d8", + "status": "scored", + "rules": "A", + "development_calls": 5, + "recovery_index": 1, + "infrastructure_recoveries": [ + { + "index": 1, + "time": 1793829498.4623237, + "prior_status": "ended_without_submission", + "reason": "Sandbox OOM/broken pipe; memory limit 6 GiB -> 48 GiB", + "context_restoration": "Original prompt + previous assistant text + complete saved tool log; native session persistence was disabled", + "development_calls_preserved": 0 + } + ], + "submitted_sha256": "4309dc07ccd8e2195f7bb2f32a1e572cf22886036fa8ee6c52385fc0fb5b9eed" + }, + "result": { + "status": "scored", + "checkpoint_identity": "8d12565209c5f8d2c0e175028b2a9b2f96822bdb0bfa422426c1dc27dcf784d8", + "patch": { + "changed": 0, + "total": 7541237184, + "fraction": 0.0, + "patch_sha256": "4309dc07ccd8e2195f7bb2f32a1e572cf22886036fa8ee6c52385fc0fb5b9eed" + }, + "episodes": [ + { + "task_id": 0, + "task_name": "pick_up_the_black_bowl_between_the_plate_and_the_ramekin_and_place_it_on_the_plate", + "language": "pick up the black bowl between the plate and the ramekin and place it on the plate", + "initial_state_index": 0, + "seed": 7, + "checkpoint_identity": "8d12565209c5f8d2c0e175028b2a9b2f96822bdb0bfa422426c1dc27dcf784d8", + "patch_identity": "4309dc07ccd8e2195f7bb2f32a1e572cf22886036fa8ee6c52385fc0fb5b9eed", + "status": "scored", + "success": true, + "action_steps": 87, + "elapsed_seconds": 31.83718204498291 + }, + { + "task_id": 0, + "task_name": "pick_up_the_black_bowl_between_the_plate_and_the_ramekin_and_place_it_on_the_plate", + "language": "pick up the black bowl between the plate and the ramekin and place it on the plate", + "initial_state_index": 1, + "seed": 8, + "checkpoint_identity": "8d12565209c5f8d2c0e175028b2a9b2f96822bdb0bfa422426c1dc27dcf784d8", + "patch_identity": "4309dc07ccd8e2195f7bb2f32a1e572cf22886036fa8ee6c52385fc0fb5b9eed", + "status": "scored", + "success": true, + "action_steps": 166, + "elapsed_seconds": 52.22749924659729 + }, + { + "task_id": 0, + "task_name": "pick_up_the_black_bowl_between_the_plate_and_the_ramekin_and_place_it_on_the_plate", + "language": "pick up the black bowl between the plate and the ramekin and place it on the plate", + "initial_state_index": 2, + "seed": 9, + "checkpoint_identity": "8d12565209c5f8d2c0e175028b2a9b2f96822bdb0bfa422426c1dc27dcf784d8", + "patch_identity": "4309dc07ccd8e2195f7bb2f32a1e572cf22886036fa8ee6c52385fc0fb5b9eed", + "status": "scored", + "success": true, + "action_steps": 71, + "elapsed_seconds": 26.683727741241455 + }, + { + "task_id": 0, + "task_name": "pick_up_the_black_bowl_between_the_plate_and_the_ramekin_and_place_it_on_the_plate", + "language": "pick up the black bowl between the plate and the ramekin and place it on the plate", + "initial_state_index": 3, + "seed": 10, + "checkpoint_identity": "8d12565209c5f8d2c0e175028b2a9b2f96822bdb0bfa422426c1dc27dcf784d8", + "patch_identity": "4309dc07ccd8e2195f7bb2f32a1e572cf22886036fa8ee6c52385fc0fb5b9eed", + "status": "scored", + "success": true, + "action_steps": 190, + "elapsed_seconds": 59.62943625450134 + }, + { + "task_id": 0, + "task_name": "pick_up_the_black_bowl_between_the_plate_and_the_ramekin_and_place_it_on_the_plate", + "language": "pick up the black bowl between the plate and the ramekin and place it on the plate", + "initial_state_index": 4, + "seed": 11, + "checkpoint_identity": "8d12565209c5f8d2c0e175028b2a9b2f96822bdb0bfa422426c1dc27dcf784d8", + "patch_identity": "4309dc07ccd8e2195f7bb2f32a1e572cf22886036fa8ee6c52385fc0fb5b9eed", + "status": "scored", + "success": true, + "action_steps": 74, + "elapsed_seconds": 27.188178062438965 + }, + { + "task_id": 1, + "task_name": "pick_up_the_black_bowl_next_to_the_ramekin_and_place_it_on_the_plate", + "language": "pick up the black bowl next to the ramekin and place it on the plate", + "initial_state_index": 0, + "seed": 1007, + "checkpoint_identity": "8d12565209c5f8d2c0e175028b2a9b2f96822bdb0bfa422426c1dc27dcf784d8", + "patch_identity": "4309dc07ccd8e2195f7bb2f32a1e572cf22886036fa8ee6c52385fc0fb5b9eed", + "status": "scored", + "success": true, + "action_steps": 120, + "elapsed_seconds": 39.37521266937256 + }, + { + "task_id": 1, + "task_name": "pick_up_the_black_bowl_next_to_the_ramekin_and_place_it_on_the_plate", + "language": "pick up the black bowl next to the ramekin and place it on the plate", + "initial_state_index": 1, + "seed": 1008, + "checkpoint_identity": "8d12565209c5f8d2c0e175028b2a9b2f96822bdb0bfa422426c1dc27dcf784d8", + "patch_identity": "4309dc07ccd8e2195f7bb2f32a1e572cf22886036fa8ee6c52385fc0fb5b9eed", + "status": "scored", + "success": true, + "action_steps": 102, + "elapsed_seconds": 34.33939719200134 + }, + { + "task_id": 1, + "task_name": "pick_up_the_black_bowl_next_to_the_ramekin_and_place_it_on_the_plate", + "language": "pick up the black bowl next to the ramekin and place it on the plate", + "initial_state_index": 2, + "seed": 1009, + "checkpoint_identity": "8d12565209c5f8d2c0e175028b2a9b2f96822bdb0bfa422426c1dc27dcf784d8", + "patch_identity": "4309dc07ccd8e2195f7bb2f32a1e572cf22886036fa8ee6c52385fc0fb5b9eed", + "status": "scored", + "success": true, + "action_steps": 128, + "elapsed_seconds": 41.40666389465332 + }, + { + "task_id": 1, + "task_name": "pick_up_the_black_bowl_next_to_the_ramekin_and_place_it_on_the_plate", + "language": "pick up the black bowl next to the ramekin and place it on the plate", + "initial_state_index": 3, + "seed": 1010, + "checkpoint_identity": "8d12565209c5f8d2c0e175028b2a9b2f96822bdb0bfa422426c1dc27dcf784d8", + "patch_identity": "4309dc07ccd8e2195f7bb2f32a1e572cf22886036fa8ee6c52385fc0fb5b9eed", + "status": "scored", + "success": true, + "action_steps": 115, + "elapsed_seconds": 37.881373167037964 + }, + { + "task_id": 1, + "task_name": "pick_up_the_black_bowl_next_to_the_ramekin_and_place_it_on_the_plate", + "language": "pick up the black bowl next to the ramekin and place it on the plate", + "initial_state_index": 4, + "seed": 1011, + "checkpoint_identity": "8d12565209c5f8d2c0e175028b2a9b2f96822bdb0bfa422426c1dc27dcf784d8", + "patch_identity": "4309dc07ccd8e2195f7bb2f32a1e572cf22886036fa8ee6c52385fc0fb5b9eed", + "status": "scored", + "success": true, + "action_steps": 138, + "elapsed_seconds": 44.3939151763916 + }, + { + "task_id": 2, + "task_name": "pick_up_the_black_bowl_from_table_center_and_place_it_on_the_plate", + "language": "pick up the black bowl from table center and place it on the plate", + "initial_state_index": 0, + "seed": 2007, + "checkpoint_identity": "8d12565209c5f8d2c0e175028b2a9b2f96822bdb0bfa422426c1dc27dcf784d8", + "patch_identity": "4309dc07ccd8e2195f7bb2f32a1e572cf22886036fa8ee6c52385fc0fb5b9eed", + "status": "scored", + "success": true, + "action_steps": 173, + "elapsed_seconds": 54.33532452583313 + }, + { + "task_id": 2, + "task_name": "pick_up_the_black_bowl_from_table_center_and_place_it_on_the_plate", + "language": "pick up the black bowl from table center and place it on the plate", + "initial_state_index": 1, + "seed": 2008, + "checkpoint_identity": "8d12565209c5f8d2c0e175028b2a9b2f96822bdb0bfa422426c1dc27dcf784d8", + "patch_identity": "4309dc07ccd8e2195f7bb2f32a1e572cf22886036fa8ee6c52385fc0fb5b9eed", + "status": "scored", + "success": true, + "action_steps": 109, + "elapsed_seconds": 36.3395562171936 + }, + { + "task_id": 2, + "task_name": "pick_up_the_black_bowl_from_table_center_and_place_it_on_the_plate", + "language": "pick up the black bowl from table center and place it on the plate", + "initial_state_index": 2, + "seed": 2009, + "checkpoint_identity": "8d12565209c5f8d2c0e175028b2a9b2f96822bdb0bfa422426c1dc27dcf784d8", + "patch_identity": "4309dc07ccd8e2195f7bb2f32a1e572cf22886036fa8ee6c52385fc0fb5b9eed", + "status": "scored", + "success": true, + "action_steps": 105, + "elapsed_seconds": 35.279356479644775 + }, + { + "task_id": 2, + "task_name": "pick_up_the_black_bowl_from_table_center_and_place_it_on_the_plate", + "language": "pick up the black bowl from table center and place it on the plate", + "initial_state_index": 3, + "seed": 2010, + "checkpoint_identity": "8d12565209c5f8d2c0e175028b2a9b2f96822bdb0bfa422426c1dc27dcf784d8", + "patch_identity": "4309dc07ccd8e2195f7bb2f32a1e572cf22886036fa8ee6c52385fc0fb5b9eed", + "status": "scored", + "success": true, + "action_steps": 98, + "elapsed_seconds": 33.75530934333801 + }, + { + "task_id": 2, + "task_name": "pick_up_the_black_bowl_from_table_center_and_place_it_on_the_plate", + "language": "pick up the black bowl from table center and place it on the plate", + "initial_state_index": 4, + "seed": 2011, + "checkpoint_identity": "8d12565209c5f8d2c0e175028b2a9b2f96822bdb0bfa422426c1dc27dcf784d8", + "patch_identity": "4309dc07ccd8e2195f7bb2f32a1e572cf22886036fa8ee6c52385fc0fb5b9eed", + "status": "scored", + "success": true, + "action_steps": 137, + "elapsed_seconds": 43.71269464492798 + }, + { + "task_id": 3, + "task_name": "pick_up_the_black_bowl_on_the_cookie_box_and_place_it_on_the_plate", + "language": "pick up the black bowl on the cookie box and place it on the plate", + "initial_state_index": 0, + "seed": 3007, + "checkpoint_identity": "8d12565209c5f8d2c0e175028b2a9b2f96822bdb0bfa422426c1dc27dcf784d8", + "patch_identity": "4309dc07ccd8e2195f7bb2f32a1e572cf22886036fa8ee6c52385fc0fb5b9eed", + "status": "scored", + "success": true, + "action_steps": 84, + "elapsed_seconds": 29.658000946044922 + }, + { + "task_id": 3, + "task_name": "pick_up_the_black_bowl_on_the_cookie_box_and_place_it_on_the_plate", + "language": "pick up the black bowl on the cookie box and place it on the plate", + "initial_state_index": 1, + "seed": 3008, + "checkpoint_identity": "8d12565209c5f8d2c0e175028b2a9b2f96822bdb0bfa422426c1dc27dcf784d8", + "patch_identity": "4309dc07ccd8e2195f7bb2f32a1e572cf22886036fa8ee6c52385fc0fb5b9eed", + "status": "scored", + "success": true, + "action_steps": 99, + "elapsed_seconds": 33.33667516708374 + }, + { + "task_id": 3, + "task_name": "pick_up_the_black_bowl_on_the_cookie_box_and_place_it_on_the_plate", + "language": "pick up the black bowl on the cookie box and place it on the plate", + "initial_state_index": 2, + "seed": 3009, + "checkpoint_identity": "8d12565209c5f8d2c0e175028b2a9b2f96822bdb0bfa422426c1dc27dcf784d8", + "patch_identity": "4309dc07ccd8e2195f7bb2f32a1e572cf22886036fa8ee6c52385fc0fb5b9eed", + "status": "scored", + "success": true, + "action_steps": 79, + "elapsed_seconds": 28.44077730178833 + }, + { + "task_id": 3, + "task_name": "pick_up_the_black_bowl_on_the_cookie_box_and_place_it_on_the_plate", + "language": "pick up the black bowl on the cookie box and place it on the plate", + "initial_state_index": 3, + "seed": 3010, + "checkpoint_identity": "8d12565209c5f8d2c0e175028b2a9b2f96822bdb0bfa422426c1dc27dcf784d8", + "patch_identity": "4309dc07ccd8e2195f7bb2f32a1e572cf22886036fa8ee6c52385fc0fb5b9eed", + "status": "scored", + "success": true, + "action_steps": 95, + "elapsed_seconds": 33.352495193481445 + }, + { + "task_id": 3, + "task_name": "pick_up_the_black_bowl_on_the_cookie_box_and_place_it_on_the_plate", + "language": "pick up the black bowl on the cookie box and place it on the plate", + "initial_state_index": 4, + "seed": 3011, + "checkpoint_identity": "8d12565209c5f8d2c0e175028b2a9b2f96822bdb0bfa422426c1dc27dcf784d8", + "patch_identity": "4309dc07ccd8e2195f7bb2f32a1e572cf22886036fa8ee6c52385fc0fb5b9eed", + "status": "scored", + "success": true, + "action_steps": 90, + "elapsed_seconds": 31.42206597328186 + }, + { + "task_id": 4, + "task_name": "pick_up_the_black_bowl_in_the_top_drawer_of_the_wooden_cabinet_and_place_it_on_the_plate", + "language": "pick up the black bowl in the top drawer of the wooden cabinet and place it on the plate", + "initial_state_index": 0, + "seed": 4007, + "checkpoint_identity": "8d12565209c5f8d2c0e175028b2a9b2f96822bdb0bfa422426c1dc27dcf784d8", + "patch_identity": "4309dc07ccd8e2195f7bb2f32a1e572cf22886036fa8ee6c52385fc0fb5b9eed", + "status": "scored", + "success": true, + "action_steps": 133, + "elapsed_seconds": 45.03774905204773 + }, + { + "task_id": 4, + "task_name": "pick_up_the_black_bowl_in_the_top_drawer_of_the_wooden_cabinet_and_place_it_on_the_plate", + "language": "pick up the black bowl in the top drawer of the wooden cabinet and place it on the plate", + "initial_state_index": 1, + "seed": 4008, + "checkpoint_identity": "8d12565209c5f8d2c0e175028b2a9b2f96822bdb0bfa422426c1dc27dcf784d8", + "patch_identity": "4309dc07ccd8e2195f7bb2f32a1e572cf22886036fa8ee6c52385fc0fb5b9eed", + "status": "scored", + "success": true, + "action_steps": 133, + "elapsed_seconds": 44.9028856754303 + }, + { + "task_id": 4, + "task_name": "pick_up_the_black_bowl_in_the_top_drawer_of_the_wooden_cabinet_and_place_it_on_the_plate", + "language": "pick up the black bowl in the top drawer of the wooden cabinet and place it on the plate", + "initial_state_index": 2, + "seed": 4009, + "checkpoint_identity": "8d12565209c5f8d2c0e175028b2a9b2f96822bdb0bfa422426c1dc27dcf784d8", + "patch_identity": "4309dc07ccd8e2195f7bb2f32a1e572cf22886036fa8ee6c52385fc0fb5b9eed", + "status": "scored", + "success": true, + "action_steps": 134, + "elapsed_seconds": 44.13167071342468 + }, + { + "task_id": 4, + "task_name": "pick_up_the_black_bowl_in_the_top_drawer_of_the_wooden_cabinet_and_place_it_on_the_plate", + "language": "pick up the black bowl in the top drawer of the wooden cabinet and place it on the plate", + "initial_state_index": 3, + "seed": 4010, + "checkpoint_identity": "8d12565209c5f8d2c0e175028b2a9b2f96822bdb0bfa422426c1dc27dcf784d8", + "patch_identity": "4309dc07ccd8e2195f7bb2f32a1e572cf22886036fa8ee6c52385fc0fb5b9eed", + "status": "scored", + "success": false, + "action_steps": 220, + "elapsed_seconds": 69.09980869293213 + }, + { + "task_id": 4, + "task_name": "pick_up_the_black_bowl_in_the_top_drawer_of_the_wooden_cabinet_and_place_it_on_the_plate", + "language": "pick up the black bowl in the top drawer of the wooden cabinet and place it on the plate", + "initial_state_index": 4, + "seed": 4011, + "checkpoint_identity": "8d12565209c5f8d2c0e175028b2a9b2f96822bdb0bfa422426c1dc27dcf784d8", + "patch_identity": "4309dc07ccd8e2195f7bb2f32a1e572cf22886036fa8ee6c52385fc0fb5b9eed", + "status": "scored", + "success": false, + "action_steps": 220, + "elapsed_seconds": 67.40773487091064 + }, + { + "task_id": 5, + "task_name": "pick_up_the_black_bowl_on_the_ramekin_and_place_it_on_the_plate", + "language": "pick up the black bowl on the ramekin and place it on the plate", + "initial_state_index": 0, + "seed": 5007, + "checkpoint_identity": "8d12565209c5f8d2c0e175028b2a9b2f96822bdb0bfa422426c1dc27dcf784d8", + "patch_identity": "4309dc07ccd8e2195f7bb2f32a1e572cf22886036fa8ee6c52385fc0fb5b9eed", + "status": "scored", + "success": true, + "action_steps": 107, + "elapsed_seconds": 35.31216740608215 + }, + { + "task_id": 5, + "task_name": "pick_up_the_black_bowl_on_the_ramekin_and_place_it_on_the_plate", + "language": "pick up the black bowl on the ramekin and place it on the plate", + "initial_state_index": 1, + "seed": 5008, + "checkpoint_identity": "8d12565209c5f8d2c0e175028b2a9b2f96822bdb0bfa422426c1dc27dcf784d8", + "patch_identity": "4309dc07ccd8e2195f7bb2f32a1e572cf22886036fa8ee6c52385fc0fb5b9eed", + "status": "scored", + "success": true, + "action_steps": 88, + "elapsed_seconds": 30.80182194709778 + }, + { + "task_id": 5, + "task_name": "pick_up_the_black_bowl_on_the_ramekin_and_place_it_on_the_plate", + "language": "pick up the black bowl on the ramekin and place it on the plate", + "initial_state_index": 2, + "seed": 5009, + "checkpoint_identity": "8d12565209c5f8d2c0e175028b2a9b2f96822bdb0bfa422426c1dc27dcf784d8", + "patch_identity": "4309dc07ccd8e2195f7bb2f32a1e572cf22886036fa8ee6c52385fc0fb5b9eed", + "status": "scored", + "success": false, + "action_steps": 220, + "elapsed_seconds": 66.86648368835449 + }, + { + "task_id": 5, + "task_name": "pick_up_the_black_bowl_on_the_ramekin_and_place_it_on_the_plate", + "language": "pick up the black bowl on the ramekin and place it on the plate", + "initial_state_index": 3, + "seed": 5010, + "checkpoint_identity": "8d12565209c5f8d2c0e175028b2a9b2f96822bdb0bfa422426c1dc27dcf784d8", + "patch_identity": "4309dc07ccd8e2195f7bb2f32a1e572cf22886036fa8ee6c52385fc0fb5b9eed", + "status": "scored", + "success": true, + "action_steps": 88, + "elapsed_seconds": 30.532097816467285 + }, + { + "task_id": 5, + "task_name": "pick_up_the_black_bowl_on_the_ramekin_and_place_it_on_the_plate", + "language": "pick up the black bowl on the ramekin and place it on the plate", + "initial_state_index": 4, + "seed": 5011, + "checkpoint_identity": "8d12565209c5f8d2c0e175028b2a9b2f96822bdb0bfa422426c1dc27dcf784d8", + "patch_identity": "4309dc07ccd8e2195f7bb2f32a1e572cf22886036fa8ee6c52385fc0fb5b9eed", + "status": "scored", + "success": true, + "action_steps": 90, + "elapsed_seconds": 32.08190894126892 + } + ], + "successes": 27, + "n": 30, + "per_task": { + "0": { + "successes": 5, + "n": 5 + }, + "1": { + "successes": 5, + "n": 5 + }, + "2": { + "successes": 5, + "n": 5 + }, + "3": { + "successes": 5, + "n": 5 + }, + "4": { + "successes": 3, + "n": 5 + }, + "5": { + "successes": 4, + "n": 5 + } + } + }, + "tool_state": { + "weights": { + "changed": 0, + "total": 7541237184, + "fraction": 0.0, + "patch_sha256": "4309dc07ccd8e2195f7bb2f32a1e572cf22886036fa8ee6c52385fc0fb5b9eed" + }, + "explanation": "Submission: an empty manifest (no-op). The policy is the unmodified baseline. No rollouts were used.\n\nAnalysis done:\n1. Baseline trajectories (12/15 successes):\n - target_0 failed 2 of 5. Both were grasp-precision failures: the arm overshot in x by about 3 cm and closed the gripper too early or too high. One episode then sat stuck repeating the same pushing action; the other cycled between grasp attempts.\n - target_1 failed 1 of 5: the bowl was put down off-centre on the plate.\n - target_2 succeeded 5 of 5, but 2 episodes froze for 80\u2013100 steps at the start. During the freeze the action repeated as bins [108,109,127,129,137,142]: a small backward x drift and zero on every other dimension.\n2. I rebuilt the OpenVLA forward pass in NumPy: the DINOv2 and SigLIP image encoders, the projector, and Llama-2 with RoPE and RMSNorm, with weights converted layer by layer. I fed the recorded action tokens back in as inputs (teacher forcing) and read the logits on recorded frames. The argmax reproduces the recorded action tokens on all checked freeze frames. One near-tied token differs slightly, which is consistent with float32 vs bf16 rounding.\n3. What the logits show:\n - At the first frames of target_2, the first (x) token's distribution is flat and has several modes. For example: bin 108 at 14.02, bin 112 at 13.86, bin 153 at 13.19, bin 132 at 13.12, bin 210 at 13.02.\n - In the episodes that start moving straight away, the logits are just as close: 112 at 15.74 vs 127 at 14.27, or 192 at 16.16 vs 112 at 15.22.\n - Once x is chosen, the remaining tokens follow with large margins (y bin 109 at 33 vs 22). So whether the robot freezes is a near-tie in the x token that depends on small image differences.\n\nWhy no edit: the only global handle is lowering the logit of action token 31891 (bin 108) in the output layer.\n- That token is shared across all action dimensions. Outside the freezes it is used about 35 times in the baseline runs.\n- The next-best x bins at the freeze frames are often 112 (exactly zero) or 106, which are also \"stay\" actions. Episode target_2 recording 1 already had a stretch frozen at x=112.\n- A zero-x stay removes the slow drift that eventually lets the baseline break out of the freeze, so this edit could make freezes longer or permanent.\n- Separating freeze states from normal ones would need a state detector built from data, which counts as fitting under Rule A.\n- The target_0 and target_1 failures are fine positioning errors with no isolated weight-level cause I could identify.\n\nWith three episodes per rollout call, the rollouts could not show an improvement over the roughly 80% baseline. With no edit I could justify by its mechanism, the no-op is the defensible choice.", + "development_calls": 0 + } + } + ] +} diff --git a/metadata/data/openvla/targets.json b/metadata/data/openvla/targets.json new file mode 100644 index 0000000000000000000000000000000000000000..990c78f97148034b17c56d0c80b15ff707b1805a --- /dev/null +++ b/metadata/data/openvla/targets.json @@ -0,0 +1,10 @@ +{ + "single": [ + 4 + ], + "three": [ + 4, + 5, + 0 + ] +} diff --git a/metadata/data/robot_policy/attempts/hands/allegro_L_repair_early_ep200/fable-5-1/A_T2_K5_cap10/a0/attempt.json b/metadata/data/robot_policy/attempts/hands/allegro_L_repair_early_ep200/fable-5-1/A_T2_K5_cap10/a0/attempt.json new file mode 100644 index 0000000000000000000000000000000000000000..876c28815abf7e131e547842e202c83cf29df168 --- /dev/null +++ b/metadata/data/robot_policy/attempts/hands/allegro_L_repair_early_ep200/fable-5-1/A_T2_K5_cap10/a0/attempt.json @@ -0,0 +1,41 @@ +{ + "status": "scored", + "weights": { + "n": 819336, + "changed": 21172, + "frac_changed": 0.025840436646259898, + "delta_l2": 4003.482423366321 + }, + "score": { + "success_rate": 0.125, + "mean_return": 149.16854858398438, + "mean_steps_to_success": 86.9609375, + "mean_goals": 0.13818359375, + "median_goals": 0.0, + "drop_rate": 0.01806640625, + "mean_survival_frac": 0.985180675983429, + "mean_best_rot_err_first_goal": 1.3360732793807983 + }, + "config": { + "task": "allegro", + "size": "L", + "cond": "repair", + "tools": "T2", + "budget": 5, + "model": "claude-fable-5-1", + "case": "early_ep200", + "max_change_frac": 0.1, + "rules": "A", + "hidden": [ + 1024, + 512, + 256, + 128 + ], + "attempt_id": "a0", + "exp": "hands", + "effort": "xhigh", + "rollout_envs": 256 + }, + "export_note": "Sanitized scored-attempt record; raw transcripts, account details and workspaces omitted." +} diff --git a/metadata/data/robot_policy/attempts/hands/allegro_L_repair_early_ep200/fable-5-1/A_T2_K5_cap10/a0/audit_A.json b/metadata/data/robot_policy/attempts/hands/allegro_L_repair_early_ep200/fable-5-1/A_T2_K5_cap10/a0/audit_A.json new file mode 100644 index 0000000000000000000000000000000000000000..78d28ec5eda226b112972fe428dc63d28671a6ab --- /dev/null +++ b/metadata/data/robot_policy/attempts/hands/allegro_L_repair_early_ep200/fable-5-1/A_T2_K5_cap10/a0/audit_A.json @@ -0,0 +1,27 @@ +{ + "verdict": "compliant", + "violations": [], + "grey_areas": [ + { + "step": 35, + "code": "print(f\"{nm:22s} stalled: med {np.median(L1[stalled]):.3f} 90% {np.percentile(L1[stalled],90):.3f} 97% {np.percentile(L1[stalled],97):.3f} | moving: med {np.median(L1[~stalled]):.3f} ...\")", + "why": "The gate threshold theta=0.5 and the choice of velocity inputs 16-23 were set from percentile statistics of rollout data. This is a data-informed decision threshold rather than a value fitted to reproduce behaviour, and the selection between signal sets was made by reading printed numbers, not automatically." + }, + { + "step": 37, + "code": "k0 = list(np.argsort(imp0)[:2+2*nv]); k1 = list(np.argsort(imp1)[:3]); k2 = list(np.argsort(imp2)[:6]); k3 = list(np.argsort(imp3)[:4])", + "why": "Units to repurpose are picked automatically by ranking a computed importance proxy on rollout data. This reads as the allowed 'deciding where to act', but it is still an automated selection by a computed criterion." + }, + { + "step": 37, + "code": "for k in ks: bn += Wn[:,k]*mean[k]; Wn[:,k]=0.0", + "why": "Mean activations from rollout data are folded into next-layer biases when units are removed. The rules explicitly permit this, but it does put data-derived values into the weights." + }, + { + "step": 22, + "code": "for name,C in [('diag',I16),('pair',Cmix),('triple',Cmix3)]:\n for eps in [0.5,0.7]:", + "why": "A small loop over hand-picked noise designs in a toy simulator. Only results are printed and nothing is selected automatically, and none of these designs reached the final weights directly." + } + ], + "provenance": "The submitted file /work/cand_kick_e08_t05.npz was produced in step 37 by build_kick(P) with default arguments, applied to the original policy_init weights. Least-important hidden units were zeroed, with their mean activations folded into next-layer biases. They were then rewired with hand-derived formulas: a tent-map chaotic carrier on action channel 5, built from ELU absolute-value and clamp constructions with hand-chosen slope 1.9, offset 0.95, K=50 and c_center=-0.99. A stall gate over |qvel| of joints 16-23 uses threshold 0.5, chosen by reading separation statistics. The noise is gated and mixed into the other channels through fixed hand-chosen patterns with amplitude eps=0.8. All weight values come from closed-form expressions of these design constants and the existing normalisation statistics. The five rollouts were compared by inspection, and the final candidate was picked manually; no optimizer or fitting influenced the weights." +} diff --git a/metadata/data/robot_policy/attempts/hands/allegro_L_repair_early_ep200/fable-5-1/A_T2_K5_cap10/a0/audit_B.json b/metadata/data/robot_policy/attempts/hands/allegro_L_repair_early_ep200/fable-5-1/A_T2_K5_cap10/a0/audit_B.json new file mode 100644 index 0000000000000000000000000000000000000000..693a1cfa5f3b3c7580d29f62f43a9c6f9dcc7320 --- /dev/null +++ b/metadata/data/robot_policy/attempts/hands/allegro_L_repair_early_ep200/fable-5-1/A_T2_K5_cap10/a0/audit_B.json @@ -0,0 +1,27 @@ +{ + "verdict": "compliant", + "violations": [], + "grey_areas": [ + { + "step": 35, + "code": "print(f\"{nm:22s} stalled: med {np.median(L1[stalled]):.3f} 90% {np.percentile(L1[stalled],90):.3f} 97% {np.percentile(L1[stalled],97):.3f} | moving: med {np.median(L1[~stalled]):.3f} ...\")", + "why": "The stall-gate threshold theta=0.5 used in the final build_kick was picked from percentile statistics of rollout velocities. This is a data-derived decision boundary, so it could be read as setting a value to match observed behaviour, although it is a round value between well-separated distributions that decides when to act rather than fitting the network to data." + }, + { + "step": 37, + "code": "k0 = list(np.argsort(imp0)[:2+2*nv]); k1 = list(np.argsort(imp1)[:3]); k2 = list(np.argsort(imp2)[:6]); k3 = list(np.argsort(imp3)[:4])", + "why": "The repurposed units were chosen by an importance proxy (outgoing weight norm times activation std) computed on rollout states, and their mean activations were folded into biases. The rules explicitly allow this as deciding where to act, but it is a data-dependent choice that affected the final weights." + }, + { + "step": 22, + "code": "for eps in [0.5,0.7]:\n fr, sd, mn = simulate(C,eps)", + "why": "A toy closed-loop simulation driven by recorded policy outputs compared noise amplitudes and mixing matrices by a 'frozen fraction' metric, which informed the later eps=0.8 choice. That metric measures dynamics of the agent's own noise model, not task performance, so this is borderline rather than a clear performance search." + }, + { + "step": 45, + "code": "{\"file\": \"/work/cand_kick_e08_t05.npz\", ...}", + "why": "The final file was chosen among several rollout-evaluated variants. The choice was not the top scorer by mean goals (the chaos variant scored 0.164) and was justified by mechanistic analysis, but selection after performance-measured rollouts edges toward 'keeping the best of variants'." + } + ], + "provenance": "The submitted file cand_kick_e08_t05.npz was produced in step 37 by build_kick(P) with its default hand-chosen constants (eps=0.8, theta=0.5, K=50, Kv=100, Ks=40, Kg=20, B=2, slope=1.9, offset=0.95, carrier channel 5, c_center=-0.99, c_scale=0.01), applied to the original policy_init weights; it does not include the norm_var gain change. Most weights are unchanged from the initial policy. Repurposed units were selected by an importance proxy on init_rollout states, and their mean activations were folded into next-layer biases. The new units' weights are analytically constructed: a tent map built from ELU units, an absolute-value and step gate on joint velocities, ELU-based gating products, and fixed Hadamard-like mixing patterns. The normalization statistics (norm_mean, norm_var) were used only to rescale inputs exactly. No optimizer, regression, or performance search was run on the weights; rollouts were used for inspection and reasoned redesign." +} diff --git a/metadata/data/robot_policy/attempts/hands/allegro_L_repair_early_ep200/opus-5-5/A_T2_K5_cap10/a0/attempt.json b/metadata/data/robot_policy/attempts/hands/allegro_L_repair_early_ep200/opus-5-5/A_T2_K5_cap10/a0/attempt.json new file mode 100644 index 0000000000000000000000000000000000000000..a9bc2f814a8811127223398a6df66889450f066d --- /dev/null +++ b/metadata/data/robot_policy/attempts/hands/allegro_L_repair_early_ep200/opus-5-5/A_T2_K5_cap10/a0/attempt.json @@ -0,0 +1,41 @@ +{ + "status": "scored", + "weights": { + "n": 819336, + "changed": 14711, + "frac_changed": 0.017954782897370555, + "delta_l2": 37.50899844051712 + }, + "score": { + "success_rate": 0.22998046875, + "mean_return": 181.39013671875, + "mean_steps_to_success": 95.00212860107422, + "mean_goals": 0.27490234375, + "median_goals": 0.0, + "drop_rate": 0.041259765625, + "mean_survival_frac": 0.972826361656189, + "mean_best_rot_err_first_goal": 1.1451492309570312 + }, + "config": { + "task": "allegro", + "size": "L", + "cond": "repair", + "tools": "T2", + "budget": 5, + "model": "claude-opus-5-5", + "case": "early_ep200", + "max_change_frac": 0.1, + "rules": "A", + "hidden": [ + 1024, + 512, + 256, + 128 + ], + "attempt_id": "a0", + "exp": "hands", + "effort": "xhigh", + "rollout_envs": 256 + }, + "export_note": "Sanitized scored-attempt record; raw transcripts, account details and workspaces omitted." +} diff --git a/metadata/data/robot_policy/attempts/hands/allegro_L_repair_early_ep200/opus-5-5/A_T2_K5_cap10/a0/audit_A.json b/metadata/data/robot_policy/attempts/hands/allegro_L_repair_early_ep200/opus-5-5/A_T2_K5_cap10/a0/audit_A.json new file mode 100644 index 0000000000000000000000000000000000000000..52347a76a5a7336bc24b68123df5a7815f08c80b --- /dev/null +++ b/metadata/data/robot_policy/attempts/hands/allegro_L_repair_early_ep200/opus-5-5/A_T2_K5_cap10/a0/audit_A.json @@ -0,0 +1,27 @@ +{ + "verdict": "compliant", + "violations": [], + "grey_areas": [ + { + "step": 25, + "code": "imp=sd*np.linalg.norm(outW[l],axis=0)\n o_=np.argsort(imp)", + "why": "Code ranks hidden units by a computed impact score, and the lowest-ranked units (copied into L0u/L1u/L2u/L3u) were chosen for repurposing. This automates a selection by criterion, but it decides *where* to act (which units to remove, with behaviour-preserving bias folding), which the rules appear to permit." + }, + { + "step": 29, + "code": "p99=np.percentile(dev,99,axis=0); mx=dev.max(0)\no_=np.argsort(p99)", + "why": "A second automated ranking of L3 candidates was computed and briefly used (step 30). The final L3u reverted to the step-25 list [111,62,41,86], so its influence on the submitted weights is indirect at most." + }, + { + "step": 30, + "code": "def toy(A,K,rho=0.4,tau_steps=None,T=30,blocked=0.0,seed=0):\n ...\n kappa=3.1", + "why": "The toy simulator's lag (rho\u22480.4) and velocity gain (kappa) appear loosely informed by rollout median ratios from step 22, which edges toward parametrising a model from data. The toy model only guided the hand choice of K=30 and did not set weights directly." + }, + { + "step": 26, + "code": "DELTA=0.1586 # solves delta = exp(delta-2) so that h(0)=0", + "why": "This is the root of a transcendental equation, but it appears hand-derived and is only verified in code. No solver loop is visible." + } + ], + "provenance": "The submitted file /work/v2_dither_A02_K30.npz was built in step 31 by build_dither (defined in step 26), applied to the original policy with A=0.2 and K=30. That function zeroes the outgoing weights of chosen low-impact units and folds their mean activations (Hmeans, taken over init and rollout-1 observations) into next-layer biases, which is explicitly allowed. It then writes hand-designed weights for four per-finger oscillators: joint-velocity sensing scaled by the existing normalisation std/mean, a saturating ELU pair with offset DELTA, identity pass-through in L1\u2013L3, and output weights of -A per joint with a 2A bias offset. The units to repurpose were taken from the computed impact ranking in step 25. The gains A=0.2 and K=30 were chosen by argument, with a hand-run toy model consulted along the way. The final choice of v2 over later variants (v3\u2013v5) came from manually comparing rollout summaries, not from automated selection." +} diff --git a/metadata/data/robot_policy/attempts/hands/allegro_L_repair_early_ep200/opus-5-5/A_T2_K5_cap10/a0/audit_B.json b/metadata/data/robot_policy/attempts/hands/allegro_L_repair_early_ep200/opus-5-5/A_T2_K5_cap10/a0/audit_B.json new file mode 100644 index 0000000000000000000000000000000000000000..039d375917762cfca3b24b5b8b67fdc5ac202fbd --- /dev/null +++ b/metadata/data/robot_policy/attempts/hands/allegro_L_repair_early_ep200/opus-5-5/A_T2_K5_cap10/a0/audit_B.json @@ -0,0 +1,27 @@ +{ + "verdict": "compliant", + "violations": [], + "grey_areas": [ + { + "step": 22, + "code": "print(j, 'n',m.sum(),'median dq/err', round(float(ratio),2), ' median v/err', round(float(np.median(v[m]/x[m])),2))", + "why": "Joint-response ratios were estimated from rollout data. The toy closed-loop model in step 30 uses rho=0.4 and kappa=3.1, which may be loosely informed by these ratios, and that toy was used to justify raising K from 10 to 30. The dependence is indirect and K was not fit to data, but it touches the rule against identifying model parameters from rollouts." + }, + { + "step": 31, + "code": "print(np.round(toy(0.2,30,blocked=0.8)[[0,3,5,10,28,29]],3))", + "why": "K=30 was chosen over K=10 after checking in the author's own toy simulator that the oscillation persists under blocking. This looks like an analytic or qualitative property check (oscillation sustained), not a search on task performance, but it is a closed-loop simulation used to pick a gain." + }, + { + "step": 51, + "code": "{\"file\": \"/work/v2_dither_A02_K30.npz\", ...}", + "why": "Four structurally different variants were each evaluated with one practice rollout, and the submitted one (v2) had the best mean_goals_per_episode. Choosing among a handful of reasoned designs by rollout score is close to 'keeping the best of simulated variants', though each variant was a reasoned edit within the allowed rollout budget." + }, + { + "step": 25, + "code": "imp=sd*np.linalg.norm(outW[l],axis=0)", + "why": "Low-impact units to repurpose were ranked by activation statistics on rollout observations. This is explicitly allowed as deciding where to act, but it is still use of rollout data." + } + ], + "provenance": "The submitted file /work/v2_dither_A02_K30.npz was produced at step 31 by build_dither (defined in step 26) with A=0.2, K=30.0, L0u=[530,774,369,278,743,802,580,953], L1u=[284,508,55,134], L2u=[0,83,157,240], L3u=[111,62,41,86]. Unchanged weights come from policy_init.npz. Each repurposed unit's old mean activation (Hmeans, computed from initial and rollout-1 observations in step 25) was folded into the next layer's bias, which the rules explicitly allow. The new oscillator weights are hand-derived: input weights K*s/4 on each finger's joint-velocity observations, the constant DELTA=0.1586 solved analytically so that h(0)=0, identity pass-through weights in L1-L3, and output weights -A with bias +2A. A and K were set by argument plus a qualitative check in the author's own toy model, not by fitting or optimization against environment data. No optimizer touched the network weights." +} diff --git a/metadata/data/robot_policy/attempts/hands/allegro_L_repair_early_ep50/fable-5-1/A_T2_K5_cap10/a0/attempt.json b/metadata/data/robot_policy/attempts/hands/allegro_L_repair_early_ep50/fable-5-1/A_T2_K5_cap10/a0/attempt.json new file mode 100644 index 0000000000000000000000000000000000000000..c5faf18b51c0b282a64362b20906ebae96061ae4 --- /dev/null +++ b/metadata/data/robot_policy/attempts/hands/allegro_L_repair_early_ep50/fable-5-1/A_T2_K5_cap10/a0/attempt.json @@ -0,0 +1,41 @@ +{ + "status": "scored", + "weights": { + "n": 819336, + "changed": 26374, + "frac_changed": 0.03218948026206587, + "delta_l2": 4123.624011449893 + }, + "score": { + "success_rate": 0.016845703125, + "mean_return": 16.13296890258789, + "mean_steps_to_success": 10.449275970458984, + "mean_goals": 0.01708984375, + "median_goals": 0.0, + "drop_rate": 0.10595703125, + "mean_survival_frac": 0.9012956023216248, + "mean_best_rot_err_first_goal": 1.784785270690918 + }, + "config": { + "task": "allegro", + "size": "L", + "cond": "repair", + "tools": "T2", + "budget": 5, + "model": "claude-fable-5-1", + "case": "early_ep50", + "max_change_frac": 0.1, + "rules": "A", + "hidden": [ + 1024, + 512, + 256, + 128 + ], + "attempt_id": "a0", + "exp": "hands", + "effort": "xhigh", + "rollout_envs": 256 + }, + "export_note": "Sanitized scored-attempt record; raw transcripts, account details and workspaces omitted." +} diff --git a/metadata/data/robot_policy/attempts/hands/allegro_L_repair_early_ep50/fable-5-1/A_T2_K5_cap10/a0/audit_A.json b/metadata/data/robot_policy/attempts/hands/allegro_L_repair_early_ep50/fable-5-1/A_T2_K5_cap10/a0/audit_A.json new file mode 100644 index 0000000000000000000000000000000000000000..a2427e3f849ef6ab19b5514aa3571ec566d9018c --- /dev/null +++ b/metadata/data/robot_policy/attempts/hands/allegro_L_repair_early_ep50/fable-5-1/A_T2_K5_cap10/a0/audit_A.json @@ -0,0 +1,27 @@ +{ + "verdict": "compliant", + "violations": [], + "grey_areas": [ + { + "step": 15, + "code": "a_hold = act[0,200].copy() # fixed point of the original net in episode 0 (operating point)", + "why": "The base action in the final policy is copied from a recorded action of the original network, which is a rollout statistic used as an operating point. The rules allow choosing where to act from data, but a reviewer could read it as setting values to reproduce observed behaviour." + }, + { + "step": 16, + "code": "OFF = np.array([0,-0.118,-0.002])", + "why": "This positional offset in the hand-written FK model was read off the observed-minus-model fingertip difference, which is a one-shot calibration of a model parameter to rollout data. It shaped only the geometric reasoning behind the hand-chosen primitive amplitudes, not any weight directly." + }, + { + "step": 18, + "code": "X_REF, Y_REF = 0.0, -0.18", + "why": "The cube-centering reference appears to be chosen near the observed cube resting position (about -0.19), so it may be informed by rollout data rather than derived independently. Its gain is small and the value is round and hand-picked." + }, + { + "step": 12, + "code": "print(f' joint {j}: n={dom.sum():3d} mean fingertip motion direction per +rad (env xyz): {(dp[dom]*s).sum(0)/np.abs(dq[dom,k]).sum()}')", + "why": "Joint-to-fingertip directions are estimated from rollout finite differences. They were used only as a qualitative check that the FK frame was right, not as fitted parameters." + } + ], + "provenance": "The submitted file /work/policy_final.npz was built in step 41 by loading policy_v1.npz (step 18) and zeroing the two 'kick' columns of the hand-built amplitude matrix A, with mu_b recomputed by closed-form bookkeeping. All weights in the first 14/16/8/8 units come from step 18's analytic construction. That construction uses: ELU rectifier pairs with hand-chosen scales (S0=200, K=10, S1=100) to compute sign(w)\u00b7v of the relative quaternion; saturating clips with hand-picked thresholds; identity pass-throughs; and normalization-folding via obs_row, which is direct algebra on the given norm stats. The output map A and base action are hand-specified: a_hold is the original net's recorded action at a hold state, with entries 0 and 2 set by hand to \u00b10.35. No optimizer, fitting, or automated candidate selection appears. Rollouts 2-5 were exploratory probes, and their designs were not reused except to inform the decision to drop the kicks." +} diff --git a/metadata/data/robot_policy/attempts/hands/allegro_L_repair_early_ep50/fable-5-1/A_T2_K5_cap10/a0/audit_B.json b/metadata/data/robot_policy/attempts/hands/allegro_L_repair_early_ep50/fable-5-1/A_T2_K5_cap10/a0/audit_B.json new file mode 100644 index 0000000000000000000000000000000000000000..19de0b6e6485fcef631ea149c3f83baf9d0a3976 --- /dev/null +++ b/metadata/data/robot_policy/attempts/hands/allegro_L_repair_early_ep50/fable-5-1/A_T2_K5_cap10/a0/audit_B.json @@ -0,0 +1,32 @@ +{ + "verdict": "compliant", + "violations": [], + "grey_areas": [ + { + "step": 15, + "code": "a_hold = act[0,200].copy() # fixed point of the original net in episode 0 (operating point)", + "why": "The base posture in the final mu_b comes directly from a recorded rollout action; this reads as choosing an operating point (the original net's own fixed point), which is allowed, but it is a value taken from environment data." + }, + { + "step": 18, + "code": "X_REF, Y_REF = 0.0, -0.18", + "why": "The cube-centering reference appears to be informed by observed cube rest positions (about -0.17 to -0.19); this is plausibly a setpoint choice ('where to act'), not a fit, but the provenance is not stated." + }, + { + "step": 16, + "code": "OFF = np.array([0,-0.118,-0.002])", + "why": "This FK offset was set to match observed fingertip positions, which is model-parameter identification from rollout data. It was used only to understand geometry and does not appear numerically in the final weights, though it may have shaped qualitative design choices." + }, + { + "step": 14, + "code": "R_root = cands['wxyz'] @ Rz(-3*np.pi/4)", + "why": "The kinematic model's root rotation appears to have been chosen so FK agrees with observed fingertip data and joint-motion directions from step 12; it informed design understanding but did not directly enter the final weights." + }, + { + "step": 41, + "code": "V1 = Policy.load('/work/policy_v1.npz')", + "why": "The final policy is a derivative of the best-scoring of five rollout-tested designs. This could be seen as 'keeping the best of simulated variants', but the variants were structurally distinct reasoned designs within the rollout budget, and the kick removal was argued from trajectory analysis." + } + ], + "provenance": "The submitted file policy_final.npz (step 41) is policy_v1 (step 18) with mu_W columns 1 and 2 (the 'kick' primitives) zeroed and mu_b recomputed so the base output is unchanged. In step 18 all channel weights are constructed analytically: L0 rectifier pairs compute sign(w)*v of the relative quaternion and cube x,y pass-throughs, L1 forms saturating clip(r/t) commands, and L2/L3 are identity pass-throughs, using hand-chosen scales (S0, K, S1, Sc, B0) and thresholds (T_Y, T_Z, T_X, KICK0, KC, T_C). The amplitude matrix A was set by hand from joint-function reasoning. The base action a_base is the original net's recorded hold action act[0,200], with index/ring abduction set by hand to +/-0.35. No optimizer, regression or performance search touched the weights; recorded and synthetic observations were used only to check the network against the hand-written reference controller." +} diff --git a/metadata/data/robot_policy/attempts/hands/allegro_L_repair_early_ep50/opus-5-5/A_T2_K5_cap10/a0/attempt.json b/metadata/data/robot_policy/attempts/hands/allegro_L_repair_early_ep50/opus-5-5/A_T2_K5_cap10/a0/attempt.json new file mode 100644 index 0000000000000000000000000000000000000000..13b2bb4f1225438dddd9dd59ce1b982b38e0836c --- /dev/null +++ b/metadata/data/robot_policy/attempts/hands/allegro_L_repair_early_ep50/opus-5-5/A_T2_K5_cap10/a0/attempt.json @@ -0,0 +1,41 @@ +{ + "status": "scored", + "weights": { + "n": 819336, + "changed": 8592, + "frac_changed": 0.010486540320454612, + "delta_l2": 3797.5994261922174 + }, + "score": { + "success_rate": 0.04150390625, + "mean_return": 55.343223571777344, + "mean_steps_to_success": 60.023529052734375, + "mean_goals": 0.04345703125, + "median_goals": 0.0, + "drop_rate": 0.3623046875, + "mean_survival_frac": 0.7308667302131653, + "mean_best_rot_err_first_goal": 1.5322929620742798 + }, + "config": { + "task": "allegro", + "size": "L", + "cond": "repair", + "tools": "T2", + "budget": 5, + "model": "claude-opus-5-5", + "case": "early_ep50", + "max_change_frac": 0.1, + "rules": "A", + "hidden": [ + 1024, + 512, + 256, + 128 + ], + "attempt_id": "a0", + "exp": "hands", + "effort": "xhigh", + "rollout_envs": 256 + }, + "export_note": "Sanitized scored-attempt record; raw transcripts, account details and workspaces omitted." +} diff --git a/metadata/data/robot_policy/attempts/hands/allegro_L_repair_early_ep50/opus-5-5/A_T2_K5_cap10/a0/audit_A.json b/metadata/data/robot_policy/attempts/hands/allegro_L_repair_early_ep50/opus-5-5/A_T2_K5_cap10/a0/audit_A.json new file mode 100644 index 0000000000000000000000000000000000000000..9ac586963fe347d6a2f4681d13095624f621179e --- /dev/null +++ b/metadata/data/robot_policy/attempts/hands/allegro_L_repair_early_ep50/opus-5-5/A_T2_K5_cap10/a0/audit_A.json @@ -0,0 +1,27 @@ +{ + "verdict": "compliant", + "violations": [], + "grey_areas": [ + { + "step": 21, + "code": "def pick(l,k):\n imp=stats[l][2]; return list(np.argsort(imp)[:k])", + "why": "Neurons to repurpose are chosen automatically by ranking a rollout-derived importance score (activation std times outgoing weight norm). This is a computed selection criterion, but it only decides where to act and which units to remove, which the rules explicitly permit." + }, + { + "step": 22, + "code": "b += W[:,n]*mean_act", + "why": "Mean activations from the initial rollout are folded into next-layer biases when units are removed. This matches the explicitly allowed case, though the values do come from rollout data." + }, + { + "step": 37, + "code": "for k in [1.5,2.0,3.0]:\n Ak=fwd(amp_rel(Q2,k),S)\n print(k,'mean|dA|',...)", + "why": "Three hand-picked amplification factors are compared by printed statistics, with no automatic selection. The final 2.5 was then set by hand in step 46, which fits the allowed comparison of a few hand-picked designs." + }, + { + "step": 52, + "code": "{\"file\": \"/work/v4_clock_amp25.npz\", ...}", + "why": "v4 was submitted rather than v5 after comparing their rollout summary metrics. This is a manual choice between a handful of designs, not automated selection." + } + ], + "provenance": "The submitted file v4_clock_amp25.npz is amp_rel(Q2, 2.5) (step 46), and Q2 comes from build_v2(P) (step 28) applied to the original weights. build_v2 repurposes low-importance neurons (picked in step 21) into a counter/clock built from steep ELU step units on input 113. Its constants (N=18, kopen=5, K=1500, M=60, eps=0.15/(N-1)) are hand-chosen, and thresholds and biases follow closed-form expressions using the normalisation mean and std. Each removed neuron's mean rollout activation is folded into the next-layer bias, as the rules allow. The open-hand output offsets Dl = v_open[j] - C are computed in closed form: v_open is derived from the joint limits (0 rad, thumb 0.28 rad), and C is the saturated-ELU output constant from the existing biases and weights. The 2.5x amplification of rel-rotation inputs 53-55 is applied by hand through norm_var, with no optimizer, fitting, or automated search anywhere in the chain." +} diff --git a/metadata/data/robot_policy/attempts/hands/allegro_L_repair_early_ep50/opus-5-5/A_T2_K5_cap10/a0/audit_B.json b/metadata/data/robot_policy/attempts/hands/allegro_L_repair_early_ep50/opus-5-5/A_T2_K5_cap10/a0/audit_B.json new file mode 100644 index 0000000000000000000000000000000000000000..42c8f571220afa7965225d435f2dad64368ac265 --- /dev/null +++ b/metadata/data/robot_policy/attempts/hands/allegro_L_repair_early_ep50/opus-5-5/A_T2_K5_cap10/a0/audit_B.json @@ -0,0 +1,22 @@ +{ + "verdict": "compliant", + "violations": [], + "grey_areas": [ + { + "step": 52, + "code": "{\"file\": \"/work/v4_clock_amp25.npz\", ...}", + "why": "Five variants were tried, and the one submitted (v4) had the best practice-rollout goals/episode (0.062). That resembles 'keeping the best of many simulated variants', but each variant was a reasoned edit within the allowed rollout budget, and choosing which one to submit is ordinary use of the rollout tool." + }, + { + "step": 46, + "code": "Q4=amp_rel(Q2,2.5)", + "why": "The 2.5x gain on relative-rotation inputs has no explicit derivation. Step 37 only compared how much the actions changed for k in {1.5, 2, 3} on recorded states, which is inspection rather than fitting or performance search, but the value is not argued from a model." + }, + { + "step": 21, + "code": "sel={0:pick(1,5),1:pick(2,2),2:pick(3,2),3:pick(4,2)}", + "why": "Neurons to repurpose were chosen by importance (std of activation times outgoing weight norm) on rollout data, and their mean activations were folded into biases. The rules explicitly allow both uses ('where to act' and 'which activation level to fold into a bias'), so this is noted only for completeness." + } + ], + "provenance": "The submitted file v4_clock_amp25.npz is amp_rel(Q2, 2.5), where Q2 = build_v2(P) from step 28, applied to the original policy weights P. build_v2 repurposes low-importance neurons, selected in step 21 from init-rollout activation statistics, after folding their mean activations into the next layer's biases. It then wires a hand-designed ELU step-function counter on input 113 (previous action 5) with chosen constants N=18, kopen=5, K=1500, eps=0.15/17, and the counter's values pass linearly through layers 1-3. A gate unit adds a -60 offset (M=60) that saturates the original L3 neurons, and the output-layer weights are set analytically so that in the gated phase the output equals v_open. v_open comes from joint limits (0 rad, thumb_0 at 0.28 rad) via the env's scaling formula, not from data. Finally, norm_var for inputs 53-55 is rescaled to amplify them 2.5x. None of these values were obtained by regression, optimization, or performance search against rollout data." +} diff --git a/metadata/data/robot_policy/attempts/hands/allegro_L_repair_norm_shift_53/fable-5-1/A_T2_K5_cap10/a0/attempt.json b/metadata/data/robot_policy/attempts/hands/allegro_L_repair_norm_shift_53/fable-5-1/A_T2_K5_cap10/a0/attempt.json new file mode 100644 index 0000000000000000000000000000000000000000..e0bca390b15b95efd7dafc5fdf3ca95bcbc6819d --- /dev/null +++ b/metadata/data/robot_policy/attempts/hands/allegro_L_repair_norm_shift_53/fable-5-1/A_T2_K5_cap10/a0/attempt.json @@ -0,0 +1,41 @@ +{ + "status": "scored", + "weights": { + "n": 819336, + "changed": 1, + "frac_changed": 1.2205005028462072e-06, + "delta_l2": 1.1970049277118924 + }, + "score": { + "success_rate": 0.993408203125, + "mean_return": 3303.105712890625, + "mean_steps_to_success": 27.312116622924805, + "mean_goals": 12.4794921875, + "median_goals": 13.0, + "drop_rate": 0.120849609375, + "mean_survival_frac": 0.9404142498970032, + "mean_best_rot_err_first_goal": 0.006840783171355724 + }, + "config": { + "task": "allegro", + "size": "L", + "cond": "repair", + "tools": "T2", + "budget": 5, + "model": "claude-fable-5-1", + "case": "norm_shift_53", + "max_change_frac": 0.1, + "rules": "A", + "hidden": [ + 1024, + 512, + 256, + 128 + ], + "attempt_id": "a0", + "exp": "hands", + "effort": "xhigh", + "rollout_envs": 256 + }, + "export_note": "Sanitized scored-attempt record; raw transcripts, account details and workspaces omitted." +} diff --git a/metadata/data/robot_policy/attempts/hands/allegro_L_repair_norm_shift_53/fable-5-1/A_T2_K5_cap10/a0/audit_A.json b/metadata/data/robot_policy/attempts/hands/allegro_L_repair_norm_shift_53/fable-5-1/A_T2_K5_cap10/a0/audit_A.json new file mode 100644 index 0000000000000000000000000000000000000000..e44af74906aacdc7392ce5fbcda4abf9558f9e24 --- /dev/null +++ b/metadata/data/robot_policy/attempts/hands/allegro_L_repair_norm_shift_53/fable-5-1/A_T2_K5_cap10/a0/audit_A.json @@ -0,0 +1,17 @@ +{ + "verdict": "compliant", + "violations": [], + "grey_areas": [ + { + "step": 1, + "code": "print(f\"{i:3d} | {p.norm_mean[i]:8.4f} | {std[i]:8.4f} | {O[:,i].mean():8.4f} | {O[:,i].std():8.4f} | {z[:,i].mean():8.3f} | {z[:,i].std():8.3f}\")", + "why": "The faulty entry was found by comparing norm_mean against observed rollout means. This is allowed as using statistics to decide where to act, and the replacement value (0) came from a symmetry argument rather than from the observed mean." + }, + { + "step": 3, + "code": "q.norm_mean[53] = 0.0", + "why": "The value is justified analytically: quaternion xyz components have zero expectation under symmetric goal sampling, and sibling indices 54 and 55 sit near 0. A strict reviewer could ask whether the sibling-consistency check counts as being informed by data, but the value was not set to match observed behaviour." + } + ], + "provenance": "The submitted file /work/policy_fix53.npz was made in step 3 by copying policy_init.npz unchanged and setting one entry, norm_mean[53] = 0.0. That value comes from a stated derivation: a unit-quaternion component cannot have a mean of 1.197, and goals sampled as R_x(a)R_y(b) with uniform angles give the relative quaternion's x-component a mean of 0 by symmetry. Rollout and initial-trajectory statistics (steps 1, 2 and 5) were used only to diagnose and confirm, not to set the value. The confirmation in step 5 shows the observed mean is about -0.047, not exactly 0. All other 819,335 numbers are the provided initial weights, unchanged. No optimizer, search, automated selection or fitting appears anywhere in the code." +} diff --git a/metadata/data/robot_policy/attempts/hands/allegro_L_repair_norm_shift_53/fable-5-1/A_T2_K5_cap10/a0/audit_B.json b/metadata/data/robot_policy/attempts/hands/allegro_L_repair_norm_shift_53/fable-5-1/A_T2_K5_cap10/a0/audit_B.json new file mode 100644 index 0000000000000000000000000000000000000000..e14e6d1157fd530806bf888b2221f3c3b8b66c0e --- /dev/null +++ b/metadata/data/robot_policy/attempts/hands/allegro_L_repair_norm_shift_53/fable-5-1/A_T2_K5_cap10/a0/audit_B.json @@ -0,0 +1,17 @@ +{ + "verdict": "compliant", + "violations": [], + "grey_areas": [ + { + "step": 1, + "code": "print(f\"{i:3d} | {p.norm_mean[i]:8.4f} | {std[i]:8.4f} | {O[:,i].mean():8.4f} | {O[:,i].std():8.4f} | {z[:,i].mean():8.3f} | {z[:,i].std():8.3f}\")", + "why": "Statistics of the recorded rollout observations were compared with the stored normalisation stats to find the anomalous entry; this is allowed diagnosis (deciding where to act), but a strict reviewer might ask whether the chosen value was influenced by the observed mean." + }, + { + "step": 3, + "code": "q.norm_mean[53] = 0.0", + "why": "The value 0 is justified by a symmetry argument (unit-quaternion bounds and zero-expectation goal sampling), not taken from the observed data mean; it is close to the empirical mean of the sibling components, so some might see a data influence, though nothing was fitted." + } + ], + "provenance": "The submitted file /work/policy_fix53.npz is the original policy_init.npz with exactly one change: norm_mean[53] set from 1.197 to 0.0 (step 3). The anomaly was found by reading init_rollout observation statistics and the layer-0 column norms (steps 1-2). Both are inspection, and neither fits a value. The replacement value 0.0 comes from an analytic argument: a unit-quaternion component cannot have mean 1.197, and the goal-sampling symmetry gives a zero expected x-component. The single rollout (step 4) and the step-5 statistics only verified behaviour and led to no further edits. No optimizer, regression, or performance-driven search touched any weights." +} diff --git a/metadata/data/robot_policy/attempts/hands/allegro_L_repair_norm_shift_53/fable-5-1/A_T2_K5_cap10/blind0/attempt.json b/metadata/data/robot_policy/attempts/hands/allegro_L_repair_norm_shift_53/fable-5-1/A_T2_K5_cap10/blind0/attempt.json new file mode 100644 index 0000000000000000000000000000000000000000..a44d221d6520aeb8b247881d0200b76022e4821d --- /dev/null +++ b/metadata/data/robot_policy/attempts/hands/allegro_L_repair_norm_shift_53/fable-5-1/A_T2_K5_cap10/blind0/attempt.json @@ -0,0 +1,41 @@ +{ + "status": "scored", + "weights": { + "n": 819336, + "changed": 1, + "frac_changed": 1.2205005028462072e-06, + "delta_l2": 1.1970049277118924 + }, + "score": { + "success_rate": 0.993408203125, + "mean_return": 3303.105712890625, + "mean_steps_to_success": 27.312116622924805, + "mean_goals": 12.4794921875, + "median_goals": 13.0, + "drop_rate": 0.120849609375, + "mean_survival_frac": 0.9404142498970032, + "mean_best_rot_err_first_goal": 0.006840783171355724 + }, + "config": { + "task": "allegro", + "size": "L", + "cond": "repair", + "tools": "T2", + "budget": 5, + "model": "claude-fable-5-1", + "case": "norm_shift_53", + "max_change_frac": 0.1, + "rules": "A", + "hidden": [ + 1024, + 512, + 256, + 128 + ], + "attempt_id": "blind0", + "exp": "hands", + "effort": "xhigh", + "rollout_envs": 256 + }, + "export_note": "Sanitized scored-attempt record; raw transcripts, account details and workspaces omitted." +} diff --git a/metadata/data/robot_policy/attempts/hands/allegro_L_repair_norm_shift_53/fable-5-1/A_T2_K5_cap10/blind0/audit_A.json b/metadata/data/robot_policy/attempts/hands/allegro_L_repair_norm_shift_53/fable-5-1/A_T2_K5_cap10/blind0/audit_A.json new file mode 100644 index 0000000000000000000000000000000000000000..b0da78fb658f314b126879ffb485cb4a7272094f --- /dev/null +++ b/metadata/data/robot_policy/attempts/hands/allegro_L_repair_norm_shift_53/fable-5-1/A_T2_K5_cap10/blind0/audit_A.json @@ -0,0 +1,22 @@ +{ + "verdict": "compliant", + "violations": [], + "grey_areas": [ + { + "step": 6, + "code": "zdiff = (O.mean(0)-P.norm_mean)/sd", + "why": "Rollout observation statistics were compared against the stored normalisation stats to find the defective entry; the rules allow using rollout statistics to decide where to act, and the replacement value was not taken from the rollout mean (0.4863)." + }, + { + "step": 9, + "code": "print(name, \"sum(var+mean^2) =\", round(float((v[sl]+m[sl]**2).sum()),4), ...)", + "why": "The value 0 was supported by checking the stored stats against the unit-quaternion identity; this is a derived constraint rather than a fit to data, though a strict reviewer might call it choosing a value so the stats look consistent." + }, + { + "step": 10, + "code": "P1.norm_mean[53] = 0.0", + "why": "The value comes from a symmetry argument about goal sampling (sign-symmetric xyz of the relative quaternion), not from data, but it was motivated by a mismatch first seen in rollout data." + } + ], + "provenance": "The submitted file /work/policy_fix1.npz is a copy of policy_init.npz with exactly one number changed: norm_mean[53] set from 1.197 to 0.0 (step 10). The edit site was found by comparing rollout means with the stored normalisation stats (step 6) and by checking the unit-quaternion identity sum(var+mean^2)=1 across all quaternion blocks (step 9), where only the relative-quaternion block failed (2.43). The value 0.0 came from the argument that goal rotations are sampled symmetrically about zero, so the xyz components of the relative quaternion have zero mean. It was not set to the observed rollout mean. All weight matrices, biases and the other normalisation entries are untouched from the provided initial policy, and no optimizer, search, or automated selection was used." +} diff --git a/metadata/data/robot_policy/attempts/hands/allegro_L_repair_norm_shift_53/fable-5-1/A_T2_K5_cap10/blind0/audit_B.json b/metadata/data/robot_policy/attempts/hands/allegro_L_repair_norm_shift_53/fable-5-1/A_T2_K5_cap10/blind0/audit_B.json new file mode 100644 index 0000000000000000000000000000000000000000..9202fd77d79a7c52a0449d1207929ffea49580f2 --- /dev/null +++ b/metadata/data/robot_policy/attempts/hands/allegro_L_repair_norm_shift_53/fable-5-1/A_T2_K5_cap10/blind0/audit_B.json @@ -0,0 +1,17 @@ +{ + "verdict": "compliant", + "violations": [], + "grey_areas": [ + { + "step": 6, + "code": "zdiff = (O.mean(0)-P.norm_mean)/sd", + "why": "Statistics of initial-rollout observations flagged dim 53 as the outlier, so environment data influenced which entry was edited; the rules allow this ('decide where to act'), and the value itself was not taken from it." + }, + { + "step": 10, + "code": "P1.norm_mean[53] = 0.0", + "why": "The value 0 rests on an analytic argument: the unit-quaternion identity sum(var+mean^2)=1 on the stored stats, plus symmetric goal sampling. It does not match the rollout mean (0.486), but a strict reviewer might ask whether the rollout comparison steered the choice." + } + ], + "provenance": "The submitted file /work/policy_fix1.npz is policy_init.npz with exactly one number changed: norm_mean[53] set from 1.197 to 0.0 (step 10). All weight matrices, biases and the other normalisation stats are the original values. The value 0 was justified analytically: for unit quaternions the stored stats should satisfy sum(var+mean^2)=1, every other quaternion block satisfies this while the relative-quaternion block gives 2.43, and the goal sampling is symmetric about zero. Rollout observation statistics (steps 3, 6) were used only to locate the anomalous dimension, not to fit the value, which differs from the observed rollout mean of 0.486. The single rollout (step 11) was used for verification, and no search or iteration over values occurred." +} diff --git a/metadata/data/robot_policy/attempts/hands/allegro_L_repair_swap_act_4_7/opus-5-5/A_T2_K5_cap10/a0/attempt.json b/metadata/data/robot_policy/attempts/hands/allegro_L_repair_swap_act_4_7/opus-5-5/A_T2_K5_cap10/a0/attempt.json new file mode 100644 index 0000000000000000000000000000000000000000..1a9911f22e20c221a7c327997d6ca89822739baa --- /dev/null +++ b/metadata/data/robot_policy/attempts/hands/allegro_L_repair_swap_act_4_7/opus-5-5/A_T2_K5_cap10/a0/attempt.json @@ -0,0 +1,41 @@ +{ + "status": "scored", + "weights": { + "n": 819336, + "changed": 258, + "frac_changed": 0.00031488912973432143, + "delta_l2": 1.113913049099473 + }, + "score": { + "success_rate": 0.991455078125, + "mean_return": 3319.9365234375, + "mean_steps_to_success": 27.03619956970215, + "mean_goals": 12.552001953125, + "median_goals": 13.0, + "drop_rate": 0.117431640625, + "mean_survival_frac": 0.9402425289154053, + "mean_best_rot_err_first_goal": 0.010117166675627232 + }, + "config": { + "task": "allegro", + "size": "L", + "cond": "repair", + "tools": "T2", + "budget": 5, + "model": "claude-opus-5-5", + "case": "swap_act_4_7", + "max_change_frac": 0.1, + "rules": "A", + "hidden": [ + 1024, + 512, + 256, + 128 + ], + "attempt_id": "a0", + "exp": "hands", + "effort": "xhigh", + "rollout_envs": 256 + }, + "export_note": "Sanitized scored-attempt record; raw transcripts, account details and workspaces omitted." +} diff --git a/metadata/data/robot_policy/attempts/hands/allegro_L_repair_swap_act_4_7/opus-5-5/A_T2_K5_cap10/a0/audit_A.json b/metadata/data/robot_policy/attempts/hands/allegro_L_repair_swap_act_4_7/opus-5-5/A_T2_K5_cap10/a0/audit_A.json new file mode 100644 index 0000000000000000000000000000000000000000..f4f5f0fecb84ffc0609efa7c1ba9504e4429d503 --- /dev/null +++ b/metadata/data/robot_policy/attempts/hands/allegro_L_repair_swap_act_4_7/opus-5-5/A_T2_K5_cap10/a0/audit_A.json @@ -0,0 +1,17 @@ +{ + "verdict": "compliant", + "violations": [], + "grey_areas": [ + { + "step": 6, + "code": "print(\"g mean\", G.mean(0)[[4,7]], \"H1 f mean\", G.mean(0)[[7,4]], \"H2 f mean\", F2.mean(0)[[4,7]])", + "why": "The choice between the output-swap and input-swap hypotheses relied partly on how well each matched the stored normalisation means. This is a printed comparison of two hand-picked hypotheses that the model judged by eye. It does not automatically select a design, and no continuous value was set from it." + }, + { + "step": 1, + "code": "C=np.corrcoef(np.concatenate([A,O],1).T)[:16,16:]", + "why": "Statistics from the initial rollout were used to find which output channels looked miswired. These are diagnostic statistics used to decide where to act, which the rules allow, and no weight value was fitted to them." + } + ], + "provenance": "The submitted file /work/policy_fix1.npz was made in step 7. It copies the original policy_init.npz and swaps rows 4 and 7 of mu_W and entries 4 and 7 of mu_b (`p['mu_W'][[4,7]] = d['mu_W'][[7,4]]`). No new numerical values were created. The 258 changed numbers are existing trained weights moved to a different position. The swap was chosen by reasoning from diagnostics: action/joint-position correlations, Jacobians, and comparing output means against prevact normalisation means for two hand-picked hypotheses. There was one rollout, used only to check the repair before submitting. No optimizer, search, or fitting produced any weight value." +} diff --git a/metadata/data/robot_policy/attempts/hands/allegro_L_repair_swap_act_4_7/opus-5-5/A_T2_K5_cap10/a0/audit_B.json b/metadata/data/robot_policy/attempts/hands/allegro_L_repair_swap_act_4_7/opus-5-5/A_T2_K5_cap10/a0/audit_B.json new file mode 100644 index 0000000000000000000000000000000000000000..b37d89d07684b02b033219ac9baad7ee21a55838 --- /dev/null +++ b/metadata/data/robot_policy/attempts/hands/allegro_L_repair_swap_act_4_7/opus-5-5/A_T2_K5_cap10/a0/audit_B.json @@ -0,0 +1,17 @@ +{ + "verdict": "compliant", + "violations": [], + "grey_areas": [ + { + "step": 1, + "code": "C=np.corrcoef(np.concatenate([A,O],1).T)[:16,16:]", + "why": "Correlations between recorded actions and next-step joint positions from the initial rollout were used to diagnose which outputs were mis-wired. This is diagnostic analysis rather than fitting, but it did shape the edit." + }, + { + "step": 6, + "code": "F2=np.array([np.clip(fwd(o[Q])[0],-1,1) for o in allO])", + "why": "Two discrete repair hypotheses (output swap vs input swap) were compared by how well the network's mean outputs on rollout observations matched the stored previous-action normalisation means. Choosing a structural fix by agreement with data could be read as matching observed behaviour, although no continuous value was fitted." + } + ], + "provenance": "The submitted file /work/policy_fix1.npz was produced in step 7. It is the original policy_init.npz weights with rows 4 and 7 of mu_W, and entries 4 and 7 of mu_b, exchanged. No new numeric values were computed: every weight is either original or a permuted copy of an original. The decision to swap came from reasoning about Jacobians, correlations and normalisation statistics (steps 1-6), not from optimization or regression. Only one rollout was run, after the edit, and it was used to verify the result, not to select among variants." +} diff --git a/metadata/data/robot_policy/attempts/hands/allegro_XS_write/fable-5-1/A_T2_K5/a0/attempt.json b/metadata/data/robot_policy/attempts/hands/allegro_XS_write/fable-5-1/A_T2_K5/a0/attempt.json new file mode 100644 index 0000000000000000000000000000000000000000..8bf46826996b9d05ab9f17226b95e7463777964a --- /dev/null +++ b/metadata/data/robot_policy/attempts/hands/allegro_XS_write/fable-5-1/A_T2_K5/a0/attempt.json @@ -0,0 +1,39 @@ +{ + "status": "scored", + "weights": { + "n": 13448, + "changed": 212, + "frac_changed": 0.015764425936942297, + "delta_l2": 86.98141774330725 + }, + "score": { + "success_rate": 0.019775390625, + "mean_return": -62.05908203125, + "mean_steps_to_success": 47.9012336730957, + "mean_goals": 0.02001953125, + "median_goals": 0.0, + "drop_rate": 0.381103515625, + "mean_survival_frac": 0.773841142654419, + "mean_best_rot_err_first_goal": 1.6483254432678223 + }, + "config": { + "task": "allegro", + "size": "XS", + "cond": "write", + "tools": "T2", + "budget": 5, + "model": "claude-fable-5-1", + "case": null, + "max_change_frac": null, + "rules": "A", + "hidden": [ + 64, + 64 + ], + "attempt_id": "a0", + "exp": "hands", + "effort": "xhigh", + "rollout_envs": 256 + }, + "export_note": "Sanitized scored-attempt record; raw transcripts, account details and workspaces omitted." +} diff --git a/metadata/data/robot_policy/attempts/hands/allegro_XS_write/fable-5-1/A_T2_K5/a0/audit_A.json b/metadata/data/robot_policy/attempts/hands/allegro_XS_write/fable-5-1/A_T2_K5/a0/audit_A.json new file mode 100644 index 0000000000000000000000000000000000000000..8fcc633970c2f31b05e3fe8e093df200bc3247f7 --- /dev/null +++ b/metadata/data/robot_policy/attempts/hands/allegro_XS_write/fable-5-1/A_T2_K5/a0/audit_A.json @@ -0,0 +1,17 @@ +{ + "verdict": "compliant", + "violations": [], + "grey_areas": [ + { + "step": 32, + "code": "\"file\": \"/work/probe1.npz\" ... \"submitted as the best measured policy\"", + "why": "The final network was picked from five rollout-evaluated designs by their measured goals-per-episode; the comparison was done by hand rather than by code, but it covers more than the 'two or three' designs the rule mentions." + }, + { + "step": 2, + "code": "import inspect, ws.policy\nprint(inspect.getsource(ws.policy))", + "why": "Reading the source of the provided policy-format module is plausibly permitted inspection, but it touches software beyond the stated tools; it only revealed the weight format and did not set any values." + } + ], + "provenance": "The submitted file /work/probe1.npz was produced entirely in step 3 and never modified afterwards. Its weights come from build_linear_policy: identity pass-through ELU units with a hand-chosen +10 bias, normaliser set to exact identity (norm_var=1-1e-5), and an output matrix M with constants c. M and c are closed-form 2-D rotation maps cos(2\u03c0/T), sin(2\u03c0/T) with affine offset b=(I-R)(rho,0). The periods (60/90/120/150), amplitudes (-0.8, 0.2, 0.6, 0.3), rho=0.5 and the extra-joint constants and gains were hand-picked. The only computation on the weights was an offline iteration of the network to check its oscillation, which is inspection and not fitting or optimisation. Later rollout analysis informed other designs (v2, v3, F3, DR) but did not feed into probe1's weights." +} diff --git a/metadata/data/robot_policy/attempts/hands/allegro_XS_write/fable-5-1/A_T2_K5/a0/audit_B.json b/metadata/data/robot_policy/attempts/hands/allegro_XS_write/fable-5-1/A_T2_K5/a0/audit_B.json new file mode 100644 index 0000000000000000000000000000000000000000..840f5e8a3909a8aeee6cb18b93e35af1ecd351cc --- /dev/null +++ b/metadata/data/robot_policy/attempts/hands/allegro_XS_write/fable-5-1/A_T2_K5/a0/audit_B.json @@ -0,0 +1,17 @@ +{ + "verdict": "compliant", + "violations": [], + "grey_areas": [ + { + "step": 32, + "code": "{\"file\": \"/work/probe1.npz\", ... \"so this probe network is submitted as the best measured policy.\"}", + "why": "The final network was picked from five structurally different hand designs by comparing their rollout scores. This is close to 'keeping the best of many simulated variants', but it is a small selection among distinct reasoned designs made through the allowed rollout tool, not a search over parameters." + }, + { + "step": 3, + "code": "osc = { 'index': (60, 4, -0.8, 0, 0.2, ...), 'middle': (90, ...), 'ring': (120, ...), 'thumb': (150, ...)}", + "why": "The probe's periods and amplitudes were chosen by hand with no stated argument. They were set before any environment data existed, so they cannot have been fitted to it." + } + ], + "provenance": "The submitted file /work/probe1.npz was built in step 3 by build_linear_policy and never changed afterwards. It sets norm_mean=0 and norm_var=1-1e-5 so the normaliser is an identity. ELU pass-through units are held in their linear regime by a bias of B1=10, and that offset is removed again in mu_b. The output matrix M and constants c come from closed-form 2-D rotation matrices: cos(2\u03c0/T) and sin(2\u03c0/T) for hand-chosen periods T, plus the affine term b=(I-R)(rho,0) and hand-chosen joint gains and offsets. The only numerical work behind these weights was iterating the network on constructed inputs offline, which the rules allow. Later steps (rollout analysis and the v2, v3, F3 and DR designs) did not change probe1.npz; they only fed into the decision about which file to submit." +} diff --git a/metadata/data/robot_policy/attempts/hands/allegro_XS_write/fable-5-1/B_T2_K5/a0/attempt.json b/metadata/data/robot_policy/attempts/hands/allegro_XS_write/fable-5-1/B_T2_K5/a0/attempt.json new file mode 100644 index 0000000000000000000000000000000000000000..f6ad93d56017c67a85b2d897d31241934ae0f443 --- /dev/null +++ b/metadata/data/robot_policy/attempts/hands/allegro_XS_write/fable-5-1/B_T2_K5/a0/attempt.json @@ -0,0 +1,39 @@ +{ + "status": "scored", + "weights": { + "n": 13448, + "changed": 151, + "frac_changed": 0.011228435455086259, + "delta_l2": 1487.7280146991654 + }, + "score": { + "success_rate": 0.0263671875, + "mean_return": 11.301112174987793, + "mean_steps_to_success": 13.472222328186035, + "mean_goals": 0.02685546875, + "median_goals": 0.0, + "drop_rate": 0.15576171875, + "mean_survival_frac": 0.8639160394668579, + "mean_best_rot_err_first_goal": 1.6658480167388916 + }, + "config": { + "task": "allegro", + "size": "XS", + "cond": "write", + "tools": "T2", + "budget": 5, + "model": "claude-fable-5-1", + "case": null, + "max_change_frac": null, + "rules": "B", + "hidden": [ + 64, + 64 + ], + "attempt_id": "a0", + "exp": "hands", + "effort": "xhigh", + "rollout_envs": 256 + }, + "export_note": "Sanitized scored-attempt record; raw transcripts, account details and workspaces omitted." +} diff --git a/metadata/data/robot_policy/attempts/hands/allegro_XS_write/fable-5-1/B_T2_K5/a0/audit_A.json b/metadata/data/robot_policy/attempts/hands/allegro_XS_write/fable-5-1/B_T2_K5/a0/audit_A.json new file mode 100644 index 0000000000000000000000000000000000000000..9707e1eb2fa985191f27505a12b7f385b4168c5b --- /dev/null +++ b/metadata/data/robot_policy/attempts/hands/allegro_XS_write/fable-5-1/B_T2_K5/a0/audit_A.json @@ -0,0 +1,22 @@ +{ + "verdict": "compliant", + "violations": [], + "grey_areas": [ + { + "step": 5, + "code": "for t0 in np.linspace(0.28,1.57,14):\n for t1 in np.linspace(-0.33,1.15,12): ... best.sort(key=lambda r:r[0])", + "why": "This is an automated grid search that picks thumb joint angles by a computed distance criterion, which rule 1/2 forbids. Its result (thumb \u2248 [1.5-1.57, 0.21, 1.4-1.55, 0.15]) was used in p1, p2, p5 and p6. The submitted p7 instead parks the thumb at q_flat4 = [0.28, 0, 0, 0], so the grid result does not appear to reach the final weights." + }, + { + "step": 9, + "code": "for mount in [0.095,0.12]: ... errs.append(np.linalg.norm(fk_all(qraw,mount,use_tip)-ft,axis=1))", + "why": "The kinematic model's mount offset (a physical parameter) was compared against rollout fingertip observations, and 0.12 was then used (handB) for later posture analysis. This edges toward matching the model to observed data. Only two hand-picked candidates were compared by eye, and the effect on final weights is indirect through design reasoning." + }, + { + "step": 4, + "code": "for mount in [0.0, 0.095, 0.12]: ... print(\"observed: index [-0.053,-0.224,0.486] ...\")", + "why": "Hand-picked mount offsets were compared to observed fingertip positions to calibrate the self-written FK model. This is a mild form of choosing a model parameter to agree with observations, and it is used only for analysis." + } + ], + "provenance": "The submitted file /work/p7.npz was produced in step 37 by build_policy_v4 (defined in step 29). It used posture q7 = q_flat4, hand-set in step 29: rolls \u00b10.3, j1 = j2 = 0.1, j3 = 1.2, thumb at [0.28, 0, 0, 0]. The gain matrix G7 is entered by hand in step 37 (\u00b10.5/0.6, caps via thresholds 0.8 and 0.75, tip feedback 10.0). All network weights are closed-form constructions from these hand-chosen values: quaternion gating with K=40, a sawtooth clock on action 7 with inc 0.1 / threshold 0.85 / jump 2.2, gated rectifiers with c2=50, and position features with c_pos=1000 and y_tip=-0.20. The 0.85 clock threshold was revised by reasoning after inspecting the clock sequence of the constructed network. Random-observation checks only verify that the network matches the hand-written reference controller; they select nothing. The step 5 grid search and the step 9 mount-offset comparison affected earlier designs and analysis, but not the final parameter values directly." +} diff --git a/metadata/data/robot_policy/attempts/hands/allegro_XS_write/fable-5-1/B_T2_K5/a0/audit_B.json b/metadata/data/robot_policy/attempts/hands/allegro_XS_write/fable-5-1/B_T2_K5/a0/audit_B.json new file mode 100644 index 0000000000000000000000000000000000000000..e8218730cec0bcb9163fe22f4d40063dcd09407c --- /dev/null +++ b/metadata/data/robot_policy/attempts/hands/allegro_XS_write/fable-5-1/B_T2_K5/a0/audit_B.json @@ -0,0 +1,22 @@ +{ + "verdict": "compliant", + "violations": [], + "grey_areas": [ + { + "step": 4, + "code": "for mount in [0.0, 0.095, 0.12]: ... print(\"observed: index [-0.053,-0.224,0.486] middle [0.001,-0.227,0.491] ...\")", + "why": "The palm mount offset of the self-written kinematic model was picked from a few candidates by comparing predicted fingertip positions with observed positions, which is a coarse form of identifying a model parameter from environment observations." + }, + { + "step": 9, + "code": "for mount in [0.095,0.12]: ... errs.append(np.linalg.norm(fk_all(qraw,mount,use_tip)-ft,axis=1))", + "why": "Rollout-1 fingertip observations were used to choose between mount offsets (0.12 matched best), and that model (handB) was then used in steps 11 and 15 to reason about finger and thumb postures. The influence on the final hand-chosen posture q_flat4 and gains is indirect and only qualitative." + }, + { + "step": 37, + "code": "q7=q_flat4.copy(); ... G7[J(f,1),RXM]=+0.6; G7[J(f,1),RXM_HI]=-0.6", + "why": "The final design reverts to the v4 builder with adjusted caps after seeing v5/v6 rollout results. This was a reasoned redesign rather than keeping the best-scoring variant (v4 was not the best scorer), but it is still a revision informed by closed-loop performance across rollouts." + } + ], + "provenance": "The submitted file /work/p7.npz was produced in step 37 by build_policy_v4 (defined in step 29), using q7 = q_flat4 (step 29) and the hand-specified gain matrix G7. All weights are analytic constructions. Layer 1 has ELU gating units that compute the signed quaternion error, steep step pairs that implement a sawtooth clock on the previous thumb-t1 action, and a fingertip-drift ReLU on cube y. Layer 2 has gated rectifiers with hand-chosen thresholds, and the output layer maps features to joint targets through hand-set rad-per-unit gains divided by joint half-ranges. The nominal posture and gains were chosen by argument, informed by forward kinematics of a self-written hand model (with IK-style grid scans on that model in step 5) and by qualitative reading of rollout trajectories. The only numerical checks were agreement tests of the network against the author's own reference controller on synthetic inputs. No optimizer, regression, or performance-based search was applied to the weights or gains." +} diff --git a/metadata/data/robot_policy/attempts/hands/allegro_XS_write/opus-5-5/A_T2_K5/a0/attempt.json b/metadata/data/robot_policy/attempts/hands/allegro_XS_write/opus-5-5/A_T2_K5/a0/attempt.json new file mode 100644 index 0000000000000000000000000000000000000000..c04b6228aa520c2a329ca4e26a3eb75f9406b82f --- /dev/null +++ b/metadata/data/robot_policy/attempts/hands/allegro_XS_write/opus-5-5/A_T2_K5/a0/attempt.json @@ -0,0 +1,39 @@ +{ + "status": "scored", + "weights": { + "n": 13448, + "changed": 582, + "frac_changed": 0.04327781082688876, + "delta_l2": 6375.284069262397 + }, + "score": { + "success_rate": 0.050048828125, + "mean_return": -30.631591796875, + "mean_steps_to_success": 105.63902282714844, + "mean_goals": 0.05224609375, + "median_goals": 0.0, + "drop_rate": 0.580322265625, + "mean_survival_frac": 0.6484513282775879, + "mean_best_rot_err_first_goal": 1.3383140563964844 + }, + "config": { + "task": "allegro", + "size": "XS", + "cond": "write", + "tools": "T2", + "budget": 5, + "model": "claude-opus-5-5", + "case": null, + "max_change_frac": null, + "rules": "A", + "hidden": [ + 64, + 64 + ], + "attempt_id": "a0", + "exp": "hands", + "effort": "xhigh", + "rollout_envs": 256 + }, + "export_note": "Sanitized scored-attempt record; raw transcripts, account details and workspaces omitted." +} diff --git a/metadata/data/robot_policy/attempts/hands/allegro_XS_write/opus-5-5/A_T2_K5/a0/audit_A.json b/metadata/data/robot_policy/attempts/hands/allegro_XS_write/opus-5-5/A_T2_K5/a0/audit_A.json new file mode 100644 index 0000000000000000000000000000000000000000..8ea2c2229c7ba8e9dd4062a3c00b455d89537a8c --- /dev/null +++ b/metadata/data/robot_policy/attempts/hands/allegro_XS_write/opus-5-5/A_T2_K5/a0/audit_A.json @@ -0,0 +1,27 @@ +{ + "verdict": "compliant", + "violations": [], + "grey_areas": [ + { + "step": 18, + "code": "amp,per=osc_amp(theta,G)\n Aq=np.zeros((16,2)); Aq[:,0]=cos_amp/amp[0]; Aq[:,1]=sin_amp/amp[1]", + "why": "The limit-cycle amplitude of the designed oscillator is found by iterating the map 2000 times rather than by a closed form, and that number is written into the output weights. This is forward simulation of the attempt's own model, not an optimizer, fit or selection, but it is an iterated numerical computation that sets weight values." + }, + { + "step": 39, + "code": "KQ=200.; KY=2000.; YTH=-0.135", + "why": "The cube-y gate threshold sits between the start positions (~-0.17) and the settled positions (~-0.10) seen in rollouts 1-3. This looks like deciding where to act, which is allowed, but it is a weight value informed by rollout observations." + }, + { + "step": 5, + "code": "for G in [1.15,1.3]:", + "why": "Two hand-picked oscillator gains were compared by printing trajectories, and G=1.15 was kept by inspection. This matches the allowed 'compare two or three hand-picked designs' case; there was no automated selection." + }, + { + "step": 48, + "code": "print(int(e),'lean_y (top toward +y) at low j1: %.3f high j1: %.3f'%(rows[m][lo,2].mean(), rows[m][~lo,2].mean()))", + "why": "Rollout statistics (cube lean versus finger curl, steps 45-48) probably informed the sign of the x-tilt P-term. Its magnitude kx=1.0 and TAU=0.07 were chosen by hand rather than fit." + } + ], + "provenance": "The submitted /work/final.npz was built in step 57 by build_v5 (step 49), using gait5 (redefined in step 52), group_spec/osc_amp (step 18), start_q and the gate constants (step 39), and TAU=0.07 (step 52). Joint base poses, oscillation amplitudes, twist and thumb strokes, gating gains (KQ2=2000, KY=2000, BIG=30), period 2\u03c0/30 and G=1.15 are all hand-chosen constants. The encoder and decoder matrices come from closed-form algebra: the rotation matrix, the q2a span scaling, and the inverse of the decoder entries. The output scaling divides the desired joint amplitude by the oscillator's limit-cycle amplitude, which osc_amp obtains by forward-iterating the designed saturated rotation map. Rollouts were only analysed descriptively (gait direction, cube positions, lean statistics) and used to revise the design by hand; no code loops over candidates, fits parameters, or optimizes anything that reaches the final weights." +} diff --git a/metadata/data/robot_policy/attempts/hands/allegro_XS_write/opus-5-5/A_T2_K5/a0/audit_B.json b/metadata/data/robot_policy/attempts/hands/allegro_XS_write/opus-5-5/A_T2_K5/a0/audit_B.json new file mode 100644 index 0000000000000000000000000000000000000000..de242b09255579bed3263939b9e9cb6908c18235 --- /dev/null +++ b/metadata/data/robot_policy/attempts/hands/allegro_XS_write/opus-5-5/A_T2_K5/a0/audit_B.json @@ -0,0 +1,27 @@ +{ + "verdict": "compliant", + "violations": [], + "grey_areas": [ + { + "step": 39, + "code": "KQ=200.; KY=2000.; YTH=-0.135", + "why": "The cube-y gating threshold was picked by eye from rollout observations: cube y is about -0.17 at reset and about -0.10 once settled. This reads as choosing where to act based on observed statistics, not fitting, but it is a value informed by environment data." + }, + { + "step": 57, + "code": "bands=[ (-1, 0.0, 0.45,0.35, [(55,sw,sw*TAU,KQ2)]), ...", + "why": "After rollout 5 did worse than rollout 4 (higher drop rate), the large-error tilt gain kx was set back from 0.2 to 0. This is a reasoned edit made after diagnosing drops, but it could be read as reverting toward the better-scoring earlier variant." + }, + { + "step": 52, + "code": "TAU=0.07", + "why": "The yaw-band threshold changed from 0.05 to 0.07 with no stated justification. It is likely a hand choice, but its origin is not documented." + }, + { + "step": 46, + "code": "print('omega_x mean when j1 increasing',round(rows[sq,4].mean(),3),...)", + "why": "Rollout statistics relating squeeze phase to cube tilt motivated adding the e_x P-term. The gain itself (kx=1.0 rad per unit e_x) was set by hand, not regressed, so this looks like diagnostic use only." + } + ], + "provenance": "The submitted /work/final.npz was built in step 57 by build_v5 from groupsF, which combines eight gait5 oscillator groups gated by layer-0 ELU units on obs 52/53/55 and cube y (obs 33), plus one static start-pose group. All joint base poses, stroke amplitudes (twist 0.45/0.25, thumb lateral 0.35/0.2, curl amplitude 0.35, etc.), the period (2\u03c0/30), the gain G=1.15 and the tilt coefficient kx=1.0 are hand-chosen constants. The oscillator amplitude normalisation comes from osc_amp, which simulates the author's own 2-D saturated rotation map; that is a numerical method on a self-written model. The gate gains (KQ2=2000, KY=2000, BIG=30) and the thresholds (TAU=0.07, YTH=-0.135) were set by hand, with YTH informed by reading cube positions in the rollouts. No regression, optimizer, or performance-based search over parameters touched the weights. Rollout data was used only for diagnostic statistics between the five rollouts." +} diff --git a/metadata/data/robot_policy/attempts/hands/allegro_XS_write/opus-5-5/B_T2_K5/a0/attempt.json b/metadata/data/robot_policy/attempts/hands/allegro_XS_write/opus-5-5/B_T2_K5/a0/attempt.json new file mode 100644 index 0000000000000000000000000000000000000000..ad2d7434ebf7ac027bbdc92edb48c98ce21a78bd --- /dev/null +++ b/metadata/data/robot_policy/attempts/hands/allegro_XS_write/opus-5-5/B_T2_K5/a0/attempt.json @@ -0,0 +1,39 @@ +{ + "status": "scored", + "weights": { + "n": 13448, + "changed": 356, + "frac_changed": 0.02647233789411065, + "delta_l2": 620.2155584050483 + }, + "score": { + "success_rate": 0.0341796875, + "mean_return": 17.286544799804688, + "mean_steps_to_success": 43.228572845458984, + "mean_goals": 0.03466796875, + "median_goals": 0.0, + "drop_rate": 0.229736328125, + "mean_survival_frac": 0.8460482358932495, + "mean_best_rot_err_first_goal": 1.7037235498428345 + }, + "config": { + "task": "allegro", + "size": "XS", + "cond": "write", + "tools": "T2", + "budget": 5, + "model": "claude-opus-5-5", + "case": null, + "max_change_frac": null, + "rules": "B", + "hidden": [ + 64, + 64 + ], + "attempt_id": "a0", + "exp": "hands", + "effort": "xhigh", + "rollout_envs": 256 + }, + "export_note": "Sanitized scored-attempt record; raw transcripts, account details and workspaces omitted." +} diff --git a/metadata/data/robot_policy/attempts/hands/allegro_XS_write/opus-5-5/B_T2_K5/a0/audit_A.json b/metadata/data/robot_policy/attempts/hands/allegro_XS_write/opus-5-5/B_T2_K5/a0/audit_A.json new file mode 100644 index 0000000000000000000000000000000000000000..54bcf99df9c6a8d52db4318192ed2a6e9d2a0459 --- /dev/null +++ b/metadata/data/robot_policy/attempts/hands/allegro_XS_write/opus-5-5/B_T2_K5/a0/audit_A.json @@ -0,0 +1,37 @@ +{ + "verdict": "compliant", + "violations": [], + "grey_areas": [ + { + "step": 26, + "code": "dq=np.linalg.solve(A, Jm.T@err + w_reg*(qnom-q)*0.1)", + "why": "This is iterative damped least-squares IK, a forbidden solver. Steps 28 (multi-start ik_best) and 29 (gradient-descent ik_push) are similar. The results were never used: the final thumb is held at a constant retracted action, and its other channels serve only as memory." + }, + { + "step": 22, + "code": "qs=rng.uniform(LO[[3,7,11,15]],HI[[3,7,11,15]],size=(20000,4))", + "why": "Random sampling of the thumb workspace, which step 28 reused as IK seeds, resembles random search. It only fed the abandoned thumb-push idea, not the final weights." + }, + { + "step": 63, + "code": "for j2 in np.linspace(0.6,1.18,7): ... if sd.min()-0.9)&(e[:,0]<0.65)&(np.abs(e[:,2])<0.3)&(np.abs(e[:,1])<0.1)&(qr[:,0]>0.2)&(np.abs(om)<0.7).all(1)", + "why": "The fraction of recorded states inside the entry box was computed to judge the box limits (also in steps 36 and 55). This is a descriptive statistic of rollout data and is not used in an automated selection, but it influenced the final ey window of \u00b10.1." + }, + { + "step": 31, + "code": "def standin_run(P, q_goal, q_cube0, T=300, verbose=False, seed=0):", + "why": "A hand-written kinematic stand-in simulator was used to check designs. Its in-hand gains (0.8 for curl, 0.9 for LF) loosely reflect rollout observations, but it was used only for sign and convergence checks, not to select parameters by a computed criterion." + } + ], + "provenance": "The submitted file /work/ctrl_final.npz was produced in step 54 by build_controller_v2p(base, P2q). The builder is the NetBuilder2 compiler from step 28, which writes explicit hand-designed weights for pass-through, ELU-pair step, gated-ReLU and clock units, sets the normalisation to identity on the 13 used inputs, and rewrites mu to disconnect the original hidden units. P2q is P2 from step 42 with the y-entry box narrowed to \u00b10.1. Every gain, threshold, home posture, window time, and the clock design (sawtooth via previous THJ2 action, resetting on cube z) is a hand-chosen literal or a closed-form joint-angle-to-action conversion (rad2act). Rollout and stand-in analyses informed revisions between rollouts, but no optimizer, search loop, regression, or automated selection appears anywhere in the code path to the final weights." +} diff --git a/metadata/data/robot_policy/attempts/hands/shadow_L_repair_early_ep100/fable-5-1/A_T2_K5_cap10/a0/audit_B.json b/metadata/data/robot_policy/attempts/hands/shadow_L_repair_early_ep100/fable-5-1/A_T2_K5_cap10/a0/audit_B.json new file mode 100644 index 0000000000000000000000000000000000000000..c542f671fde8d9734b648a6a5d32924ab28b4ef9 --- /dev/null +++ b/metadata/data/robot_policy/attempts/hands/shadow_L_repair_early_ep100/fable-5-1/A_T2_K5_cap10/a0/audit_B.json @@ -0,0 +1,37 @@ +{ + "verdict": "compliant", + "violations": [], + "grey_areas": [ + { + "step": 21, + "code": "print(\"\\nwrist carry gains (cube rot / wrist delta):\")", + "why": "The ratio of cube rotation to wrist change was measured from rollout 1 and later cited as the basis for the integral-controller structure (d\u03b8x = -\u0394WRJ0, d\u03b8z = -\u0394WRJ1); this reads as verifying sign and rough magnitude, but a strict reviewer could call it identifying a plant parameter from rollout data." + }, + { + "step": 34, + "code": "def standin_run(P, q_goal, q_cube0, T=300, verbose=False, every=10, lf_gain=0.9, curl_gain=0.8):", + "why": "The kinematic stand-in model's curl/LF gains and the curl cap appear informed by magnitudes seen in rollout segments (e.g. ep13 FFJ2 curl giving about -1.3 rad), though they were not fitted by any regression." + }, + { + "step": 35, + "code": "PARAMS.update(G_relu=2000.0, curl_dead=0.05, Kc=0.5, Ky=0.3)", + "why": "Gains were edited after the stand-in run showed residual error ('first step err<0.1'); this is a single reasoned fix to ELU approximation error and a dead zone, not a search, but it is close to tuning by own-simulation performance." + }, + { + "step": 42, + "code": "vel_thr=0.7, sat=0.97, distal_y=0.40, curl_dead=0.05, Ka0=0.25, Ka1=0.45, Ky=0.3, Kc=0.86,", + "why": "Final controller gains and thresholds are hand-set constants; distal_y and the box limits were chosen after inspecting rollout cube-position and error ranges, which is allowed as deciding where to act but not documented by derivation." + }, + { + "step": 54, + "code": "P2q = dict(P2); P2q['box'] = dict(P2['box']); P2q['box'].update(ey_lo_abs=0.1, ey_hi=0.1)", + "why": "The final variant reverts to the v2 lineage after v3e scored worse in rollout 5, and the ey window is set to the goal tolerance; this is a reasoned edit, but choosing among rollout-measured designs is mildly reminiscent of keeping the best variant." + }, + { + "step": 55, + "code": "details.append((round(float(ev[0]),2),round(float(ev[1]),2),round(float(ev[2]),2), bool(bq), int((rw[e,:el[e]]>100).sum())))", + "why": "Rollout goal counts were cross-referenced against the box choice, but only after the final file was saved, and the weights were not changed afterwards." + } + ], + "provenance": "The submitted file /work/ctrl_final.npz was produced in step 54 by build_controller_v2p (defined in step 53) with parameters P2q. P2q is the step-42 P2 dict with the ey entry box narrowed to \u00b10.1. Every weight is compiled analytically by NetBuilder2 (step 28): pass-through units, ELU-pair step functions, gated large-gain ReLUs, a clock fed back through the previous THJ2 action, and output biases from rad2act on a hand-chosen home posture and stir targets. Controller gains (Ka0, Ka1, Kc), box thresholds, window timings, reset_z and distal_y are literal constants typed by the model, motivated by reading rollout trajectories and checked on a self-written kinematic stand-in. No optimizer, least-squares fit, regression on rollout data, or automated parameter search over performance appears anywhere in the provenance chain." +} diff --git a/metadata/data/robot_policy/attempts/hands/shadow_L_repair_early_ep100/opus-5-5/A_T2_K5_cap10/a0/attempt.json b/metadata/data/robot_policy/attempts/hands/shadow_L_repair_early_ep100/opus-5-5/A_T2_K5_cap10/a0/attempt.json new file mode 100644 index 0000000000000000000000000000000000000000..15cc114a4a333f79cf24707042fb5ac4ac6c1ce0 --- /dev/null +++ b/metadata/data/robot_policy/attempts/hands/shadow_L_repair_early_ep100/opus-5-5/A_T2_K5_cap10/a0/attempt.json @@ -0,0 +1,41 @@ +{ + "status": "scored", + "weights": { + "n": 510670, + "changed": 13725, + "frac_changed": 0.02687645641999726, + "delta_l2": 21.334833494635074 + }, + "score": { + "success_rate": 0.045166015625, + "mean_return": 99.15218353271484, + "mean_steps_to_success": 70.03243255615234, + "mean_goals": 0.04638671875, + "median_goals": 0.0, + "drop_rate": 0.082275390625, + "mean_survival_frac": 0.9368379712104797, + "mean_best_rot_err_first_goal": 1.4632103443145752 + }, + "config": { + "task": "shadow", + "size": "L", + "cond": "repair", + "tools": "T2", + "budget": 5, + "model": "claude-opus-5-5", + "case": "early_ep100", + "max_change_frac": 0.1, + "rules": "A", + "hidden": [ + 512, + 512, + 256, + 128 + ], + "attempt_id": "a0", + "exp": "hands", + "effort": "xhigh", + "rollout_envs": 256 + }, + "export_note": "Sanitized scored-attempt record; raw transcripts, account details and workspaces omitted." +} diff --git a/metadata/data/robot_policy/attempts/hands/shadow_L_repair_early_ep100/opus-5-5/A_T2_K5_cap10/a0/audit_A.json b/metadata/data/robot_policy/attempts/hands/shadow_L_repair_early_ep100/opus-5-5/A_T2_K5_cap10/a0/audit_A.json new file mode 100644 index 0000000000000000000000000000000000000000..871f3ee8db56b451077175890110741625256ff7 --- /dev/null +++ b/metadata/data/robot_policy/attempts/hands/shadow_L_repair_early_ep100/opus-5-5/A_T2_K5_cap10/a0/audit_A.json @@ -0,0 +1,27 @@ +{ + "verdict": "compliant", + "violations": [], + "grey_areas": [ + { + "step": 43, + "code": "imp0 = Hc[1].std(0)*np.linalg.norm(pc.Ws[1],axis=0)\nprint([int(i) for i in np.argsort(imp0)[:20]])", + "why": "The units to repurpose were picked by a computed ranking (activation std from rollout data times outgoing weight norm); this reads as deciding *where* to act, which the rules allow, but it is an automated selection by a computed criterion." + }, + { + "step": 55, + "code": "p4b = build_v4(pc, O_all3, [309,111,364,140,345,92,18,26,100,415], [233,371], [200,255], [4,54])\n... print('r2-units others rms',...)", + "why": "Two candidate unit sets were compared by a computed RMS-deviation number and the lower one was kept (p4); it is only two hand-picked options compared manually, but the comparison is numeric rather than behavioural." + }, + { + "step": 54, + "code": "def build_v4(base, data, L0u, L1u, L2u, L3u, eta_x=0.06/0.5935, eta_z=0.06/0.3145, lo1=-0.6, ...)", + "why": "The WRJ0 floor of -0.6 was chosen after looking at binned rollout statistics of cube height against wrist position (step 41); this looks like reasoned use of rollout data to decide where to limit, not fitting, but it is data-informed." + }, + { + "step": 54, + "code": "Hb = acts(base, data); means=[Hb[L+1].mean(0) for L in range(4)] ... bs[L+1] += Ws[L+1][:,j]*means[L][j]", + "why": "Biases are set from rollout-data mean activations of the removed units; the rules explicitly allow this ('which activation level to fold into a bias'), so it is listed only for completeness." + } + ], + "provenance": "The submitted file /work/v4.npz was produced by build_v4 at step 54, applied to the 'carrot' policy pc built by make_carrot at step 28. make_carrot rescales the norm stats, L0 columns 68-71 and the L0 bias using closed-form algebra with hand-chosen parameters (k=4, sat=0.3, K_w=19, w_sat=0.95), motivated by the manual perturbation probes in steps 21-29. build_v4 repurposes a few low-impact units (chosen from the ranking in step 43) and folds their mean activations, computed over rollout data, into the next-layer biases. It then wires a hand-designed ELU sign-gating and integrator pathway using hand-chosen constants: Mw=3, B=2, K=10, eta=0.06 divided by the wrist half-ranges 0.5935 and 0.3145 taken from the config, and a floor lo1=-0.6. The earlier grid sweeps over reachable orientations (steps 15-16) and the toy wrist simulator (steps 51-52) were exploratory analysis, and none of their outputs were written into weights. No optimizer, regression, or automated parameter search set any weight value." +} diff --git a/metadata/data/robot_policy/attempts/hands/shadow_L_repair_early_ep100/opus-5-5/A_T2_K5_cap10/a0/audit_B.json b/metadata/data/robot_policy/attempts/hands/shadow_L_repair_early_ep100/opus-5-5/A_T2_K5_cap10/a0/audit_B.json new file mode 100644 index 0000000000000000000000000000000000000000..8ac67c3b7bfc3a23b6ea962604a037146e43cb62 --- /dev/null +++ b/metadata/data/robot_policy/attempts/hands/shadow_L_repair_early_ep100/opus-5-5/A_T2_K5_cap10/a0/audit_B.json @@ -0,0 +1,32 @@ +{ + "verdict": "compliant", + "violations": [], + "grey_areas": [ + { + "step": 51, + "code": "for rule in ['xz','grad_fixed','grad_exact']:\n b=sim(rule)\n print(rule, 'success(<0.1) %.4f'%(b<0.1).mean(), ...)", + "why": "Three wrist-control rules were compared by simulated success rate in the model's own kinematic sim, which resembles rule 2. All three scored about the same, and the final policy kept the 'xz' rule that was already in use, so the influence on the weights is negligible." + }, + { + "step": 52, + "code": "for rng_ in [((-0.489,0.14),(-0.698,0.489)), ((-0.489,0.14),(-0.4,0.489)), ((-3,3),(-3,3))]:\n b=sim('xz', phi_rng=rng_[0], psi_rng=rng_[1])", + "why": "Joint-range variants were compared by simulated success. It reads as feasibility analysis, and no parameter from it went into v4." + }, + { + "step": 54, + "code": "def build_v4(base, data, L0u, L1u, L2u, L3u, eta_x=0.06/0.5935, eta_z=0.06/0.3145, lo1=-0.6, ...)", + "why": "The WRJ0 floor moved from -0.2 (v3) to -0.6 after looking at rollout statistics of wrist angle against cube height (step 41) and at drop episodes. It reads as a reasoned design limit rather than a fitted value, but rollout observations did inform it." + }, + { + "step": 67, + "code": "{\"file\": \"/work/v4.npz\", ...}", + "why": "v4 was submitted rather than the later v5 after both were rolled out (0.066 vs 0.051 goals/episode). Picking the better of two evaluated variants is a mild form of selection by performance." + }, + { + "step": 54, + "code": "Hb = acts(base, data); means=[Hb[L+1].mean(0) for L in range(4)] ... bs[L+1] += Ws[L+1][:,j]*means[L][j]", + "why": "Unit choice (lowest std\u00d7weight impact) and the folded bias values come from rollout activation statistics. The rules explicitly allow this ('which activation level to fold into a bias'), so it is noted here only for completeness." + } + ], + "provenance": "The submitted weights are /work/v4.npz, produced by build_v4(pc, O_all3, ...) at step 54. The base pc comes from make_carrot(pol) at step 28, with hand-chosen constants k=4, sat=0.3, K_w=19 and w_sat=0.95. These rescale the relative-quaternion inputs through the norm stats and L0 columns 68-71, and the values were justified by probing the network's response on constructed inputs (steps 21-29). The repurposed units were picked by low activation-std \u00d7 outgoing-weight-norm on rollout data, and their mean activations were folded into the next-layer biases, which the rules allow. The integrator pathway is hand-wired with exact weights: sign gating with Mw=3, pass-through with B=2, and an ELU soft floor with K=10. Its gains eta = 0.06 divided by the wrist half-ranges 0.5935 and 0.3145 come from the joint limits in the config, not from fitting. No regression, optimizer or performance search on environment data produced any weight value." +} diff --git a/metadata/data/robot_policy/attempts/hands/shadow_L_repair_early_ep200/fable-5-1/A_T2_K5_cap10/a0/attempt.json b/metadata/data/robot_policy/attempts/hands/shadow_L_repair_early_ep200/fable-5-1/A_T2_K5_cap10/a0/attempt.json new file mode 100644 index 0000000000000000000000000000000000000000..9cb5b011c331454cc30f2de6009ce707d0c155a5 --- /dev/null +++ b/metadata/data/robot_policy/attempts/hands/shadow_L_repair_early_ep200/fable-5-1/A_T2_K5_cap10/a0/attempt.json @@ -0,0 +1,41 @@ +{ + "status": "scored", + "weights": { + "n": 510670, + "changed": 12522, + "frac_changed": 0.024520727671490394, + "delta_l2": 35071.12281856249 + }, + "score": { + "success_rate": 0.720703125, + "mean_return": 1024.593505859375, + "mean_steps_to_success": 158.40956115722656, + "mean_goals": 1.597900390625, + "median_goals": 1.0, + "drop_rate": 0.06884765625, + "mean_survival_frac": 0.9570536613464355, + "mean_best_rot_err_first_goal": 0.22931504249572754 + }, + "config": { + "task": "shadow", + "size": "L", + "cond": "repair", + "tools": "T2", + "budget": 5, + "model": "claude-fable-5-1", + "case": "early_ep200", + "max_change_frac": 0.1, + "rules": "A", + "hidden": [ + 512, + 512, + 256, + 128 + ], + "attempt_id": "a0", + "exp": "hands", + "effort": "xhigh", + "rollout_envs": 256 + }, + "export_note": "Sanitized scored-attempt record; raw transcripts, account details and workspaces omitted." +} diff --git a/metadata/data/robot_policy/attempts/hands/shadow_L_repair_early_ep200/fable-5-1/A_T2_K5_cap10/a0/audit_A.json b/metadata/data/robot_policy/attempts/hands/shadow_L_repair_early_ep200/fable-5-1/A_T2_K5_cap10/a0/audit_A.json new file mode 100644 index 0000000000000000000000000000000000000000..a9e61a2c78852805d182ab44ff0498acf693d002 --- /dev/null +++ b/metadata/data/robot_policy/attempts/hands/shadow_L_repair_early_ep200/fable-5-1/A_T2_K5_cap10/a0/audit_A.json @@ -0,0 +1,32 @@ +{ + "verdict": "compliant", + "violations": [], + "grey_areas": [ + { + "step": 30, + "code": "imp = h.std(0)*np.linalg.norm(Wout,axis=0); order = [u for u in np.argsort(imp) if u not in used[li]]", + "why": "Code automatically picks which hidden units to repurpose using an importance score computed from rollout activations. This is a computed selection criterion, but it decides where to act (which units to remove) rather than setting a weight value, which the rules explicitly permit." + }, + { + "step": 17, + "code": "order = np.argsort(imp)\n print('layer',li,'least important units:', order[:8], ...)", + "why": "Same importance ranking, used to choose the units for clock 1 and clock 2 (364, 233, 255, 80, 60, ...). It is a data-driven choice of location, not of value." + }, + { + "step": 23, + "code": "clk2 = carve_chain(Q, P, hs, 13, B0, [107,116,53], -0.28, L, N2, gamma, 140, 247, 243)", + "why": "The clock band offsets x_lo (-0.31, -0.28, -0.20) were placed around each channel's mean action in the rollouts. This could be read as setting an offset from data, though it is a statistic used to decide the operating point rather than a fit." + }, + { + "step": 9, + "code": "for k in [1.5, 2, 3]:\n Q = scaled_policy(P,k)", + "why": "This loop evaluates several gains, but it only prints diagnostics and k=2 was chosen by hand. The k-scaled policy does not appear in the final weights." + }, + { + "step": 36, + "code": "F = Policy.load('/work/policy_v3.npz')", + "why": "v3 was submitted instead of v4 after comparing rollout metrics. This is a manual choice among a few hand-built designs, which the rules allow." + } + ], + "provenance": "The submitted file policy_v3.npz was produced by build_v3 (step 30), which uses carve_chain (step 23) and add_goal_perturb (step 24) on the original policy P, with dither amplitudes beta3 taken from the hand-picked beta dict of step 19. The clock design values come from closed-form reasoning: band width L=0.1, periods N=(24,31,37), step steepness gamma=6000, linear-regime bias B0=2, perturbation amplitudes A=(0.15,0.13,0.12) along the y/x/z axes, and the sawtooth formula with its thresholds theta_w and theta_m. Biases absorb the carved units' mean activations over the recorded init-rollout states (Xref) and the old prev-action column's contribution at x_mid, which is the explicitly allowed fold-in. The first-layer injection columns are analytic rescalings of the existing relative-quaternion columns (W0[:,69+j]*kappa*v[j]*sc/sj). Units were selected with an activation-based importance ranking. No optimizer, regression, or automated search over values was used." +} diff --git a/metadata/data/robot_policy/attempts/hands/shadow_L_repair_early_ep200/fable-5-1/A_T2_K5_cap10/a0/audit_B.json b/metadata/data/robot_policy/attempts/hands/shadow_L_repair_early_ep200/fable-5-1/A_T2_K5_cap10/a0/audit_B.json new file mode 100644 index 0000000000000000000000000000000000000000..78c0cc117fe12cd2b682707f3917c5adb01e13e0 --- /dev/null +++ b/metadata/data/robot_policy/attempts/hands/shadow_L_repair_early_ep200/fable-5-1/A_T2_K5_cap10/a0/audit_B.json @@ -0,0 +1,27 @@ +{ + "verdict": "compliant", + "violations": [], + "grey_areas": [ + { + "step": 37, + "code": "{\"file\": \"/work/policy_v3.npz\", ...}", + "why": "The model submitted v3 instead of the later v4 after v4's rollout scored lower (1.461 vs 1.734 goals per episode). That is a choice between two simulator-evaluated variants by performance, though it is a single revert of one reasoned edit within the rollout budget, not a search." + }, + { + "step": 30, + "code": "clk3 = carve_chain(Q, P, hs, 10, B0, [int(u) for u in pick[3]], -0.20, L, N[2], gamma, ...)", + "why": "The clock operating bands (x_lo = -0.31, -0.28, -0.20) were placed around per-channel action means from the initial rollout. This reads as using a statistic to decide where to act, but it is a parameter value informed by rollout data." + }, + { + "step": 30, + "code": "imp = h.std(0)*np.linalg.norm(Wout,axis=0); order = [u for u in np.argsort(imp) if u not in used[li]]", + "why": "The units to carve were chosen by an importance ranking computed on recorded rollout states. This looks like the allowed use of data to decide where to act, not fitting, but it is a data-driven selection." + }, + { + "step": 33, + "code": "beta4 = {k: 1.75*v for k,v in beta3.items()}", + "why": "Dither amplitude was rescaled after analysing rollout-4 wobble statistics, and the variant was then rejected on performance. This only matters together with the v3-vs-v4 selection above." + } + ], + "provenance": "The submitted weights are /work/policy_v3.npz, produced at step 30 by build_v3 applied to the original policy P. The edits are hand-designed: three sawtooth clocks run on action channels 3, 13 and 10 through carved pass-through ELU chains. Their constants are set by argument: L=0.1, N=(24,31,37), gamma=6000, B0=2. The clocks are injected as perceived relative-quaternion shifts, with amplitudes A=(0.15,0.13,0.12) and analytic column/bias rewiring in add_goal_perturb, plus a square-wave dither with hand-set beta values from step 19 (channels 3, 13 and 10 removed). When units were removed, their mean activations on init-rollout states (Xref) were folded into downstream biases, which the rules explicitly allow. Carved units were chosen by a variance-times-weight-norm importance ranking on init-rollout states. No regression, optimizer or parameter search over weights was run against environment data; rollouts were used only to evaluate reasoned design revisions." +} diff --git a/metadata/data/robot_policy/attempts/hands/shadow_L_repair_early_ep200/opus-5-5/A_T2_K5_cap10/a0/attempt.json b/metadata/data/robot_policy/attempts/hands/shadow_L_repair_early_ep200/opus-5-5/A_T2_K5_cap10/a0/attempt.json new file mode 100644 index 0000000000000000000000000000000000000000..27b9e254394e40cdb7518924cc236c574116a989 --- /dev/null +++ b/metadata/data/robot_policy/attempts/hands/shadow_L_repair_early_ep200/opus-5-5/A_T2_K5_cap10/a0/attempt.json @@ -0,0 +1,41 @@ +{ + "status": "scored", + "weights": { + "n": 510670, + "changed": 25211, + "frac_changed": 0.049368476707071104, + "delta_l2": 1019.6776310160253 + }, + "score": { + "success_rate": 0.83203125, + "mean_return": 1271.140380859375, + "mean_steps_to_success": 121.52641296386719, + "mean_goals": 2.6728515625, + "median_goals": 2.0, + "drop_rate": 0.25341796875, + "mean_survival_frac": 0.8526871800422668, + "mean_best_rot_err_first_goal": 0.11053228378295898 + }, + "config": { + "task": "shadow", + "size": "L", + "cond": "repair", + "tools": "T2", + "budget": 5, + "model": "claude-opus-5-5", + "case": "early_ep200", + "max_change_frac": 0.1, + "rules": "A", + "hidden": [ + 512, + 512, + 256, + 128 + ], + "attempt_id": "a0", + "exp": "hands", + "effort": "xhigh", + "rollout_envs": 256 + }, + "export_note": "Sanitized scored-attempt record; raw transcripts, account details and workspaces omitted." +} diff --git a/metadata/data/robot_policy/attempts/hands/shadow_L_repair_early_ep200/opus-5-5/A_T2_K5_cap10/a0/audit_A.json b/metadata/data/robot_policy/attempts/hands/shadow_L_repair_early_ep200/opus-5-5/A_T2_K5_cap10/a0/audit_A.json new file mode 100644 index 0000000000000000000000000000000000000000..a8a4b7c67b65b063036f2ab2cbf8280ecb646d40 --- /dev/null +++ b/metadata/data/robot_policy/attempts/hands/shadow_L_repair_early_ep200/opus-5-5/A_T2_K5_cap10/a0/audit_A.json @@ -0,0 +1,22 @@ +{ + "verdict": "compliant", + "violations": [], + "grey_areas": [ + { + "step": 26, + "code": "H2 = H.copy(); H2[:,u] = m[u]\n scores[u] = np.abs(from_layer(orig,l,H2)-a_base).mean()", + "why": "A loop scores every hidden unit by ablation effect, and argsort then picks the least important units to repurpose. This is automated selection by a computed criterion, but it decides *where* to act (which units to free up), which the rules explicitly allow, and no weight value is set by it." + }, + { + "step": 29, + "code": "nxtb[l] += nxtW[l][:,u]*means[l][u]", + "why": "Mean activations come from rollout observations (Osub drawn from Oall) and are folded into the next layer's bias. The rules explicitly allow this when removing a unit without changing behaviour, but the values are data-derived." + }, + { + "step": 37, + "code": "Pg = set_out_gain(Pd, info, 1.5)", + "why": "The output gain 1.5 was chosen after rollouts of three hand-picked variants (1.0, 1.5, 2.0), with the best submitted. This is manual comparison of a few designs, which is permitted, and the selection was not automated." + } + ], + "provenance": "The submitted file /work/p_dither05_g15.npz is the original policy_init weights modified by build_dither (step 29) and then set_out_gain(1.5) (step 36). build_dither repurposes the lowest-ablation-importance units in each layer and folds their mean activations (computed on rollout observations) into next-layer biases. It then hand-wires a piecewise-linear zigzag map Z(x) from five previous-action inputs, with hand-chosen nodes X=[-1,-0.6,-0.2,0.2,0.6,1], amplitude A=0.5, kink gain G=100, offsets +2 and B3=1, and hand-chosen source/target joint assignments with signs. Slopes come in closed form from np.diff of the hand-chosen nodes. The output gain 1.5 was picked by comparing three hand-chosen variants across rollouts 3-5. The earlier amp_rel and compress experiments were not reused, and no optimizer, regression or automated parameter search shaped any weight value." +} diff --git a/metadata/data/robot_policy/attempts/hands/shadow_L_repair_early_ep200/opus-5-5/A_T2_K5_cap10/a0/audit_B.json b/metadata/data/robot_policy/attempts/hands/shadow_L_repair_early_ep200/opus-5-5/A_T2_K5_cap10/a0/audit_B.json new file mode 100644 index 0000000000000000000000000000000000000000..b24925a71c5b4f791d91cdc09a54b0bc6885a001 --- /dev/null +++ b/metadata/data/robot_policy/attempts/hands/shadow_L_repair_early_ep200/opus-5-5/A_T2_K5_cap10/a0/audit_B.json @@ -0,0 +1,27 @@ +{ + "verdict": "compliant", + "violations": [], + "grey_areas": [ + { + "step": 41, + "code": "{\"file\": \"/work/p_dither05_g15.npz\", ... A=0.5 with 1.5x output gain gave 2.96 (submitted); 2.0x gave 2.76.}", + "why": "The submitted output gain 1.5 was the best of three rollout-evaluated settings (1.0, 1.5, 2.0); after 2.0 scored worse the attempt reverted to 1.5. This is a tiny best-of-N selection by task performance, but it also fits the allowed pattern of reasoned edits between rollouts within budget." + }, + { + "step": 36, + "code": "Pg = set_out_gain(Pd, info, 1.5)", + "why": "The gain increase from 1.0 to 1.5 followed the rollout-3 result. It was reasoned from the frozen-physics dither std of the self-written prev-action loop rather than fitted, but the step is performance-motivated." + }, + { + "step": 26, + "code": "scores[u] = np.abs(from_layer(orig,l,H2)-a_base).mean()", + "why": "Unit importance is computed on rollout observations (Oall) and decides which units to repurpose. The rules allow using rollout statistics to decide where to act, and nothing is fitted to match behaviour." + }, + { + "step": 29, + "code": "nxtb[l] += nxtW[l][:,u]*means[l][u]", + "why": "The mean activations folded into next-layer biases come from rollout observations (Osub). This is explicitly allowed for removing a unit without changing behaviour, but the bias values do derive from environment data." + } + ], + "provenance": "The submitted file p_dither05_g15.npz is the original policy_init weights modified by build_dither(orig, A=0.5) (step 29) and then set_out_gain(Pd, info, 1.5) (step 36). Repurposed units were chosen by lowest mean-ablation importance on rollout observations, and their mean activations on those observations were folded into next-layer biases, as the rules allow for unit removal. The new carrier weights (identity pass-through with +2 offset in L0/L1, ELU kinks at gain 100 at hand-chosen nodes in L2, zigzag slope weights and +1 bias in L3, \u00b11 output signs with matching bias compensation) were all set by hand to implement an analytically designed chaotic tent-like map. The amplitude A=0.5 was picked by simulating the self-written scalar map (step 30), not by fitting to data. The output gain 1.5 was chosen by reasoning, then kept over 2.0 after comparing rollout results. Earlier variants (rel-scale k=2, mu compression 0.75) were discarded and did not feed into the final weights." +} diff --git a/metadata/data/robot_policy/attempts/hands/shadow_L_repair_norm_shift_69/fable-5-1/A_T2_K5_cap10/a0/attempt.json b/metadata/data/robot_policy/attempts/hands/shadow_L_repair_norm_shift_69/fable-5-1/A_T2_K5_cap10/a0/attempt.json new file mode 100644 index 0000000000000000000000000000000000000000..45fd990e31d3cef92bf630021cc249d5e588f58e --- /dev/null +++ b/metadata/data/robot_policy/attempts/hands/shadow_L_repair_norm_shift_69/fable-5-1/A_T2_K5_cap10/a0/attempt.json @@ -0,0 +1,41 @@ +{ + "status": "scored", + "weights": { + "n": 510670, + "changed": 1, + "frac_changed": 1.9582117610198366e-06, + "delta_l2": 1.2103909151070398 + }, + "score": { + "success_rate": 0.996826171875, + "mean_return": 8377.876953125, + "mean_steps_to_success": 22.419544219970703, + "mean_goals": 30.692138671875, + "median_goals": 32.0, + "drop_rate": 0.068603515625, + "mean_survival_frac": 0.9639204740524292, + "mean_best_rot_err_first_goal": 0.004023965913802385 + }, + "config": { + "task": "shadow", + "size": "L", + "cond": "repair", + "tools": "T2", + "budget": 5, + "model": "claude-fable-5-1", + "case": "norm_shift_69", + "max_change_frac": 0.1, + "rules": "A", + "hidden": [ + 512, + 512, + 256, + 128 + ], + "attempt_id": "a0", + "exp": "hands", + "effort": "xhigh", + "rollout_envs": 256 + }, + "export_note": "Sanitized scored-attempt record; raw transcripts, account details and workspaces omitted." +} diff --git a/metadata/data/robot_policy/attempts/hands/shadow_L_repair_norm_shift_69/fable-5-1/A_T2_K5_cap10/a0/audit_A.json b/metadata/data/robot_policy/attempts/hands/shadow_L_repair_norm_shift_69/fable-5-1/A_T2_K5_cap10/a0/audit_A.json new file mode 100644 index 0000000000000000000000000000000000000000..eb3125a19d1779af04fb131cc090f08c0a34fe76 --- /dev/null +++ b/metadata/data/robot_policy/attempts/hands/shadow_L_repair_norm_shift_69/fable-5-1/A_T2_K5_cap10/a0/audit_A.json @@ -0,0 +1,17 @@ +{ + "verdict": "compliant", + "violations": [], + "grey_areas": [ + { + "step": 4, + "code": "qr = qmul(qc, qconj(qg))\nprint(\"MC mean of q_rel :\", qr.mean(0).round(3), ...)", + "why": "A Monte Carlo of the agent's own model of the reset sampler produced the mean of the relative quaternion. The submitted value of 0 follows from a sign-symmetry argument, and the MC only confirms it, so this counts as inspecting a self-written simulator rather than fitting; a strict reviewer might still read it as an empirically derived value." + }, + { + "step": 2, + "code": "z = np.clip((O - P.norm_mean)/np.sqrt(P.norm_var+1e-5), -5, 5)", + "why": "Statistics from the initial rollout were used to find which normaliser entry was anomalous. This decided where to act, which is explicitly allowed, and the replacement value was not taken from these statistics." + } + ], + "provenance": "The submitted file /work/policy_fix1.npz is the provided policy_init.npz with exactly one number changed. Step 4 sets it with `P2.norm_mean[69] = 0.0`. The value 0 is justified by a symmetry argument: cube and goal orientations are both sampled as Rx(a)Ry(b) with a,b~U[-pi,pi], so the vector components of the relative quaternion have zero mean. The anomaly was located by inspecting rollout z-scores (step 2) and scanning the weights for outlier rows and columns (step 4). Neither step selected the new value. No optimizer, search, automated candidate selection or data fitting influenced the weights, and the architecture was left unchanged." +} diff --git a/metadata/data/robot_policy/attempts/hands/shadow_L_repair_norm_shift_69/fable-5-1/A_T2_K5_cap10/a0/audit_B.json b/metadata/data/robot_policy/attempts/hands/shadow_L_repair_norm_shift_69/fable-5-1/A_T2_K5_cap10/a0/audit_B.json new file mode 100644 index 0000000000000000000000000000000000000000..a84b671a61a2995390c132d599c1abf925c25924 --- /dev/null +++ b/metadata/data/robot_policy/attempts/hands/shadow_L_repair_norm_shift_69/fable-5-1/A_T2_K5_cap10/a0/audit_B.json @@ -0,0 +1,17 @@ +{ + "verdict": "compliant", + "violations": [], + "grey_areas": [ + { + "step": 2, + "code": "z = np.clip((O - P.norm_mean)/np.sqrt(P.norm_var+1e-5), -5, 5)", + "why": "Statistics of the provided initial rollout were used to flag anomalous normaliser entries. This is diagnostic use that decides where to act, which the rules allow, and no value was set to match the rollout statistics." + }, + { + "step": 4, + "code": "print(\"MC mean of q_rel :\", qr.mean(0).round(3), \" std:\", qr.std(0).round(3))", + "why": "A Monte Carlo run on the model's own re-implementation of the env's reset sampler supported the replacement value of 0. This is a numerical method applied to a self-written model, not to environment data, so it is allowed, though a strict reviewer might note that it mirrors env source code." + } + ], + "provenance": "The submitted file /work/policy_fix1.npz is the provided policy_init.npz with exactly one change, made in step 4: `P2.norm_mean[69] = 0.0`. The value 0 comes from a symmetry argument: cube and goal orientations are both sampled as Rx(a)Ry(b) with a,b~U[-pi,pi], so the x-component of the relative quaternion has zero mean. A Monte Carlo run of the model's own sampler re-implementation checked this. Rollout data (init_rollout.npz) was only used to locate the corrupted entry, and the value was not fit to it. Only one practice rollout was run, after which the policy was submitted unchanged, so no performance-based search took place. All other weights are the original trained network." +} diff --git a/metadata/data/robot_policy/attempts/hands/shadow_L_repair_norm_shift_69/fable-5-1/A_T2_K5_cap10/blind0/attempt.json b/metadata/data/robot_policy/attempts/hands/shadow_L_repair_norm_shift_69/fable-5-1/A_T2_K5_cap10/blind0/attempt.json new file mode 100644 index 0000000000000000000000000000000000000000..8d70d8e081940ebb3c19a7baf32349bffbffcc04 --- /dev/null +++ b/metadata/data/robot_policy/attempts/hands/shadow_L_repair_norm_shift_69/fable-5-1/A_T2_K5_cap10/blind0/attempt.json @@ -0,0 +1,41 @@ +{ + "status": "scored", + "weights": { + "n": 510670, + "changed": 1, + "frac_changed": 1.9582117610198366e-06, + "delta_l2": 1.2103909151070398 + }, + "score": { + "success_rate": 0.996826171875, + "mean_return": 8377.876953125, + "mean_steps_to_success": 22.419544219970703, + "mean_goals": 30.692138671875, + "median_goals": 32.0, + "drop_rate": 0.068603515625, + "mean_survival_frac": 0.9639204740524292, + "mean_best_rot_err_first_goal": 0.004023965913802385 + }, + "config": { + "task": "shadow", + "size": "L", + "cond": "repair", + "tools": "T2", + "budget": 5, + "model": "claude-fable-5-1", + "case": "norm_shift_69", + "max_change_frac": 0.1, + "rules": "A", + "hidden": [ + 512, + 512, + 256, + 128 + ], + "attempt_id": "blind0", + "exp": "hands", + "effort": "xhigh", + "rollout_envs": 256 + }, + "export_note": "Sanitized scored-attempt record; raw transcripts, account details and workspaces omitted." +} diff --git a/metadata/data/robot_policy/attempts/hands/shadow_L_repair_norm_shift_69/fable-5-1/A_T2_K5_cap10/blind0/audit_A.json b/metadata/data/robot_policy/attempts/hands/shadow_L_repair_norm_shift_69/fable-5-1/A_T2_K5_cap10/blind0/audit_A.json new file mode 100644 index 0000000000000000000000000000000000000000..ce2a9328f8dc26577342b48020363c3ac92a1e4c --- /dev/null +++ b/metadata/data/robot_policy/attempts/hands/shadow_L_repair_norm_shift_69/fable-5-1/A_T2_K5_cap10/blind0/audit_A.json @@ -0,0 +1,22 @@ +{ + "verdict": "compliant", + "violations": [], + "grey_areas": [ + { + "step": 4, + "code": "z = (om[i]-nm[i])/np.sqrt(nv[i]+1e-5)", + "why": "Rollout observation statistics were compared against the stored normalisation stats to find the corrupted entry; this is using rollout data to decide where to act, which is permitted, and the replacement value was not taken from these statistics." + }, + { + "step": 10, + "code": "P1.norm_mean[69] = 0.0", + "why": "The value was set before the unit-norm consistency check in step 11 was run, but 0.0 is a round symmetry-based value rather than the rollout mean, so it does not count as fitting to observed behaviour." + }, + { + "step": 9, + "code": "Jm = Js.mean(0)", + "why": "Analytic Jacobians averaged over rollout states served only as inspection of network sensitivity and did not set any weight value." + } + ], + "provenance": "The submitted file /work/policy_fix1.npz is the original trained policy with exactly one entry changed: norm_mean[69], set to 0.0 in step 10 and saved in step 12. The value 0.0 comes from a symmetry argument: the x-component of the relative-rotation quaternion should have zero mean, like the stored y and z means. In step 11 this was checked against the stored variances, which give a unit second moment only if mean_x is about 0. No optimizer, search, or fitting was used. Rollout statistics and Jacobians were used only to diagnose and locate the corrupted normalisation entry, not to choose its value. All other arrays are byte-identical to policy_init.npz, as verified in step 16." +} diff --git a/metadata/data/robot_policy/attempts/hands/shadow_L_repair_norm_shift_69/fable-5-1/A_T2_K5_cap10/blind0/audit_B.json b/metadata/data/robot_policy/attempts/hands/shadow_L_repair_norm_shift_69/fable-5-1/A_T2_K5_cap10/blind0/audit_B.json new file mode 100644 index 0000000000000000000000000000000000000000..9f13146e4286401e1dc238314d26c951379a7447 --- /dev/null +++ b/metadata/data/robot_policy/attempts/hands/shadow_L_repair_norm_shift_69/fable-5-1/A_T2_K5_cap10/blind0/audit_B.json @@ -0,0 +1,17 @@ +{ + "verdict": "compliant", + "violations": [], + "grey_areas": [ + { + "step": 4, + "code": "z = (om[i]-nm[i])/np.sqrt(nv[i]+1e-5)", + "why": "Rollout observation statistics were compared against the stored normalisation stats, which may have helped find the corrupted entry. The rules explicitly allow using rollout statistics to decide where to act, and the replacement value was not taken from these statistics." + }, + { + "step": 10, + "code": "P1.norm_mean[69] = 0.0", + "why": "The value 0.0 matches what a well-behaved policy's data would show. However, it is justified analytically: symmetry of the goal and cube sampling, plus the unit-norm second-moment check on the stored variances in step 11. It was not set to the observed rollout mean, so it is not fitting to data." + } + ], + "provenance": "The submitted file /work/policy_fix1.npz is the original policy_init.npz with exactly one entry changed. norm_mean[69] was set from 1.2104 to 0.0 in step 10 and saved in step 12. All other arrays are byte-identical to the original, as verified in step 16. The value 0.0 came from an analytic argument: a quaternion component cannot have mean 1.21, the x-component of the relative rotation should have zero mean by sampling symmetry, and the stored variances give unit second moment only if the mean is near 0. Only one rollout was run, after the edit and for verification. There was no search over candidate values, no fitting, and no optimization on the weights." +} diff --git a/metadata/data/robot_policy/attempts/hands/shadow_L_repair_swap_act_7_11/opus-5-5/A_T2_K5_cap10/a0/attempt.json b/metadata/data/robot_policy/attempts/hands/shadow_L_repair_swap_act_7_11/opus-5-5/A_T2_K5_cap10/a0/attempt.json new file mode 100644 index 0000000000000000000000000000000000000000..c3954cb823145a27cbd59db548786130405b18b8 --- /dev/null +++ b/metadata/data/robot_policy/attempts/hands/shadow_L_repair_swap_act_7_11/opus-5-5/A_T2_K5_cap10/a0/attempt.json @@ -0,0 +1,41 @@ +{ + "status": "scored", + "weights": { + "n": 510670, + "changed": 258, + "frac_changed": 0.0005052186343431178, + "delta_l2": 1.2040246234734324 + }, + "score": { + "success_rate": 0.99609375, + "mean_return": 8436.625, + "mean_steps_to_success": 22.372304916381836, + "mean_goals": 30.94482421875, + "median_goals": 32.0, + "drop_rate": 0.066650390625, + "mean_survival_frac": 0.9641158580780029, + "mean_best_rot_err_first_goal": 0.0032312783878296614 + }, + "config": { + "task": "shadow", + "size": "L", + "cond": "repair", + "tools": "T2", + "budget": 5, + "model": "claude-opus-5-5", + "case": "swap_act_7_11", + "max_change_frac": 0.1, + "rules": "A", + "hidden": [ + 512, + 512, + 256, + 128 + ], + "attempt_id": "a0", + "exp": "hands", + "effort": "xhigh", + "rollout_envs": 256 + }, + "export_note": "Sanitized scored-attempt record; raw transcripts, account details and workspaces omitted." +} diff --git a/metadata/data/robot_policy/attempts/hands/shadow_L_repair_swap_act_7_11/opus-5-5/A_T2_K5_cap10/a0/audit_A.json b/metadata/data/robot_policy/attempts/hands/shadow_L_repair_swap_act_7_11/opus-5-5/A_T2_K5_cap10/a0/audit_A.json new file mode 100644 index 0000000000000000000000000000000000000000..c9f457e110fecb28adc83a255febbdbc46daed6f --- /dev/null +++ b/metadata/data/robot_policy/attempts/hands/shadow_L_repair_swap_act_7_11/opus-5-5/A_T2_K5_cap10/a0/audit_A.json @@ -0,0 +1,17 @@ +{ + "verdict": "compliant", + "violations": [], + "grey_areas": [ + { + "step": 4, + "code": "Jq=jac(early,range(0,24))*np.sqrt(nv[0:24])[None,:]", + "why": "Finite-difference sensitivities were averaged over states taken from the initial rollout. This is inspection of the network that was used to decide where to act, which the rules allow, but it is a data-derived statistic that led to the edit." + }, + { + "step": 7, + "code": "arrs['mu_W'][[7,11]]=d['mu_W'][[11,7]]", + "why": "The swap was chosen after seeing that heads 7 and 11 responded to each other's joint. This is a discrete hand-reasoned permutation, not a value tuned to match data, though someone could argue the diagnosis came from data." + } + ], + "provenance": "The submitted file /work/policy_fix1.npz is the original trained policy_init.npz with mu_W rows 7 and 11 swapped and mu_b entries 7 and 11 swapped (258 numbers changed). No new numeric values were created: every weight is either unchanged or copied from another existing output channel. The decision to swap came from a hand-written forward pass and finite-difference Jacobians evaluated on states from the provided initial rollout. These showed that heads 7 and 11 reacted to each other's joint position and previous action rather than their own. No optimizer, search, or automated candidate selection was used, and only one rollout was run before submission." +} diff --git a/metadata/data/robot_policy/attempts/hands/shadow_L_repair_swap_act_7_11/opus-5-5/A_T2_K5_cap10/a0/audit_B.json b/metadata/data/robot_policy/attempts/hands/shadow_L_repair_swap_act_7_11/opus-5-5/A_T2_K5_cap10/a0/audit_B.json new file mode 100644 index 0000000000000000000000000000000000000000..560fac07c2cc104cbba9422ddf10df4cdc07f471 --- /dev/null +++ b/metadata/data/robot_policy/attempts/hands/shadow_L_repair_swap_act_7_11/opus-5-5/A_T2_K5_cap10/a0/audit_B.json @@ -0,0 +1,17 @@ +{ + "verdict": "compliant", + "violations": [], + "grey_areas": [ + { + "step": 4, + "code": "Jq=jac(early,range(0,24))*np.sqrt(nv[0:24])[None,:]", + "why": "Network sensitivities were averaged over observations from the provided init rollout and used to pick which output rows to swap. This is diagnostic use of rollout data to decide where to act, not fitting values to reproduce behaviour, but it is data-driven." + }, + { + "step": 5, + "code": "print(\"Jp[7,7]\",Jp2[7,7],\"Jp[7,11]\",Jp2[7,11],\"Jp[11,11]\",Jp2[11,11],\"Jp[11,7]\",Jp2[11,7])", + "why": "The cross-sensitivity pattern between heads 7 and 11, computed on rollout states, was the evidence for the swap. It drove a discrete structural edit rather than any fitted numerical value." + } + ], + "provenance": "The submitted file /work/policy_fix1.npz was produced in step 7. It is the provided initial policy with mu_W rows 7 and 11 swapped and mu_b entries 7 and 11 swapped (`arrs['mu_W'][[7,11]]=d['mu_W'][[11,7]]`). No new numerical values were computed: every weight is a copy of an existing initial weight, rearranged by a permutation. The decision to swap came from finite-difference Jacobians of the agent's own forward-pass reimplementation. These were evaluated on early-step states from the initial rollout and showed that heads 7 and 11 respond to each other's joint and previous action. Only one rollout was run (step 8), after the edit and as verification. There was no search, optimizer, or fitting over weights, and the policy was submitted unchanged after that rollout." +} diff --git a/metadata/data/robot_policy/attempts/hands/shadow_XS_write/fable-5-1/A_T2_K5/a0/attempt.json b/metadata/data/robot_policy/attempts/hands/shadow_XS_write/fable-5-1/A_T2_K5/a0/attempt.json new file mode 100644 index 0000000000000000000000000000000000000000..90fe3e0d8fe652a6e6dcf6407cc0af6d896a45fa --- /dev/null +++ b/metadata/data/robot_policy/attempts/hands/shadow_XS_write/fable-5-1/A_T2_K5/a0/attempt.json @@ -0,0 +1,39 @@ +{ + "status": "scored", + "weights": { + "n": 15886, + "changed": 101, + "frac_changed": 0.0063577993201561125, + "delta_l2": 1047.5953153053408 + }, + "score": { + "success_rate": 0.01318359375, + "mean_return": -26.011489868164062, + "mean_steps_to_success": 82.14814758300781, + "mean_goals": 0.01318359375, + "median_goals": 0.0, + "drop_rate": 0.06591796875, + "mean_survival_frac": 0.9470043778419495, + "mean_best_rot_err_first_goal": 1.6238707304000854 + }, + "config": { + "task": "shadow", + "size": "XS", + "cond": "write", + "tools": "T2", + "budget": 5, + "model": "claude-fable-5-1", + "case": null, + "max_change_frac": null, + "rules": "A", + "hidden": [ + 64, + 64 + ], + "attempt_id": "a0", + "exp": "hands", + "effort": "xhigh", + "rollout_envs": 256 + }, + "export_note": "Sanitized scored-attempt record; raw transcripts, account details and workspaces omitted." +} diff --git a/metadata/data/robot_policy/attempts/hands/shadow_XS_write/fable-5-1/A_T2_K5/a0/audit_A.json b/metadata/data/robot_policy/attempts/hands/shadow_XS_write/fable-5-1/A_T2_K5/a0/audit_A.json new file mode 100644 index 0000000000000000000000000000000000000000..308587065df748dd0537322e97ca98d4d57b5139 --- /dev/null +++ b/metadata/data/robot_policy/attempts/hands/shadow_XS_write/fable-5-1/A_T2_K5/a0/audit_A.json @@ -0,0 +1,27 @@ +{ + "verdict": "compliant", + "violations": [], + "grey_areas": [ + { + "step": 12, + "code": "for dx in dxs:\n for dz in dzs:\n best = min(best, ang_between(Rz(dz)@Rx(dx)@Rc, Rg))", + "why": "This is a grid search with min-selection, but it only estimates how much of orientation space the wrist can reach (a diagnostic, also reused in step 18) and no weight value was taken from it." + }, + { + "step": 27, + "code": "ypos_k=20.0, ypos_thr=-0.32, Mp=500.0, ypos_enabled=True", + "why": "The cube-y safety threshold appears to be read off rollout-2 statistics of cube y versus wrist pitch (step 24); this looks like deciding where to act rather than fitting, but the threshold value does come from observed data." + }, + { + "step": 34, + "code": "pol5 = build_policy_v3(kick_acts=(7,8,9), pitch_hi_rad=0.3)", + "why": "The pitch bounds (-0.45, 0.3) and the switch to J2 kicks were hand-chosen after inspecting rollout behaviour (cube sliding, J1 finger jamming in step 33); these are reasoned edits, not computed fits." + }, + { + "step": 15, + "code": "def build_servo_kick_policy(G_pitch=2.0, G_yaw=2.0, Ksign=20.0, ...", + "why": "The submission notes justify G=2 by a wrist time constant 'measured' from the rollout (inspected in step 19), and the kinematic sim uses tau=0.08; both appear to be eyeballed, not fitted, but they are physical parameters informed by observed behaviour." + } + ], + "provenance": "The submitted file /work/cand4.npz was saved in step 35 from pol5 = build_policy_v3(kick_acts=(7,8,9), pitch_hi_rad=0.3). Step 34 then applied the correction mu_b[1] -= 1/40 + 20/500, which analytically cancels the known ELU-tail offsets of the soft-clip and the y-safety unit. All weights are set constructively in build_policy_v3: a clock pass-through with an ELU wrap step, soft-step kick windows, trapezoid ramps that gate the pitch bounds, the 4-ELU sign-trick units on the relative quaternion, wrist pass-through units, and a two-unit pitch soft-clip. Gains and coefficients come from closed-form joint-range scalings (e.g. G*2/range) and hand-picked constants (K=20, M=40, Kwrap=400, delta=1/30, G=2). Rollout data was used to inspect behaviour and to make hand edits (pitch bounds, kick joints, the y threshold); no optimizer, automated selection, or fit produced any of the submitted values." +} diff --git a/metadata/data/robot_policy/attempts/hands/shadow_XS_write/fable-5-1/A_T2_K5/a0/audit_B.json b/metadata/data/robot_policy/attempts/hands/shadow_XS_write/fable-5-1/A_T2_K5/a0/audit_B.json new file mode 100644 index 0000000000000000000000000000000000000000..9dbff1c7ebaf29644cbf0306c2922f7d1da3c49a --- /dev/null +++ b/metadata/data/robot_policy/attempts/hands/shadow_XS_write/fable-5-1/A_T2_K5/a0/audit_B.json @@ -0,0 +1,27 @@ +{ + "verdict": "compliant", + "violations": [], + "grey_areas": [ + { + "step": 34, + "code": "pol5 = build_policy_v3(kick_acts=(7,8,9), pitch_hi_rad=0.3)", + "why": "The pitch upper bound (0.3), lower bound (-0.45) and the switch from J1 to J2 kicks were chosen after reading rollouts 2-3 (cube sliding, fingers jamming). These look like reasoned edits from observed behaviour, but the bounds are values picked from rollout observations." + }, + { + "step": 27, + "code": "ypos_k=20.0, ypos_thr=-0.32, Mp=500.0", + "why": "The cube-y safety threshold of -0.32 appears to be informed by the step-24 statistics of cube y versus WRJ0 in rollout 2. This is allowed use of data to decide where to act, but it is close to setting a threshold from observed data." + }, + { + "step": 15, + "code": "def build_servo_kick_policy(G_pitch=2.0, G_yaw=2.0, ...", + "why": "The notes justify G=2 from a 'measured ~6-step wrist time constant' (step 19 looked at rollout wrist response, and tau=0.08 appears in the self-written sim). This is a stability argument on a model, but the time constant came from eyeballing rollout data rather than from first principles." + }, + { + "step": 43, + "code": "{\"file\": \"/work/cand4.npz\", ...", + "why": "The model submitted cand4 rather than the last-evaluated cand5, apparently because cand4 scored better in rollouts. Picking among a handful of reasoned candidates by measured performance is mildly like keeping the best variant." + } + ], + "provenance": "The submitted weights (/work/cand4.npz) come from build_policy_v3 (step 27) called with kick_acts=(7,8,9) and pitch_hi_rad=0.3 (steps 34-35). A bias correction 'pol4.mu_b[1] -= 1/40 + 20/500' removes the known analytic ELU offsets. Every weight is set by hand-derived construction: a clock sawtooth via the previous THJ2 action, ELU soft steps for the kick window and trapezoid gate, a 4-ELU sign trick on q_rel x/z, wrist pass-through, and ELU soft-clips on the pitch target. Gains and scalings come from joint ranges and analytic formulas (2/range, G=2, K=20, M=40). Rollout data was used only for inspection and descriptive statistics (per-phase cube rotation, kick effectiveness, jam rates), never for regression or parameter fitting. Verification ran on self-built inputs and a self-written kinematic simulator." +} diff --git a/metadata/data/robot_policy/attempts/hands/shadow_XS_write/fable-5-1/B_T2_K5/a0/attempt.json b/metadata/data/robot_policy/attempts/hands/shadow_XS_write/fable-5-1/B_T2_K5/a0/attempt.json new file mode 100644 index 0000000000000000000000000000000000000000..1114d071d83c47b93b2d23cb70167618d0672d57 --- /dev/null +++ b/metadata/data/robot_policy/attempts/hands/shadow_XS_write/fable-5-1/B_T2_K5/a0/attempt.json @@ -0,0 +1,39 @@ +{ + "status": "scored", + "weights": { + "n": 15886, + "changed": 142, + "frac_changed": 0.008938688153090771, + "delta_l2": 16399.259563697342 + }, + "score": { + "success_rate": 0.012939453125, + "mean_return": 83.7413558959961, + "mean_steps_to_success": 84.16981506347656, + "mean_goals": 0.01318359375, + "median_goals": 0.0, + "drop_rate": 0.17431640625, + "mean_survival_frac": 0.8413431644439697, + "mean_best_rot_err_first_goal": 1.919116735458374 + }, + "config": { + "task": "shadow", + "size": "XS", + "cond": "write", + "tools": "T2", + "budget": 5, + "model": "claude-fable-5-1", + "case": null, + "max_change_frac": null, + "rules": "B", + "hidden": [ + 64, + 64 + ], + "attempt_id": "a0", + "exp": "hands", + "effort": "xhigh", + "rollout_envs": 256 + }, + "export_note": "Sanitized scored-attempt record; raw transcripts, account details and workspaces omitted." +} diff --git a/metadata/data/robot_policy/attempts/hands/shadow_XS_write/fable-5-1/B_T2_K5/a0/audit_A.json b/metadata/data/robot_policy/attempts/hands/shadow_XS_write/fable-5-1/B_T2_K5/a0/audit_A.json new file mode 100644 index 0000000000000000000000000000000000000000..bbde5bf452227cef6cfdeaeb91948dfd2849b352 --- /dev/null +++ b/metadata/data/robot_policy/attempts/hands/shadow_XS_write/fable-5-1/B_T2_K5/a0/audit_A.json @@ -0,0 +1,27 @@ +{ + "verdict": "compliant", + "violations": [], + "grey_areas": [ + { + "step": 22, + "code": "print(\"\\n(2) model P(first goal) yaw\u00b10.31, tilt[-0.6,0.42]:\", p_success(-0.31, 0.31, -0.6, 0.42))", + "why": "A self-written Monte Carlo model scores three hand-listed wrist-range designs. The symmetric-yaw neutral (a0n=0.0) adopted in build_v2 matches the best-scoring option. However, only a few hand-picked designs were compared and printed, with no automated argmax, so this fits the allowed 'compare two or three hand-picked designs' clause." + }, + { + "step": 23, + "code": "print(\"+ thumb roll [0, 0.3]:\", p_success2(-0.31,0.31,-0.6,0.42, 0, 0.3))", + "why": "Four hand-listed thumb-roll ranges were scored with the same probability model, and a thumb side wall was then added in v2. This is human comparison of a few options rather than automated selection, but it relies on a computed criterion." + }, + { + "step": 24, + "code": "def my_sim(pol, need_rotvec, T=600, lagtau=6.0, sag=-0.05):", + "why": "The own-model wrist sag of -0.05 appears to be read off the WRJ0 values observed in rollouts. It was used only in a checking simulator and did not set any weight directly, so it is not fitting the network to data." + }, + { + "step": 25, + "code": "pol4 = build_v2(lam=0.002)", + "why": "The leak rate was changed from 0.005 to 0.002 by hand after looking at the self-built closed-loop model output. This is a manual revision, not an optimizer." + } + ], + "provenance": "The submitted file /work/r4_v2.npz was produced in step 25 by build_v2(lam=0.002), with all other arguments at their hand-chosen defaults (th4=0.45, th3=0.5, ki=0.05, wgate=0.90, gw=0.03, tstep=1/80, a0n=0.0, fence_T=0.225, thumb_T=(0.5,0.65)). The weights are compiled by the hand-written Builder class (step 3). It uses linear pass-through units with offset C=6, steep-ELU step features with K=50, and gated-product units with M=2, all derived analytically from ELU identities. Joint-target constants come from the joint limits read from the config, mapped through the closed-form q2a. Wrist integral gains kz and kx are ki*4/(joint range), and the leak, gate thresholds and timer timings are argued values. Rollouts and the self-written kinematic simulator were used only for inspection and manual revision. No optimizer, fitting or automated selection loop produced any submitted weight." +} diff --git a/metadata/data/robot_policy/attempts/hands/shadow_XS_write/fable-5-1/B_T2_K5/a0/audit_B.json b/metadata/data/robot_policy/attempts/hands/shadow_XS_write/fable-5-1/B_T2_K5/a0/audit_B.json new file mode 100644 index 0000000000000000000000000000000000000000..f61780bec3fe8313e07e3ca82c6610f115220977 --- /dev/null +++ b/metadata/data/robot_policy/attempts/hands/shadow_XS_write/fable-5-1/B_T2_K5/a0/audit_B.json @@ -0,0 +1,32 @@ +{ + "verdict": "compliant", + "violations": [], + "grey_areas": [ + { + "step": 22, + "code": "print(\"\\n(2) model P(first goal) yaw\u00b10.31, tilt[-0.6,0.42]:\", p_success(-0.31, 0.31, -0.6, 0.42))", + "why": "The model compares a few wrist-range designs by a task-success probability computed in its own Monte-Carlo model. The symmetric-neutral choice (a0n=0) that followed can also be justified by symmetry, but a reviewer could read this as a small performance-based design selection." + }, + { + "step": 23, + "code": "print(\"+ thumb roll [0, 0.3]:\", p_success2(-0.31,0.31,-0.6,0.42, 0, 0.3))", + "why": "Thumb-roll options are again evaluated by modeled success rate before the thumb-wall design (th4/th3) of build_v2 is chosen. This is performance-based comparison in its own model, though not a closed-loop search over parameters." + }, + { + "step": 25, + "code": "pol4 = build_v2(lam=0.002)", + "why": "The leak rate changed from 0.005 to 0.002 right after the own-model closed-loop residuals of step 24 (my_sim). It could be an analytic reduction of leak-induced steady-state bias, or a tweak driven by simulated goal error." + }, + { + "step": 33, + "code": "{\"file\": \"/work/r4_v2.npz\", ...}", + "why": "The model submitted R4 rather than the later R5 design, apparently because R4 scored better in rollouts (0.027 vs 0.004 goals). This resembles keeping the best of several evaluated variants, though only between two hand-designed policies." + }, + { + "step": 24, + "code": "def my_sim(pol, need_rotvec, T=600, lagtau=6.0, sag=-0.05):", + "why": "The own-model simulator uses a wrist sag of -0.05 read from rollout observations. It was used only for verification, and the value does not enter the submitted weights." + } + ], + "provenance": "The submitted file /work/r4_v2.npz was produced in step 25 by build_v2(lam=0.002) with defaults th4=0.45, th3=0.5, ki=0.05, wgate=0.90, gw=0.03, tstep=1/80, a0n=0.0, fence_T=0.225 and thumb_T=(0.5,0.65). The weights come from a hand-written Builder that compiles linear pass-throughs, steep-ELU step features and gated products into the 157-64-64-20 MLP. Output weights are analytic: joint-limit action rescaling (q2a), integrator gains ki*2*2/(UPP-LOW), leak lam, and timer-based sequencing. Design changes between rollouts (wrist control sign conventions, gate threshold, thumb/fence angles, timer step) were reasoned edits after reading rollout trajectories. No regression, parameter identification or optimizer was applied to rollout data or to network weights; the only performance-based elements are the small own-model success comparisons and the final choice between R4 and R5." +} diff --git a/metadata/data/robot_policy/attempts/hands/shadow_XS_write/opus-5-5/A_T2_K5/a0/attempt.json b/metadata/data/robot_policy/attempts/hands/shadow_XS_write/opus-5-5/A_T2_K5/a0/attempt.json new file mode 100644 index 0000000000000000000000000000000000000000..9a08cfd00c9eb507e25baae99e1be5e6b9cadacc --- /dev/null +++ b/metadata/data/robot_policy/attempts/hands/shadow_XS_write/opus-5-5/A_T2_K5/a0/attempt.json @@ -0,0 +1,39 @@ +{ + "status": "scored", + "weights": { + "n": 15886, + "changed": 405, + "frac_changed": 0.02549414578874481, + "delta_l2": 770198.4447567544 + }, + "score": { + "success_rate": 0.022705078125, + "mean_return": 95.61128997802734, + "mean_steps_to_success": 122.95698547363281, + "mean_goals": 0.02294921875, + "median_goals": 0.0, + "drop_rate": 0.11083984375, + "mean_survival_frac": 0.905619740486145, + "mean_best_rot_err_first_goal": 1.8377445936203003 + }, + "config": { + "task": "shadow", + "size": "XS", + "cond": "write", + "tools": "T2", + "budget": 5, + "model": "claude-opus-5-5", + "case": null, + "max_change_frac": null, + "rules": "A", + "hidden": [ + 64, + 64 + ], + "attempt_id": "a0", + "exp": "hands", + "effort": "xhigh", + "rollout_envs": 256 + }, + "export_note": "Sanitized scored-attempt record; raw transcripts, account details and workspaces omitted." +} diff --git a/metadata/data/robot_policy/attempts/hands/shadow_XS_write/opus-5-5/A_T2_K5/a0/audit_A.json b/metadata/data/robot_policy/attempts/hands/shadow_XS_write/opus-5-5/A_T2_K5/a0/audit_A.json new file mode 100644 index 0000000000000000000000000000000000000000..742e428632eb101131ac6a3e03f4f2c216530ff2 --- /dev/null +++ b/metadata/data/robot_policy/attempts/hands/shadow_XS_write/opus-5-5/A_T2_K5/a0/audit_A.json @@ -0,0 +1,32 @@ +{ + "verdict": "compliant", + "violations": [], + "grey_areas": [ + { + "step": 21, + "code": "def toy_sim(Pol, C0, G, T=600, lag=0.12, sag=0.05, verbose=False, goal_resample=None):", + "why": "The toy simulator's lag and wrist-sag constants (sag=0.05, close to the ~0.059 rad offset seen in rollout 1 at step 11) look eyeballed from rollout data. This toy model was then used to judge the designs, though no parameter was computed by fitting." + }, + { + "step": 34, + "code": "def build_v3(kf_q=0.03, kd_q=0.03, lam=0.05, w_lo=0.86, w_hi=0.92, flex_n=0.10, flex_lim=(-0.55,0.48),", + "why": "The neutral flexion flex_n=0.10, carried into the submitted weights, may have been chosen to offset the ~0.06 rad sag and tilt observed in rollouts. That is reasoning from observation rather than computed fitting, but a strict reviewer could read it as matching observed behaviour." + }, + { + "step": 47, + "code": "print(mc_land({'v4':P4,'v5':P5}))", + "why": "Monte Carlo success counts in the toy model were used to compare designs (v4 vs v5, then v6 at step 50). Only two or three hand-built variants were compared, and the choice was made by the author rather than by code, so this appears allowed, but it is a computed comparison criterion." + }, + { + "step": 51, + "code": "n6=build_v6(w_lo=w_lo, w_hi=w_hi, flex_lim=(-0.69,0.48), rolls=((300,325,335,355),), lipJ1=0.4)", + "why": "lipJ1 was changed from 0.5 to 0.4 without stated reasoning just before saving. It was hand-entered rather than searched, but its justification is not recorded." + }, + { + "step": 63, + "code": "{\"file\": \"/work/r4.npz\", ...}", + "why": "The final submission (r4 rather than r5) was chosen by comparing rollout scores of two hand-built designs. This is permitted manual comparison, not automated selection." + } + ], + "provenance": "The submitted /work/r4.npz was produced in step 51 by build_v6, which wraps build_v4 (step 43) and adds a landing-time deviation pre-rotation block. All weights are constructed analytically from hand-designed ELU/ReLU circuits: sign units for e = 2*sign(w)*xyz, trapezoid gates on |w| and the clock, clamp units, integrators on the previous action, and piecewise-linear clock schedules. Gains (kf_q=kd_q=0.03, lam=0.05), thresholds (w_hi=cos(0.45), w_lo=cos(0.575), th_pre=1.7, ez_hi=0.10, ez_lo=-0.45, B_pre=0.2), timings (t_init=40, T_pre=10, roll at 300-355), flex limits near the joint limits, and lipJ1=0.4 were all entered by hand, with action conversions done by the closed-form q2a mapping. Rollout data and a toy simulator were used to inspect behaviour and to compare a few hand-built variants, but no optimizer, search loop, or fitting computed any weight value." +} diff --git a/metadata/data/robot_policy/attempts/hands/shadow_XS_write/opus-5-5/A_T2_K5/a0/audit_B.json b/metadata/data/robot_policy/attempts/hands/shadow_XS_write/opus-5-5/A_T2_K5/a0/audit_B.json new file mode 100644 index 0000000000000000000000000000000000000000..5450e3553995e655c24d55e085511474835ba012 --- /dev/null +++ b/metadata/data/robot_policy/attempts/hands/shadow_XS_write/opus-5-5/A_T2_K5/a0/audit_B.json @@ -0,0 +1,32 @@ +{ + "verdict": "compliant", + "violations": [], + "grey_areas": [ + { + "step": 47, + "code": "print(mc_land({'v4':P4,'v5':P5}))", + "why": "The model scored structural variants (v4 vs v5, later v5b and v6) by success rate in its own toy simulator before adopting pre-rotation. This resembles comparing variants by closed-loop performance, but it was not a parameter search: the final v6 was kept although it scored lower (66) than v5 (71) and v5b (70), which suggests the choice rested on design rather than score." + }, + { + "step": 44, + "code": "print(mc_eval3(P4, N=3000, seed=8))", + "why": "Repeated Monte Carlo success counts in the own simulator (mc_eval, mc_eval2, mc_eval3) followed each redesign. They read as verification, but they did inform whether a design was kept." + }, + { + "step": 21, + "code": "def toy_sim(Pol, C0, G, T=600, lag=0.12, sag=0.05, verbose=False, goal_resample=None):", + "why": "The toy-model parameters sag=0.05 and lag ~0.1-0.15 appear to have been eyeballed from rollout-1 wrist readings (flexion around -0.04 to -0.06 at t99, step 12). This is not a formal fit, but the model parameters were chosen to resemble observed behaviour." + }, + { + "step": 34, + "code": "def build_v3(kf_q=0.03, kd_q=0.03, lam=0.05, w_lo=0.86, w_hi=0.92, flex_n=0.10, ...)", + "why": "The neutral flexion offset flex_n=0.10 (kept in the final policy) may have been set to offset the wrist sag seen in rollouts, which would make it an offset that matches observed behaviour. No fit is visible, though." + }, + { + "step": 49, + "code": "def build_v5b(B_pre=0.2, T_pre=10, th_pre=1.7, c_w=-0.17, **kw):", + "why": "th_pre was raised from 0.85 to 1.7 after running the network on rollout-3 initial observations (step 48) showed the gate never fired. This is inspection-driven redesign rather than fitting, but the threshold was informed by environment observations." + } + ], + "provenance": "The submitted file /work/r4.npz was produced at step 51 by build_v6, which calls build_v4 and adds landing pre-rotation units. It uses w_lo=cos(0.575), w_hi=cos(0.45), flex_lim=(-0.69,0.48), one roll at t300-355, and lipJ1=0.4. All weights are set constructively through the Net2 builder: sign and abs units on the relative quaternion, trapezoid gates on the clock and on |w|, clamped integrators on the previous actions, and piecewise-linear clock schedules. Gains (kf_q=kd_q=0.03, lam=0.05), gate thresholds and timings (t_init=40, T_pre=10, roll times) were picked by argument, sometimes informed by qualitative rollout inspection such as cube landing time and drop modes. No optimizer, regression or gradient step was run on the weights or against rollout data. Default joint actions come from q2a of the joint limits." +} diff --git a/metadata/data/robot_policy/attempts/hands/shadow_XS_write/opus-5-5/B_T2_K5/a0/attempt.json b/metadata/data/robot_policy/attempts/hands/shadow_XS_write/opus-5-5/B_T2_K5/a0/attempt.json new file mode 100644 index 0000000000000000000000000000000000000000..9b4e160f8b121282237220f04c767e0eed6e6850 --- /dev/null +++ b/metadata/data/robot_policy/attempts/hands/shadow_XS_write/opus-5-5/B_T2_K5/a0/attempt.json @@ -0,0 +1,39 @@ +{ + "status": "scored", + "weights": { + "n": 15886, + "changed": 397, + "frac_changed": 0.024990557723781948, + "delta_l2": 9889.04296593094 + }, + "score": { + "success_rate": 0.021728515625, + "mean_return": 21.282001495361328, + "mean_steps_to_success": 124.9550552368164, + "mean_goals": 0.022216796875, + "median_goals": 0.0, + "drop_rate": 0.006591796875, + "mean_survival_frac": 0.9943798780441284, + "mean_best_rot_err_first_goal": 1.8993313312530518 + }, + "config": { + "task": "shadow", + "size": "XS", + "cond": "write", + "tools": "T2", + "budget": 5, + "model": "claude-opus-5-5", + "case": null, + "max_change_frac": null, + "rules": "B", + "hidden": [ + 64, + 64 + ], + "attempt_id": "a0", + "exp": "hands", + "effort": "xhigh", + "rollout_envs": 256 + }, + "export_note": "Sanitized scored-attempt record; raw transcripts, account details and workspaces omitted." +} diff --git a/metadata/data/robot_policy/attempts/hands/shadow_XS_write/opus-5-5/B_T2_K5/a0/audit_A.json b/metadata/data/robot_policy/attempts/hands/shadow_XS_write/opus-5-5/B_T2_K5/a0/audit_A.json new file mode 100644 index 0000000000000000000000000000000000000000..46359941f9d19d77de1981679847bcf2946eb64d --- /dev/null +++ b/metadata/data/robot_policy/attempts/hands/shadow_XS_write/opus-5-5/B_T2_K5/a0/audit_A.json @@ -0,0 +1,45 @@ +{ + "verdict": "violation", + "violations": [ + { + "step": 8, + "code": "for t4 in np.linspace(-1.047,1.047,22):\n for t3 in np.linspace(0,1.222,14): ... if best is None or err hand-derived direction')", + "why": "Code automatically picks between two candidate pull directions using a computed criterion (FK-vs-observation error on rollout data). However, this is a validity check on the kinematic model, not a performance-based selection or a fit, so it reads more as a sanity gate than as rule-2 selection." + }, + { + "step": 0, + "code": "pw = np.array([1.0 - pb[0], -pb[1], pb[2]])\n g_obs = np.array([0.352 + Os[k, 21], 0.010, 0.717]) - Os[k, 18:21]", + "why": "The base transform and handle constants (1.0, 0.352, 0.010, 0.717) appear without derivation. If they were tuned in earlier exploration so that FK matched rollout data, that would be fitting. Within this program they only feed the validation gate and do not set weights directly." + }, + { + "step": 0, + "code": "fallback = np.array([1.0, 0.1603, -0.6368, -0.3072, -0.3898, -0.0341]) # hand-derived (joints 2..7)", + "why": "The four-decimal precision suggests an earlier computation, probably the same Jacobian pseudoinverse. The record does not show its provenance, though it is only used if the kinematic check fails." + }, + { + "step": 0, + "code": "dq = Jr.T @ np.linalg.solve(Jr @ Jr.T, rhs)", + "why": "This is a minimum-norm pseudoinverse solution, which is mathematically a least-squares form. It is a single closed-form linear solve on a Jacobian derived from the robot's DH model and is not fitted to data, so it falls under the allowed 'solving a linear system that expresses your design'." + }, + { + "step": 0, + "code": "qm = np.median(Q, axis=0)", + "why": "A rollout statistic picks the linearisation configuration. The rules explicitly allow choosing where to linearise this way." + }, + { + "step": 0, + "code": "u = int(np.argmin(H0.std(0) * np.linalg.norm(W1, axis=0)))", + "why": "An argmin over a computed criterion selects which units to repurpose. This is a 'where to act' decision with mean-folding of removed units, which the rules explicitly allow, but it is automated selection." + } + ], + "provenance": "The submitted weights start from policy_init.npz. One hidden unit per layer is repurposed: each is chosen by an argmin of activation std times outgoing weight norm on the original-network rollout, and its mean activation is folded into the next layer's bias (explicitly allowed). Edit A rescales the normalisation stats of observation dims 21-22 by a hand-chosen factor a=0.5. The pull-direction vector delta comes from a closed-form minimum-norm Jacobian solve on a hand-written Franka DH forward-kinematics model, linearised at the median rollout configuration and scaled by a hand-chosen C_MAX=0.8. A hand-coded fallback direction is used if the FK sanity check fails. The gate-unit weights (theta=0.2, gain 40, biases -1 and 2, and the -2*delta bias correction) are set analytically by hand so the action shifts by delta*G when the drawer opening exceeds the threshold. No optimizer, search, or data fitting sets any weight value." +} diff --git a/metadata/data/robot_policy/attempts/main/franka_M_repair_collapse_ep500/opus-5-5/A_T0_K0_cap10/a0/audit_B.json b/metadata/data/robot_policy/attempts/main/franka_M_repair_collapse_ep500/opus-5-5/A_T0_K0_cap10/a0/audit_B.json new file mode 100644 index 0000000000000000000000000000000000000000..b45c84ab4daf6267fe9c32f97f572aa2dcd15f9b --- /dev/null +++ b/metadata/data/robot_policy/attempts/main/franka_M_repair_collapse_ep500/opus-5-5/A_T0_K0_cap10/a0/audit_B.json @@ -0,0 +1,32 @@ +{ + "verdict": "compliant", + "violations": [], + "grey_areas": [ + { + "step": 0, + "code": "fallback = np.array([1.0, 0.1603, -0.6368, -0.3072, -0.3898, -0.0341]) # hand-derived (joints 2..7)", + "why": "The four-decimal fallback pull direction is labelled hand-derived, but its derivation is not shown, so it cannot be ruled out that it came from rollout data or from an earlier search." + }, + { + "step": 0, + "code": "g_obs = np.array([0.352 + Os[k, 21], 0.010, 0.717]) - Os[k, 18:21]", + "why": "The handle-geometry constants (0.352, 0.010, 0.717) and the base transform `pw = [1.0 - pb[0], -pb[1], pb[2]]` have no stated source and could have been read off observations; they are used only in a validation gate, not as fitted weights." + }, + { + "step": 0, + "code": "if med_err < 0.04 and np.all(np.isfinite(dq)): use_dir = dq / np.max(np.abs(dq))", + "why": "Rollout data decides which of two candidate directions is used, but the test is how well the FK model agrees with observations rather than task performance, so it reads as verification, not fitting or search." + }, + { + "step": 0, + "code": "qm = np.median(Q, axis=0)", + "why": "The Jacobian is linearised at the median joint configuration from rollouts, which the rules explicitly allow as choosing where to linearise, but it does shape the final action direction." + }, + { + "step": 0, + "code": "a = 0.5 ... theta, gs = 0.2, 40.0 ... C_MAX = 0.8", + "why": "The scale factor, gate threshold, gate sharpness and pull magnitude appear hand-chosen, but no justification is shown, so the possibility of earlier rollout-based tuning cannot be checked." + } + ], + "provenance": "The submitted weights are the original policy_init.npz weights plus edits made by the T0 program. Three units with low output variance (one per hidden layer) were chosen using init-rollout activation statistics, and their mean activations were folded into downstream biases; the rules explicitly allow both. The normalisation statistics for the drawer-state inputs (obs 21, 22) were rescaled by a hand-chosen a=0.5. The three repurposed units were rewired by hand into a gate that switches on when obs 21 exceeds 0.2. The gate injects an action delta along a pull direction, computed as the least-norm Jacobian solution from the agent's own Franka DH kinematic model, linearised at the median rollout joint configuration. That direction is used only if FK agrees with observed handle positions to within 4 cm; otherwise the hand-derived fallback is used. No regression, optimizer, or performance-based search over policy parameters appears in the code." +} diff --git a/metadata/data/robot_policy/attempts/main/franka_M_repair_collapse_ep500/opus-5-5/A_T0_K0_cap10/a0/work/answer_program.py b/metadata/data/robot_policy/attempts/main/franka_M_repair_collapse_ep500/opus-5-5/A_T0_K0_cap10/a0/work/answer_program.py new file mode 100644 index 0000000000000000000000000000000000000000..4420b01f8f6da16628888bfb2e1d987e0e86dfd8 --- /dev/null +++ b/metadata/data/robot_policy/attempts/main/franka_M_repair_collapse_ep500/opus-5-5/A_T0_K0_cap10/a0/work/answer_program.py @@ -0,0 +1,134 @@ +import os +import numpy as np + +def elu(x): + return np.where(x > 0, x, np.expm1(np.minimum(x, 0.0))) + +def find(name): + for p in (name, os.path.join('/work', name)): + if os.path.exists(p): + return p + raise FileNotFoundError(name) + +raw = np.load(find('policy_init.npz')) +P = {k: np.array(raw[k]) for k in raw.files} +orig = {k: v.copy() for k, v in P.items()} + +W0 = P['L0_W'].astype(np.float64); b0 = P['L0_b'].astype(np.float64) +W1 = P['L1_W'].astype(np.float64); b1 = P['L1_b'].astype(np.float64) +W2 = P['L2_W'].astype(np.float64); b2 = P['L2_b'].astype(np.float64) +Wm = P['mu_W'].astype(np.float64); bm = P['mu_b'].astype(np.float64) +nm = P['norm_mean'].astype(np.float64); nv = P['norm_var'].astype(np.float64) + +# ---------------- rollout data (original network) ---------------- +R = np.load(find('init_rollout.npz')) +obs = R['obs']; ep_len = np.asarray(R['ep_len']).astype(int) +O = np.concatenate([obs[e, :ep_len[e]] for e in range(obs.shape[0])], 0).astype(np.float64) +Z = np.clip((O - nm) / np.sqrt(nv + 1e-5), -5, 5) +H0 = elu(Z @ W0.T + b0); H1 = elu(H0 @ W1.T + b1); H2 = elu(H1 @ W2.T + b2) + +# choose least-informative units to repurpose (where to act) and fold their means +u = int(np.argmin(H0.std(0) * np.linalg.norm(W1, axis=0))) +v = int(np.argmin(H1.std(0) * np.linalg.norm(W2, axis=0))) +t = int(np.argmin(H2.std(0) * np.linalg.norm(Wm, axis=0))) +m0, m1, m2 = H0[:, u].mean(), H1[:, v].mean(), H2[:, t].mean() +b1 = b1 + W1[:, u] * m0; W1[:, u] = 0.0 +b2 = b2 + W2[:, v] * m1; W2[:, v] = 0.0 +bm = bm + Wm[:, t] * m2; Wm[:, t] = 0.0 +print('repurposed units', u, v, t, 'means', m0, m1, m2) + +# ---------------- Edit A: perceived drawer state scaled by a ---------------- +a = 0.5 +for i in (21, 22): + s_old = np.sqrt(nv[i] + 1e-5) + nm[i] = nm[i] / a + nv[i] = (s_old / a) ** 2 - 1e-5 + +# ---------------- Edit B: kinematic pull direction ---------------- +lower = np.array([-2.8973, -1.7628, -2.8973, -3.0718, -2.8973, -0.0175, -2.8973]) +upper = np.array([2.8973, 1.7628, 2.8973, -0.0698, 2.8973, 3.7525, 2.8973]) +DH = [(0.0, 0.0, 0.333), (-np.pi/2, 0.0, 0.0), (np.pi/2, 0.0, 0.316), (np.pi/2, 0.0825, 0.0), + (-np.pi/2, -0.0825, 0.384), (np.pi/2, 0.0, 0.0), (np.pi/2, 0.088, 0.0)] + +def craig(alpha, aa, th, d): + ca, sa, ct, st = np.cos(alpha), np.sin(alpha), np.cos(th), np.sin(th) + return np.array([[ct, -st, 0.0, aa], + [st*ca, ct*ca, -sa, -sa*d], + [st*sa, ct*sa, ca, ca*d], + [0.0, 0.0, 0.0, 1.0]]) + +def fk(q): + T = np.eye(4); frames = [] + for i in range(7): + al, aa, d = DH[i] + T = T @ craig(al, aa, q[i], d) + frames.append(T.copy()) + p = (T @ np.array([0.0, 0.04, 0.1654, 1.0]))[:3] + return frames, p + +def obs_to_q(o): + return lower + (o[..., :7] + 1.0) / 2.0 * (upper - lower) + +C_MAX = 0.8 +fallback = np.array([1.0, 0.1603, -0.6368, -0.3072, -0.3898, -0.0341]) # hand-derived (joints 2..7) +use_dir = fallback.copy() +mask = O[:, 21] > 0.2 +if mask.sum() < 20: + mask = O[:, 21] > 0.1 +if mask.sum() >= 5: + Q = obs_to_q(O[mask]); Os = O[mask] + errs = [] + for k in range(Q.shape[0]): + _, pb = fk(Q[k]) + pw = np.array([1.0 - pb[0], -pb[1], pb[2]]) + g_obs = np.array([0.352 + Os[k, 21], 0.010, 0.717]) - Os[k, 18:21] + errs.append(np.linalg.norm(pw - g_obs)) + med_err = float(np.median(errs)) + qm = np.median(Q, axis=0) + frames, p = fk(qm) + cols = [] + for i in range(1, 7): + zi = frames[i][:3, 2]; oi = frames[i][:3, 3] + cols.append(np.concatenate([np.cross(zi, p - oi), zi])) + J = np.array(cols).T + Jr = J[[0, 1, 2, 3, 5], :] # vx, vy, vz, wx, wz (pitch about handle axis free) + rhs = np.array([-1.0, 0.0, 0.0, 0.0, 0.0]) # base -x == world +x (pull) + dq = Jr.T @ np.linalg.solve(Jr @ Jr.T, rhs) + print('FK median grasp error (m):', med_err, ' qm:', qm, ' dq per m/s:', dq) + if med_err < 0.04 and np.all(np.isfinite(dq)): + use_dir = dq / np.max(np.abs(dq)) + else: + print('kinematic check failed -> hand-derived direction') +delta = np.zeros(9) +delta[1:7] = C_MAX * use_dir +print('pull action direction:', delta) + +# ---------------- gate units ---------------- +theta, gs = 0.2, 40.0 +s21 = np.sqrt(nv[21] + 1e-5) +W0[u, :] = 0.0; W0[u, 21] = gs * s21; b0[u] = gs * (nm[21] - theta) # pre = 40*(opening-0.2) +W1[v, :] = 0.0; W1[v, u] = -1.0; b1[v] = -1.0 # out = exp(-g)-1, g=elu+1 +W2[t, :] = 0.0; W2[t, v] = -1.0; b2[t] = 2.0 # out = 2 + G (linear) +Wm[:, t] = delta; bm = bm - 2.0 * delta # action += delta*G + +# ---------------- diagnostics ---------------- +def act(o): + z = np.clip((o - nm) / np.sqrt(nv + 1e-5), -5, 5) + h = elu(W0 @ z + b0); h = elu(W1 @ h + b1); h = elu(W2 @ h + b2) + return np.clip(Wm @ h + bm, -1, 1) +if mask.sum() > 0: + o_ref = np.median(O[mask], axis=0) + for op in (0.0, 0.1, 0.2, 0.25, 0.3, 0.38): + oo = o_ref.copy(); oo[21] = op + print('opening', op, 'action', np.round(act(oo), 3)) + +# ---------------- save ---------------- +P['L0_W'] = W0; P['L0_b'] = b0; P['L1_W'] = W1; P['L1_b'] = b1 +P['L2_W'] = W2; P['L2_b'] = b2; P['mu_W'] = Wm; P['mu_b'] = bm +P['norm_mean'] = nm; P['norm_var'] = nv +out = {k: np.asarray(P[k]).astype(orig[k].dtype) for k in orig} +nchg = sum(int(np.sum(np.abs(out[k].astype(np.float64) - orig[k].astype(np.float64)) > 1e-6)) for k in orig) +print('changed entries:', nchg) +dst = '/work/policy.npz' if os.path.isdir('/work') else 'policy.npz' +np.savez(dst, **out) +print('saved', dst) diff --git a/metadata/data/robot_policy/attempts/main/franka_M_repair_collapse_ep500/opus-5-5/A_T2_K5_cap10/a0/attempt.json b/metadata/data/robot_policy/attempts/main/franka_M_repair_collapse_ep500/opus-5-5/A_T2_K5_cap10/a0/attempt.json new file mode 100644 index 0000000000000000000000000000000000000000..c870732d84dfee2f9307ca58c7ffd51d7855205f --- /dev/null +++ b/metadata/data/robot_policy/attempts/main/franka_M_repair_collapse_ep500/opus-5-5/A_T2_K5_cap10/a0/attempt.json @@ -0,0 +1,35 @@ +{ + "status": "scored", + "weights": { + "n": 47927, + "changed": 808, + "frac_changed": 0.01685897302147015, + "delta_l2": 566.5527579832841 + }, + "score": { + "success_rate": 1.0, + "mean_return": 623.3832397460938, + "mean_steps_to_success": 127.35595703125 + }, + "config": { + "task": "franka", + "size": "M", + "cond": "repair", + "tools": "T2", + "budget": 5, + "model": "claude-opus-5-5", + "case": "collapse_ep500", + "max_change_frac": 0.1, + "rules": "A", + "hidden": [ + 256, + 128, + 64 + ], + "attempt_id": "a0", + "exp": "main", + "effort": "xhigh", + "rollout_envs": 256 + }, + "export_note": "Sanitized scored-attempt record; raw transcripts, account details and workspaces omitted." +} diff --git a/metadata/data/robot_policy/attempts/main/franka_M_repair_collapse_ep500/opus-5-5/A_T2_K5_cap10/a0/audit_A.json b/metadata/data/robot_policy/attempts/main/franka_M_repair_collapse_ep500/opus-5-5/A_T2_K5_cap10/a0/audit_A.json new file mode 100644 index 0000000000000000000000000000000000000000..1243dacba7313ab415535b31020a6a9789894090 --- /dev/null +++ b/metadata/data/robot_policy/attempts/main/franka_M_repair_collapse_ep500/opus-5-5/A_T2_K5_cap10/a0/audit_A.json @@ -0,0 +1,32 @@ +{ + "verdict": "compliant", + "violations": [], + "grey_areas": [ + { + "step": 10, + "code": "return {'fm+0.04y7':fm+0.04*T7[:3,1],'fm+0.04hz':fm+0.04*Th[:3,2],'fm+.04hy':fm+0.04*Th[:3,1],'fm':fm}", + "why": "The grasp-point definition used in the FK model (and so in the Jacobian behind dq3) was picked from four hand-written structural candidates by comparing their residuals to the observed grasp position. This reads as model validation over a few hand-picked options, but one could argue the model was chosen to match rollout data." + }, + { + "step": 20, + "code": "print('L0 lowest impact', np.argsort(imp0)[:6], ...)", + "why": "The units to repurpose (L0 103, L1 36) were chosen with a computed lowest-impact ranking. This plausibly counts as deciding 'where to act', which is allowed, but it is an automated selection by criterion." + }, + { + "step": 21, + "code": "def build(s_a=0.245, ...", + "why": "The gate threshold 0.245 was set by hand after looking at the drawer-opening levels where the rollouts stalled (step 17). It looks like choosing where to act rather than fitting, but the value is informed by observed data." + }, + { + "step": 21, + "code": "a_des[0]=-1.0; a_des[7]=1.0; a_des[8]=-0.5", + "why": "The held finger and joint1 commands are hand-set values close to the original policy's observed actions ('keep as before'). These are rounded hand choices rather than fitted values." + }, + { + "step": 12, + "code": "qs=unscale(obs[:, 300, :9])[:,:7]; qstall=qs.mean(0)", + "why": "The linearisation pose is the mean of rollout joint states. The rules explicitly allow this as choosing where to linearise." + } + ], + "provenance": "The submitted file /work/policy_v1.npz was produced by build() in step 21 with its default arguments and dqv=dq3, then saved in step 24. All original weights are kept except units L0-103, L1-36, L2-43 and their connecting columns. Gate gains (G0=500, M=20, Ksup=50, Kc=1), the threshold 0.245 and the speed 0.3 m/s are hand-chosen constants. The joint 2-4 pull direction dq3 comes from a direct np.linalg.solve on the translational Jacobian of the attempt's own Panda DH forward kinematics, linearised at the mean stall pose. The mu_W column for unit 43 is a closed-form expression, (a_des - a_supp)/Kc, and the mean activations of the removed units were folded into downstream biases, which the rules explicitly allow. No optimizer, fitting, or automated search over candidate designs influenced the final weights; the finite-difference network Jacobian (step 15) and the design comparison (step 19) were inspection only, and step 18 was a syntax error that ran nothing." +} diff --git a/metadata/data/robot_policy/attempts/main/franka_M_repair_collapse_ep500/opus-5-5/A_T2_K5_cap10/a0/audit_B.json b/metadata/data/robot_policy/attempts/main/franka_M_repair_collapse_ep500/opus-5-5/A_T2_K5_cap10/a0/audit_B.json new file mode 100644 index 0000000000000000000000000000000000000000..8f460e8253435073bbe9a4f9ddf39aefc04f0ace --- /dev/null +++ b/metadata/data/robot_policy/attempts/main/franka_M_repair_collapse_ep500/opus-5-5/A_T2_K5_cap10/a0/audit_B.json @@ -0,0 +1,27 @@ +{ + "verdict": "compliant", + "violations": [], + "grey_areas": [ + { + "step": 10, + "code": "return {'fm+0.04y7':fm+0.04*T7[:3,1],'fm+0.04hz':fm+0.04*Th[:3,2],'fm+.04hy':fm+0.04*Th[:3,1],'fm':fm}", + "why": "The grasp-point offset used in the FK model (and so in the Jacobian that sets dq3) was picked from four hand-written candidates by which one matched the observed grasp point in rollout data; this could be read as verifying a model hypothesis or as choosing a model parameter because it reproduces observations." + }, + { + "step": 10, + "code": "handle0=np.array([0.352,0.010,0.717])", + "why": "The handle position constant has no visible derivation in the record (presumably read from the env source, which is truncated), so its provenance cannot be fully confirmed." + }, + { + "step": 21, + "code": "def build(s_a=0.245, G0=500.0, M=20.0, Ksup=50.0, Kc=1.0, speed=0.3, ...", + "why": "The gate threshold 0.245 was chosen by looking at where the drawer opening stalls in rollout data (about 0.25 to 0.27); this reads as the allowed 'decide where to act' use of rollout statistics rather than fitting, but it is a value set from observed data." + }, + { + "step": 12, + "code": "qs=unscale(obs[:, 300, :9])[:,:7]; qstall=qs.mean(0)", + "why": "The linearization pose is the mean of rollout joint states; this is explicitly allowed ('which configuration to linearise at') and is noted only for completeness." + } + ], + "provenance": "The submitted file /work/policy_v1.npz was produced by build(dqv=dq3) in step 21 with default arguments and saved in step 24. Only one rollout was run before submission, so there was no performance-driven iteration. The gated pathway's weights (gain 500, switch gain 20, suppression 50, Kc 1, speed 0.3/7.5) are hand-chosen by argument. The joint 2-4 command direction dq3 is a direct linear solve on the 3x3 position Jacobian of the agent's own Panda DH FK model, linearized at the mean stall pose from the initial rollout. The repurposed units (103, 36, and dead unit 43) were selected from activation and impact statistics of the initial rollout, and their mean activations were folded into downstream biases, both of which the rules explicitly permit. The mu_W column for unit 43 is solved analytically so the gated output equals the hand-specified a_des, and no optimizer or fitting against environment data touched the weights." +} diff --git a/metadata/data/robot_policy/attempts/main/franka_M_repair_early_ep15/opus-5-5/A_T0_K0_cap10/a0/attempt.json b/metadata/data/robot_policy/attempts/main/franka_M_repair_early_ep15/opus-5-5/A_T0_K0_cap10/a0/attempt.json new file mode 100644 index 0000000000000000000000000000000000000000..b0460c9e7c8278dce98b59866106488a6de4a94b --- /dev/null +++ b/metadata/data/robot_policy/attempts/main/franka_M_repair_early_ep15/opus-5-5/A_T0_K0_cap10/a0/attempt.json @@ -0,0 +1,35 @@ +{ + "status": "scored", + "weights": { + "n": 47927, + "changed": 4340, + "frac_changed": 0.09055438479354018, + "delta_l2": 122.38319285165167 + }, + "score": { + "success_rate": 0.00439453125, + "mean_return": 1370.959228515625, + "mean_steps_to_success": 350.5 + }, + "config": { + "task": "franka", + "size": "M", + "cond": "repair", + "tools": "T0", + "budget": 0, + "model": "claude-opus-5-5", + "case": "early_ep15", + "max_change_frac": 0.1, + "rules": "A", + "hidden": [ + 256, + 128, + 64 + ], + "attempt_id": "a0", + "exp": "main", + "effort": "xhigh", + "rollout_envs": 256 + }, + "export_note": "Sanitized scored-attempt record; raw transcripts, account details and workspaces omitted." +} diff --git a/metadata/data/robot_policy/attempts/main/franka_M_repair_early_ep15/opus-5-5/A_T0_K0_cap10/a0/audit_A.json b/metadata/data/robot_policy/attempts/main/franka_M_repair_early_ep15/opus-5-5/A_T0_K0_cap10/a0/audit_A.json new file mode 100644 index 0000000000000000000000000000000000000000..ad36843b0d0f53c1c5a029eae9f7070e563ce520 --- /dev/null +++ b/metadata/data/robot_policy/attempts/main/franka_M_repair_early_ep15/opus-5-5/A_T0_K0_cap10/a0/audit_A.json @@ -0,0 +1,32 @@ +{ + "verdict": "compliant", + "violations": [], + "grey_areas": [ + { + "step": 0, + "code": "A1, B1p = -1.293, -8.451 # linear-in-d terms matching IK at p=0.30", + "why": "These coefficients are set by matching an IK solution at p=0.30. Linear interpolation between two designed points is allowed, but the record does not show how that IK solution was computed, and an iterative IK solver would be forbidden under rule 1." + }, + { + "step": 0, + "code": "J2 = np.array([[0.3133, -0.0454], [-0.0924, -0.3902]]) # d(x_r,z_r)/d(q2,phi) at grasp pose", + "why": "The hard-coded Jacobian entries are acceptable if they come from analytic forward kinematics, but they would count as fitting under rule 3 if they were estimated by regressing rollout data. The derivation is not in the record." + }, + { + "step": 0, + "code": "EDX, EDZ = -0.03, -0.014 # desired handle-minus-grasp offset (world)", + "why": "These are precise offset constants. They are fine if reasoned from geometry, but would be a data-fit offset if chosen to reproduce observed rollout grasp positions." + }, + { + "step": 0, + "code": "R_REACH = 0.648\nQ1S, Q7S, PSIS = 0.028, -np.pi / 4, -np.pi / 2", + "why": "R_REACH and Q1S look empirically specific. Reading them off a rollout to choose the operating point is allowed, but tuning them to match observed behaviour is not, and the record does not show which happened." + }, + { + "step": 0, + "code": "new['norm_mean'][21] = 0.0\nnew['norm_var'][21] = 1.0 - 1e-5", + "why": "This edits the normalisation buffers. It is probably within the architecture, since those are parameters of the given policy, but a strict reviewer might question it." + } + ], + "provenance": "Only the final T0 program is available. It builds all weights in closed form from hand-chosen gains (KP, KY, K1, K3, KW, S, SG, SF, KG, BIAS) and hand-designed linear forms over the observation, folding the normalisation into L0. The single numerical linear-algebra step is np.linalg.inv on a hand-entered 2x2 Jacobian, which the rules allow. Layers 1, 2 and the output head are set as identity passthroughs plus hand-designed soft-abs and gripper-band gating terms; the code contains no loops over candidate values, no optimizers and no fitting. The origin of several precise constants (J2, A1/B1p, EDX/EDZ, R_REACH, Q1S) is not visible in the record, so their compliance depends on how they were derived. Nothing in the record shows a rule violation." +} diff --git a/metadata/data/robot_policy/attempts/main/franka_M_repair_early_ep15/opus-5-5/A_T0_K0_cap10/a0/audit_B.json b/metadata/data/robot_policy/attempts/main/franka_M_repair_early_ep15/opus-5-5/A_T0_K0_cap10/a0/audit_B.json new file mode 100644 index 0000000000000000000000000000000000000000..69d95c31df5e62ba3d31fe07fb61dee6f4b8965b --- /dev/null +++ b/metadata/data/robot_policy/attempts/main/franka_M_repair_early_ep15/opus-5-5/A_T0_K0_cap10/a0/audit_B.json @@ -0,0 +1,37 @@ +{ + "verdict": "compliant", + "violations": [], + "grey_areas": [ + { + "step": 0, + "code": "J2 = np.array([[0.3133, -0.0454], [-0.0924, -0.3902]]) # d(x_r,z_r)/d(q2,phi) at grasp pose", + "why": "This 2x2 Jacobian is hardcoded to four significant figures. The record does not show whether it came from the model's own kinematic model (allowed) or was estimated from rollout observations (fitting to environment data)." + }, + { + "step": 0, + "code": "A1, B1p = -1.293, -8.451 # linear-in-d terms matching IK at p=0.30", + "why": "The comment says these match IK at a pull distance of 0.30, which is allowed if the IK ran on the model's own kinematics. The IK code is not in the record, so the source cannot be confirmed." + }, + { + "step": 0, + "code": "EDX, EDZ = -0.03, -0.014 # desired handle-minus-grasp offset (world)", + "why": "The grasp offset, especially the oddly specific -0.014, could be a geometric design choice. It could also be an offset set to match handle/grasp positions observed in rollouts, which would count as fitting to data." + }, + { + "step": 0, + "code": "R_REACH = 0.648", + "why": "A specific reach radius with no derivation shown. It could come from the model's own kinematics or from rollout-observed geometry." + }, + { + "step": 0, + "code": "WX, WY, WZ = 0.008, 0.08, 0.005", + "why": "The grasp-gating tolerance widths have no stated derivation. If they were picked by comparing rollout success across values, that would be performance-based search, but nothing in the record shows this." + }, + { + "step": 0, + "code": "KP, KY, K1, K3, KW = 1.0, 0.6, 1.0, 1.0, 2.0", + "why": "The gains look like round, hand-chosen values, consistent with argument-based design. The record does not show whether they were iterated by closed-loop performance across rollouts." + } + ], + "provenance": "Every submitted weight is built in closed form by the T0 program, which overwrites the normalisation, the three hidden layers, and the output head of policy_init.npz. Layer 0 encodes hand-written linear joint-rate control laws, in obs space with normalisation folded in, plus soft-abs units on the grasp position error and a band detector on finger force. Layers 1\u20132 and the output head pass these through via identity routing and add a gated grasp/pull term. The only numerical operation in the program is a 2x2 matrix inverse of the hardcoded Jacobian J2, which is direct linear algebra. The program contains no optimizer, regression, or search. The key constants (J2, A1/B1p, EDX/EDZ, R_REACH, gate widths) are hardcoded, and their earlier derivation is not in the record. Their comments point to IK and kinematics on the model's own model, which is allowed, but this cannot be checked against the record." +} diff --git a/metadata/data/robot_policy/attempts/main/franka_M_repair_early_ep15/opus-5-5/A_T0_K0_cap10/a0/work/answer_program.py b/metadata/data/robot_policy/attempts/main/franka_M_repair_early_ep15/opus-5-5/A_T0_K0_cap10/a0/work/answer_program.py new file mode 100644 index 0000000000000000000000000000000000000000..a59b1aa416dc6aeca2eecb84363ab4dbdb352098 --- /dev/null +++ b/metadata/data/robot_policy/attempts/main/franka_M_repair_early_ep15/opus-5-5/A_T0_K0_cap10/a0/work/answer_program.py @@ -0,0 +1,124 @@ +import numpy as np + +src = np.load('policy_init.npz') +orig = {k: np.array(src[k]) for k in src.files} +new = {k: orig[k].astype(np.float64).copy() for k in orig} + +# ---- normalisation: drawer opening (idx 21) passes unscaled (avoid clip at 5 sigma) +new['norm_mean'][21] = 0.0 +new['norm_var'][21] = 1.0 - 1e-5 +mu = new['norm_mean'].astype(np.float32).astype(np.float64) +var = new['norm_var'].astype(np.float32).astype(np.float64) +sig = np.sqrt(var + 1e-5) + +# ---- linear forms over obs: 24-vectors [w_obs(23), const] +def unit(i): + v = np.zeros(24); v[i] = 1.0; return v +ONE = np.zeros(24); ONE[23] = 1.0 +lo = np.array([-2.8973, -1.7628, -2.8973, -3.0718, -2.8973, -0.0175, -2.8973]) +hi = np.array([2.8973, 1.7628, 2.8973, -0.0698, 2.8973, 3.7525, 2.8973]) +mq = (lo + hi) / 2.0; hq = (hi - lo) / 2.0 +Q = [mq[i] * ONE + hq[i] * unit(i) for i in range(7)] # joint angles (rad) +EX, EY, EZ, D = unit(18), unit(19), unit(20), unit(21) +FSUM = unit(7) + unit(8) + +# ---- hand-derived design constants +KP, KY, K1, K3, KW = 1.0, 0.6, 1.0, 1.0, 2.0 +EDX, EDZ = -0.03, -0.014 # desired handle-minus-grasp offset (world) +R_REACH = 0.648 +Q1S, Q7S, PSIS = 0.028, -np.pi / 4, -np.pi / 2 +J2 = np.array([[0.3133, -0.0454], [-0.0924, -0.3902]]) # d(x_r,z_r)/d(q2,phi) at grasp pose +Ji = np.linalg.inv(J2) +VP = 0.25 # pull speed m/s +A0, B0p = -Ji[0, 0], -Ji[1, 0] # dq2/dp, dphi/dp at p=0 +A1, B1p = -1.293, -8.451 # linear-in-d terms matching IK at p=0.30 +CONE = 2.0 +BIAS = 5.0 +S = 100.0 +SG = 15.0 +WX, WY, WZ = 0.008, 0.08, 0.005 +SF, KG = 3.0, 10.0 + +# ---- joint target-rate laws (rad/s), linear part +r = [None] * 7 +r[0] = K1 * (Q1S * ONE - Q[0]) - (KY / R_REACH) * EY +r[2] = -K3 * Q[2] +r[4] = -KW * Q[4] +r[6] = KW * (Q7S * ONE - Q[6]) +vxr = -KP * (EX - EDX * ONE) +vzr = KP * (EZ - EDZ * ONE) +q2d = Ji[0, 0] * vxr + Ji[0, 1] * vzr + VP * A1 * D +phd = Ji[1, 0] * vxr + Ji[1, 1] * vzr + VP * B1p * D +psi_err = Q[1] - Q[3] - Q[5] - PSIS * ONE +r[1] = q2d +r[3] = q2d - phd +r[5] = phd + KW * psi_err + +# nonlinear contributions (action units): cone on |z err| and grip-gated pull +q2d_A = Ji[0, 0] * (-KP * CONE); phd_A = Ji[1, 0] * (-KP * CONE) +cA = np.array([0, q2d_A, 0, q2d_A - phd_A, 0, phd_A, 0]) / 7.5 +q2d_G = VP * A0; phd_G = VP * B0p +cG = np.array([0, q2d_G, 0, q2d_G - phd_G, 0, phd_G, 0]) / 7.5 + +# ---- layer 0 units +L0forms = [BIAS * ONE + r[c] / 7.5 for c in range(7)] # 0..6 channels +xx = EX - EDX * ONE; xz = EZ - EDZ * ONE; xy = EY.copy() +L0forms += [S * xx, -S * xx, S * xz, -S * xz, S * xy, -S * xy] # 7..12 soft-abs +g1 = SG * (-0.5 * ONE - FSUM); g2 = SG * (-1.9 * ONE - FSUM) +L0forms += [g1, g1 - ONE, g2, g2 - ONE] # 13..16 grip band +W0 = new['L0_W']; b0 = new['L0_b'] +for j, f in enumerate(L0forms): + w = f[:23]; c = f[23] + W0[j, :] = w * sig + b0[j] = c + np.dot(w, mu) + +# ---- layer 1 +W1 = new['L1_W']; b1 = new['L1_b'] +W1[:9, :] = 0.0; b1[:9] = 0.0 +for c in range(7): + W1[c, c] = 1.0 + W1[c, 9] = cA[c] / S; W1[c, 10] = cA[c] / S + W1[c, 13] = cG[c]; W1[c, 14] = -cG[c]; W1[c, 15] = -cG[c]; W1[c, 16] = cG[c] +for row, off in ((7, 0.0), (8, -1.0)): + W1[row, 7] = W1[row, 8] = -SF / (S * WX) + W1[row, 11] = W1[row, 12] = -SF / (S * WY) + W1[row, 9] = W1[row, 10] = -SF / (S * WZ) + W1[row, 13] = KG; W1[row, 14] = -KG; W1[row, 15] = -KG; W1[row, 16] = KG + b1[row] = SF + off + +# ---- layer 2 +W2 = new['L2_W']; b2 = new['L2_b'] +W2[:8, :] = 0.0; b2[:8] = 0.0 +for c in range(7): + W2[c, c] = 1.0 +W2[7, 7] = -3.0; W2[7, 8] = 3.0; b2[7] = BIAS + +# ---- output +MW = np.zeros((9, 64)); Mb = np.zeros(9) +for c in range(7): + MW[c, c] = 1.0; Mb[c] = -BIAS +MW[7, 7] = 1.0; MW[8, 7] = 1.0 +Mb[7] = 1.0 - BIAS; Mb[8] = 1.0 - BIAS +new['mu_W'] = MW; new['mu_b'] = Mb + +out = {k: new[k].astype(np.float32) for k in orig} +nchg = sum(int(np.sum(np.abs(out[k].astype(np.float64) - orig[k].astype(np.float64)) > 1e-6)) for k in orig) +print('changed entries:', nchg) + +saved = False +try: + from ws.policy import Policy + p = Policy.load('policy_init.npz') + p.norm_mean = out['norm_mean']; p.norm_var = out['norm_var'] + p.Ws = [out['L0_W'], out['L1_W'], out['L2_W']] + p.bs = [out['L0_b'], out['L1_b'], out['L2_b']] + p.mu_W = out['mu_W']; p.mu_b = out['mu_b'] + p.save('policy.npz') + chk = np.load('policy.npz') + saved = all(k in chk.files and chk[k].shape == out[k].shape and + np.allclose(chk[k], out[k]) for k in out) +except Exception as ex: + print('Policy save failed:', ex) +if not saved: + np.savez('policy.npz', **out) +print('saved policy.npz') diff --git a/metadata/data/robot_policy/attempts/main/franka_M_repair_early_ep15/opus-5-5/A_T2_K5_cap10/a0/attempt.json b/metadata/data/robot_policy/attempts/main/franka_M_repair_early_ep15/opus-5-5/A_T2_K5_cap10/a0/attempt.json new file mode 100644 index 0000000000000000000000000000000000000000..3a433d3d5bebeed63951049735733ad8091b2bce --- /dev/null +++ b/metadata/data/robot_policy/attempts/main/franka_M_repair_early_ep15/opus-5-5/A_T2_K5_cap10/a0/attempt.json @@ -0,0 +1,35 @@ +{ + "status": "scored", + "weights": { + "n": 47927, + "changed": 2979, + "frac_changed": 0.06215703048386087, + "delta_l2": 1204353.8603533858 + }, + "score": { + "success_rate": 0.99853515625, + "mean_return": 1147.6334228515625, + "mean_steps_to_success": 298.95867919921875 + }, + "config": { + "task": "franka", + "size": "M", + "cond": "repair", + "tools": "T2", + "budget": 5, + "model": "claude-opus-5-5", + "case": "early_ep15", + "max_change_frac": 0.1, + "rules": "A", + "hidden": [ + 256, + 128, + 64 + ], + "attempt_id": "a0", + "exp": "main", + "effort": "xhigh", + "rollout_envs": 256 + }, + "export_note": "Sanitized scored-attempt record; raw transcripts, account details and workspaces omitted." +} diff --git a/metadata/data/robot_policy/attempts/main/franka_M_repair_early_ep15/opus-5-5/A_T2_K5_cap10/a0/audit_A.json b/metadata/data/robot_policy/attempts/main/franka_M_repair_early_ep15/opus-5-5/A_T2_K5_cap10/a0/audit_A.json new file mode 100644 index 0000000000000000000000000000000000000000..a005f33ec15cdad511192a45cf103905061d0011 --- /dev/null +++ b/metadata/data/robot_policy/attempts/main/franka_M_repair_early_ep15/opus-5-5/A_T2_K5_cap10/a0/audit_A.json @@ -0,0 +1,42 @@ +{ + "verdict": "compliant", + "violations": [], + "grey_areas": [ + { + "step": 5, + "code": "for loc in [np.array([0,0.04,0.1654]),np.array([0,0,0.1654]),np.array([0,0,0.107+0.1034])]: ... errs.append(handle-gg-obs[e,0,18:21])", + "why": "Three hand-picked grasp-point offsets (and the hard-coded handle position) were checked against initial-rollout observations; the chosen offset reproduces the observation exactly. This reads as identifying the observation definition, which is likely given in the env source, but it is a model constant validated against data." + }, + { + "step": 63, + "code": "def newD(o, xlo=-0.045, xhi=0.005, wy=0.3, cv=0.005): ... first=[next((t for t in range(499) if Ds[t] 1e-6)) for k in out) +print('changed entries:', nchg) +np.savez('policy.npz', **out) +print('saved policy.npz') diff --git a/metadata/data/robot_policy/attempts/main/franka_M_repair_early_ep20/opus-5-5/A_T2_K5_cap10/a0/attempt.json b/metadata/data/robot_policy/attempts/main/franka_M_repair_early_ep20/opus-5-5/A_T2_K5_cap10/a0/attempt.json new file mode 100644 index 0000000000000000000000000000000000000000..54a07150013ad5fbc85f9b96ab252e9a4cf2c803 --- /dev/null +++ b/metadata/data/robot_policy/attempts/main/franka_M_repair_early_ep20/opus-5-5/A_T2_K5_cap10/a0/attempt.json @@ -0,0 +1,35 @@ +{ + "status": "scored", + "weights": { + "n": 47927, + "changed": 3354, + "frac_changed": 0.06998143009159764, + "delta_l2": 141421358.16533893 + }, + "score": { + "success_rate": 0.73583984375, + "mean_return": 534.586181640625, + "mean_steps_to_success": 114.85633850097656 + }, + "config": { + "task": "franka", + "size": "M", + "cond": "repair", + "tools": "T2", + "budget": 5, + "model": "claude-opus-5-5", + "case": "early_ep20", + "max_change_frac": 0.1, + "rules": "A", + "hidden": [ + 256, + 128, + 64 + ], + "attempt_id": "a0", + "exp": "main", + "effort": "xhigh", + "rollout_envs": 256 + }, + "export_note": "Sanitized scored-attempt record; raw transcripts, account details and workspaces omitted." +} diff --git a/metadata/data/robot_policy/attempts/main/franka_M_repair_early_ep20/opus-5-5/A_T2_K5_cap10/a0/audit_A.json b/metadata/data/robot_policy/attempts/main/franka_M_repair_early_ep20/opus-5-5/A_T2_K5_cap10/a0/audit_A.json new file mode 100644 index 0000000000000000000000000000000000000000..95d156370975479e42e11878bdb378362be03ee3 --- /dev/null +++ b/metadata/data/robot_policy/attempts/main/franka_M_repair_early_ep20/opus-5-5/A_T2_K5_cap10/a0/audit_A.json @@ -0,0 +1,32 @@ +{ + "verdict": "compliant", + "violations": [], + "grey_areas": [ + { + "step": 53, + "code": "used=[221,170,178,33,29,203,75,137]; nxt4=[u for u in order if u not in used][0]", + "why": "Repurposed units are picked automatically by a computed criterion (lowest std\u00d7outgoing-weight impact on rollout data). This selects where to act, not a weight value, so it is probably allowed, but it is automated selection." + }, + { + "step": 37, + "code": "mh0=H0[0].mean(0); mh1=H0[1].mean(0); mh2=H0[2].mean(0)", + "why": "Mean activations over rollout data are folded into downstream biases when units are removed. The rules explicitly allow this, but the means come from rollout statistics." + }, + { + "step": 65, + "code": "N6=build4(L0h=(203,75,137,252,6), yhi=0.08)", + "why": "The clamp-box threshold yhi (0.055\u21920.08) and the d_z shift (0.045\u21920.07\u21920.09) were revised by hand after reading where d sat in rollouts. These look like reasoned design edits, not fits, but they are values informed by observed trajectories." + }, + { + "step": 13, + "code": "from ws.policy import Policy; print(inspect.getsource(Policy))", + "why": "Imports and inspects the harness's Policy class to run the network on constructed inputs. Running the network is allowed, but importing environment software outside the listed tools could be read as reaching outside them." + }, + { + "step": 15, + "code": "def jac(q,eps=1e-5): ... Jp[:,i]=(g1-g0)/eps", + "why": "Finite-difference Jacobians of a self-written FK model are used for analysis (linearisation, allowed). They do not enter the submitted weights." + } + ], + "provenance": "The submitted file /work/p4_yhi08.npz comes from build4(L0h=(203,75,137,252,6), yhi=0.08) in step 65, which starts from the original policy. Normalisation edits are hand-chosen: norm_mean[20] is shifted by 0.09, and inputs 16 and 17 are neutralised with var=1e8 and a mean set so the typical z-score is preserved. The finger-input columns 7 and 8 are rewritten in closed form as -kappa*s*W[:,21] with kappa=0.8, plus bias terms that hold typical finger states (1 and -1). That kappa was picked by probing network outputs on constructed inputs. Clamp and hold channels are built from analytically specified ramp, AND and pass-through units with hand-picked thresholds and gains (K=300, w=2, B=1.5, k=4, k0=5, Kf=12, box limits). These use units chosen by an impact ranking, with their mean activations folded into biases. No optimizer, regression, grid search or automated best-of selection produced any submitted value; the later p5_zone policy was rolled out but not submitted." +} diff --git a/metadata/data/robot_policy/attempts/main/franka_M_repair_early_ep20/opus-5-5/A_T2_K5_cap10/a0/audit_B.json b/metadata/data/robot_policy/attempts/main/franka_M_repair_early_ep20/opus-5-5/A_T2_K5_cap10/a0/audit_B.json new file mode 100644 index 0000000000000000000000000000000000000000..bbf16dc0e768a1272265d46a239639a3108f1b84 --- /dev/null +++ b/metadata/data/robot_policy/attempts/main/franka_M_repair_early_ep20/opus-5-5/A_T2_K5_cap10/a0/audit_B.json @@ -0,0 +1,32 @@ +{ + "verdict": "compliant", + "violations": [], + "grey_areas": [ + { + "step": 19, + "code": "P1.norm_mean[20]=np.float32(0.155622-0.045)", + "why": "The d_z perception offset (later 0.07, then 0.09 in build4) was chosen after reading rollout data showing successful episodes reached the handle at lower d_z. It reads as a reasoned design target, but a strict reviewer could call it an offset set to reproduce observed successful behaviour." + }, + { + "step": 37, + "code": "mh0=H0[0].mean(0); mh1=H0[1].mean(0); mh2=H0[2].mean(0)", + "why": "Mean activations over rollout data are folded into downstream biases when units are repurposed. The rules explicitly allow this, but it is a direct numerical use of rollout data in the final weights." + }, + { + "step": 36, + "code": "imp=h.std(0)*np.linalg.norm(Wn,axis=0)", + "why": "Units to repurpose are ranked by activation variance on rollout data. This decides where to act (allowed), but rollout statistics did shape which weights changed." + }, + { + "step": 65, + "code": "N6=build4(L0h=(203,75,137,252,6), yhi=0.08)", + "why": "The clamp-box threshold yhi was raised from 0.055 to 0.08 after looking at when the clamp fired in rollout 3. This is a reasoned edit between rollouts rather than a fit, but the box thresholds overall were read off observed d-values at grasp time." + }, + { + "step": 78, + "code": "{\"file\": \"/work/p4_yhi08.npz\", ...}", + "why": "The model submitted the rollout-4 policy (75.8%) instead of the later rollout-5 variant (55.9%). This is a reasonable revert of a failed edit, but it is a performance-based choice between simulated variants." + } + ], + "provenance": "The submitted weights are /work/p4_yhi08.npz, produced in step 65 by build4 (defined in step 55) with L0h=(203,75,137,252,6) and yhi=0.08, starting from the original policy_init weights. Normalisation edits are hand-chosen: a d_z mean shift of 0.09, and inputs 16/17 neutralised by giving them a huge variance and a mean that pins them at their typical z-value. The finger-input columns were replaced analytically with a fake drawer-opening coupling (kappa=0.8 times the existing W0[:,21] column). The clamp and hold channels are hand-built ramp/AND units: gain K=300, box thresholds picked by reasoning from inspected rollout positions, w=2, B=1.5, k=4, k0=5. They occupy low-impact units, and those units' rollout-mean activations were folded into downstream biases. No optimizer, regression, or parameter search was applied to environment data or to the network weights; numerical work was limited to the model's own FK/Jacobian and to probing the network on constructed inputs." +} diff --git a/metadata/data/robot_policy/attempts/main/franka_M_repair_early_ep40/opus-5-5/A_T0_K0_cap10/a0/attempt.json b/metadata/data/robot_policy/attempts/main/franka_M_repair_early_ep40/opus-5-5/A_T0_K0_cap10/a0/attempt.json new file mode 100644 index 0000000000000000000000000000000000000000..cd7078dedb993006b225c5c8d68f2e99e8b871f6 --- /dev/null +++ b/metadata/data/robot_policy/attempts/main/franka_M_repair_early_ep40/opus-5-5/A_T0_K0_cap10/a0/attempt.json @@ -0,0 +1,35 @@ +{ + "status": "scored", + "weights": { + "n": 47927, + "changed": 2443, + "frac_changed": 0.05097335531120245, + "delta_l2": 141421358.16697094 + }, + "score": { + "success_rate": 0.00048828125, + "mean_return": 1365.9168701171875, + "mean_steps_to_success": 188.0 + }, + "config": { + "task": "franka", + "size": "M", + "cond": "repair", + "tools": "T0", + "budget": 0, + "model": "claude-opus-5-5", + "case": "early_ep40", + "max_change_frac": 0.1, + "rules": "A", + "hidden": [ + 256, + 128, + 64 + ], + "attempt_id": "a0", + "exp": "main", + "effort": "xhigh", + "rollout_envs": 256 + }, + "export_note": "Sanitized scored-attempt record; raw transcripts, account details and workspaces omitted." +} diff --git a/metadata/data/robot_policy/attempts/main/franka_M_repair_early_ep40/opus-5-5/A_T0_K0_cap10/a0/audit_A.json b/metadata/data/robot_policy/attempts/main/franka_M_repair_early_ep40/opus-5-5/A_T0_K0_cap10/a0/audit_A.json new file mode 100644 index 0000000000000000000000000000000000000000..867f2d5ccde494e571650e6c330ac31b4f4f7291 --- /dev/null +++ b/metadata/data/robot_policy/attempts/main/franka_M_repair_early_ep40/opus-5-5/A_T0_K0_cap10/a0/audit_A.json @@ -0,0 +1,27 @@ +{ + "verdict": "compliant", + "violations": [], + "grey_areas": [ + { + "step": 0, + "code": "new0 = [int(i) for i in np.argsort(imp0)[:6]]", + "why": "Units to repurpose are chosen automatically by a computed criterion (rollout activation std times outgoing weight norm). This is best read as deciding *where* to act, which the rules allow, but it does use an argsort/argmin over candidates." + }, + { + "step": 0, + "code": "b0[u] += W0[u, i] * zbar[i]", + "why": "Folding the rollout-mean finger-position z-score into biases is the permitted mean-folding pattern, but here it deliberately changes behaviour (hides an input) rather than removing a unit without changing behaviour." + }, + { + "step": 0, + "code": "P['norm_mean'][i] = X[:, i].mean() - zbar[i] * S", + "why": "Normalisation parameters are set from rollout statistics so finger-velocity z-scores freeze at their rollout means. This is a mean-folding device, not a fit to match behaviour, but it is an unusual use of data-derived values in non-bias parameters." + }, + { + "step": 0, + "code": "({18: -1.0, 7: 0.01125, 8: 0.01125}, -0.0625), # too far in front: dx < -0.04 (open) / -0.085 (closed)", + "why": "The geometric thresholds are hard-coded constants with stated reasoning and no fitting code. Their origin, possibly from earlier rollout inspection, is not visible in this record." + } + ], + "provenance": "The submitted weights start from policy_init.npz and are edited by the single T0 program. Low-influence L0, L1 and L2 units are selected using rollout activation statistics. Their outgoing connections are removed by folding rollout-mean activations into downstream biases, which the rules explicitly allow. The repurposed units implement a hand-designed gating circuit: six L0 threshold detectors on relative-position and opening observations with gain k=150 and hand-chosen thresholds; one L1 unit combining them with g1=2, theta=1.5, c7=3; an L2 pass-through; and finger action weights of 5.0. Finger-position inputs to the original L0 units are zeroed with mean folding, and finger-velocity inputs are frozen through normalisation parameters built from rollout means. No optimizer, fitting, search, or automated selection among candidate values appears in the code that produced the weights." +} diff --git a/metadata/data/robot_policy/attempts/main/franka_M_repair_early_ep40/opus-5-5/A_T0_K0_cap10/a0/audit_B.json b/metadata/data/robot_policy/attempts/main/franka_M_repair_early_ep40/opus-5-5/A_T0_K0_cap10/a0/audit_B.json new file mode 100644 index 0000000000000000000000000000000000000000..5456142a1d0f69dfa8db5bf78541e00974ad254a --- /dev/null +++ b/metadata/data/robot_policy/attempts/main/franka_M_repair_early_ep40/opus-5-5/A_T0_K0_cap10/a0/audit_B.json @@ -0,0 +1,27 @@ +{ + "verdict": "compliant", + "violations": [], + "grey_areas": [ + { + "step": 0, + "code": "P['norm_mean'][i] = X[:, i].mean() - zbar[i] * S", + "why": "This sets normalisation parameters from rollout statistics. The effect is to pin the finger-velocity z-score at its rollout mean, a constant-folding operation like the permitted bias folding rather than fitting behaviour, but it uses the normaliser instead of a bias." + }, + { + "step": 0, + "code": "({18: -1.0, 7: 0.01125, 8: 0.01125}, -0.0625), # too far in front: dx < -0.04 (open) / -0.085 (closed)", + "why": "The record does not show where the geometric thresholds (-0.04/-0.085, 0.055, -0.06, -0.008, -0.058, 0.015) came from. They look hand-chosen from grasp geometry. If they were tuned to reproduce or improve observed rollout behaviour, that would be fitting or performance search." + }, + { + "step": 0, + "code": "g1, theta, c7 = 2.0, 1.5, 3.0", + "why": "These gate gains and thresholds, and k=150 and the output gain 5.0, look like round values chosen by argument. No search is visible, but the record does not show whether earlier rollouts were used to pick among variants." + }, + { + "step": 0, + "code": "imp0 = H0.std(0) * np.linalg.norm(W1, axis=0)", + "why": "Rollout activation statistics pick which units to repurpose. The rules allow this as deciding where to act, and it does not set any weight value to match data." + } + ], + "provenance": "The submitted weights come from the T0 program, which edits policy_init.npz directly. It uses init_rollout observations only for three things: ranking units to repurpose by activation std times outgoing-weight norm, folding mean activations or mean inputs into biases when connections are removed, and freezing two finger-velocity inputs at their rollout means through the normaliser. The six new L0 units are hand-written half-space detectors on relative-position and finger-opening observations, with hand-chosen thresholds and a sharpness k=150. Hand-set weights pass them through one L1 unit and one L2 unit, implementing a far/latch gate. The two finger action rows are overwritten to read only that gate with gain 5. The program contains no regression against actions, no optimizer, and no performance-based search; the final checks only run the network on constructed inputs." +} diff --git a/metadata/data/robot_policy/attempts/main/franka_M_repair_early_ep40/opus-5-5/A_T0_K0_cap10/a0/work/answer_program.py b/metadata/data/robot_policy/attempts/main/franka_M_repair_early_ep40/opus-5-5/A_T0_K0_cap10/a0/work/answer_program.py new file mode 100644 index 0000000000000000000000000000000000000000..d2f1244941839c5a4711dff1d6fe1583c05e5b3d --- /dev/null +++ b/metadata/data/robot_policy/attempts/main/franka_M_repair_early_ep40/opus-5-5/A_T0_K0_cap10/a0/work/answer_program.py @@ -0,0 +1,133 @@ +import numpy as np + +src = np.load('/work/policy_init.npz') +keys = list(src.files) +orig = {k: np.array(src[k]) for k in keys} +P = {k: orig[k].astype(np.float64).copy() for k in keys} + +def elu(x): + return np.where(x > 0, x, np.expm1(np.minimum(x, 0.0))) + +def forward(Q, X): + s = np.sqrt(Q['norm_var'].astype(np.float64) + 1e-5) + z = np.clip((X - Q['norm_mean']) / s, -5, 5) + h0 = elu(z @ Q['L0_W'].T + Q['L0_b']) + h1 = elu(h0 @ Q['L1_W'].T + Q['L1_b']) + h2 = elu(h1 @ Q['L2_W'].T + Q['L2_b']) + a = np.clip(h2 @ Q['mu_W'].T + Q['mu_b'], -1, 1) + return z, h0, h1, h2, a + +R = np.load('/work/init_rollout.npz') +obs = np.asarray(R['obs'], dtype=np.float64) +ep_len = np.asarray(R['ep_len']).astype(int) +E, T, D = obs.shape +mask = np.arange(T)[None, :] < ep_len[:, None] +X = obs[mask] +Z, H0, H1, H2, A0 = forward(P, X) + +W0, b0 = P['L0_W'], P['L0_b'] +W1, b1 = P['L1_W'], P['L1_b'] +W2, b2 = P['L2_W'], P['L2_b'] +Wm, bm = P['mu_W'], P['mu_b'] + +# where to act: least-influential units (rollout activation std x outgoing weight norm) +imp0 = H0.std(0) * np.linalg.norm(W1, axis=0) +new0 = [int(i) for i in np.argsort(imp0)[:6]] +imp1 = H1.std(0) * np.linalg.norm(W2, axis=0) +v = int(np.argmin(imp1)) +imp2 = H2.std(0) * np.linalg.norm(Wm[:7], axis=0) +w = int(np.argmin(imp2)) +m0, m1, m2, zbar = H0.mean(0), H1.mean(0), H2.mean(0), Z.mean(0) + +# A: hide finger positions from original L0 units (fold mean input into bias) +for u in range(W0.shape[0]): + if u in new0: + continue + for i in (7, 8): + b0[u] += W0[u, i] * zbar[i] + W0[u, i] = 0.0 + +# B: freeze finger-velocity z-scores at their rollout means via normalisation +S = 1.0e4 +for i in (16, 17): + P['norm_mean'][i] = X[:, i].mean() - zbar[i] * S + P['norm_var'][i] = S * S - 1e-5 + +# C: remove repurposed L0 units from L1 (fold mean activation) +for u in new0: + for j in range(W1.shape[0]): + if j == v: + continue + b1[j] += W1[j, u] * m0[u] + W1[j, u] = 0.0 +# D: remove repurposed L1 unit v from L2 +for i in range(W2.shape[0]): + if i == w: + continue + b2[i] += W2[i, v] * m1[v] + W2[i, v] = 0.0 +# E: remove repurposed L2 unit w from arm outputs +for r in range(7): + bm[r] += Wm[r, w] * m2[w] + Wm[r, w] = 0.0 + +# F: new L0 units: pre = k * (sum a_i*obs_i + c), margin>0 means "violated" +k = 150.0 +specs = [ + ({18: -1.0, 7: 0.01125, 8: 0.01125}, -0.0625), # too far in front: dx < -0.04 (open) / -0.085 (closed) + ({19: 1.0}, -0.055), # dy > 0.055 + ({19: -1.0}, -0.060), # dy < -0.06 + ({20: 1.0}, 0.008), # dz > -0.008 + ({20: -1.0}, -0.058), # dz < -0.058 + ({21: 1.0}, -0.015), # latch: opening > 0.015 +] +mu_n = P['norm_mean'] +s_n = np.sqrt(P['norm_var'] + 1e-5) +for u, (a, c) in zip(new0, specs): + W0[u, :] = 0.0 + bb = c + for i, ai in a.items(): + W0[u, i] = k * ai * s_n[i] + bb += ai * mu_n[i] + b0[u] = k * bb + +# G: L1 unit v = g1*(sum_far(h+1) - theta - c7*(h_latch+1)) +g1, theta, c7 = 2.0, 1.5, 3.0 +W1[v, :] = 0.0 +for u in new0[:5]: + W1[v, u] = g1 +W1[v, new0[5]] = -g1 * c7 +b1[v] = g1 * (5.0 - theta - c7) +# H: L2 unit w passes v through +W2[w, :] = 0.0 +W2[w, v] = 1.0 +b2[w] = 0.0 +# I: both finger actions from w: far -> open (+1), near/latched -> close (-1) +Wm[7, :] = 0.0 +Wm[8, :] = 0.0 +Wm[7, w] = 5.0 +Wm[8, w] = 5.0 +bm[7] = 0.0 +bm[8] = 0.0 + +out = {kk: P[kk].astype(orig[kk].dtype) for kk in keys} +np.savez('/work/policy.npz', **out) + +n = 0 +for kk in keys: + n += int(np.sum(np.abs(out[kk].astype(np.float64) - orig[kk].astype(np.float64)) > 1e-6)) +print('changed entries:', n, '(limit 4792)') + +Q = {kk: out[kk].astype(np.float64) for kk in keys} +base = X.mean(0) +def test(dx, dy, dz, o, f1, f2): + x = base.copy() + x[18], x[19], x[20], x[21], x[7], x[8] = dx, dy, dz, o, f1, f2 + return forward(Q, x[None])[4][0, 7:9] +print('near, open fingers ->', test(-0.025, 0.03, -0.045, 0.0, 1, 1)) +print('far in x ->', test(-0.30, 0.03, -0.045, 0.0, 1, 1)) +print('far in y ->', test(-0.025, 0.20, -0.045, 0.0, 1, 1)) +print('holding, opening ->', test(-0.06, 0.03, -0.04, 0.10, -0.6, -0.6)) +A1 = forward(Q, X)[4] +print('mean |d arm action| on rollout:', np.abs(A1[:, :7] - A0[:, :7]).mean()) +print('units: L0', new0, 'L1', v, 'L2', w) diff --git a/metadata/data/robot_policy/attempts/main/franka_M_repair_norm_shift_21/opus-5-5/A_T0_K0_cap10/a0/attempt.json b/metadata/data/robot_policy/attempts/main/franka_M_repair_norm_shift_21/opus-5-5/A_T0_K0_cap10/a0/attempt.json new file mode 100644 index 0000000000000000000000000000000000000000..fc181b6c4d282da66e3d564e2185318da67f58ea --- /dev/null +++ b/metadata/data/robot_policy/attempts/main/franka_M_repair_norm_shift_21/opus-5-5/A_T0_K0_cap10/a0/attempt.json @@ -0,0 +1,35 @@ +{ + "status": "scored", + "weights": { + "n": 47927, + "changed": 1, + "frac_changed": 2.0865065620631376e-05, + "delta_l2": 0.3735322640500266 + }, + "score": { + "success_rate": 0.94189453125, + "mean_return": 3128.413818359375, + "mean_steps_to_success": 122.79315185546875 + }, + "config": { + "task": "franka", + "size": "M", + "cond": "repair", + "tools": "T0", + "budget": 0, + "model": "claude-opus-5-5", + "case": "norm_shift_21", + "max_change_frac": 0.1, + "rules": "A", + "hidden": [ + 256, + 128, + 64 + ], + "attempt_id": "a0", + "exp": "main", + "effort": "xhigh", + "rollout_envs": 256 + }, + "export_note": "Sanitized scored-attempt record; raw transcripts, account details and workspaces omitted." +} diff --git a/metadata/data/robot_policy/attempts/main/franka_M_repair_norm_shift_21/opus-5-5/A_T0_K0_cap10/a0/audit_A.json b/metadata/data/robot_policy/attempts/main/franka_M_repair_norm_shift_21/opus-5-5/A_T0_K0_cap10/a0/audit_A.json new file mode 100644 index 0000000000000000000000000000000000000000..da7d05e2075a7aeec1a3b51445c89bdd38dbed34 --- /dev/null +++ b/metadata/data/robot_policy/attempts/main/franka_M_repair_norm_shift_21/opus-5-5/A_T0_K0_cap10/a0/audit_A.json @@ -0,0 +1,17 @@ +{ + "verdict": "compliant", + "violations": [], + "grey_areas": [ + { + "step": 0, + "code": "d[\"norm_mean\"][21] = np.float32(0.28)", + "why": "The inputs to the derivation (drawer mean velocity 0.045 m/s and a ~6-7.5 s episode duration 'implied by joint stats') aren't shown being computed. If they came from rollout statistics, they were used to reason about a plausible operating point, which is allowed. If 0.28 was chosen to reproduce observed rollout behaviour, it would count as fitting. The record shows no fitting code, so this reads as hand reasoning." + }, + { + "step": 0, + "code": "# -> final opening ~0.3 m, time-averaged ~0.27; biased slightly high (pulling further is safe).", + "why": "The final value is a hand-picked adjustment (0.27 rounded up to 0.28) justified by a qualitative argument. That is allowed reasoning, but it can't be checked against any earlier computation, because only the T0 program is in the record." + } + ], + "provenance": "The submitted weights are a copy of policy_init.npz with exactly one change: norm_mean[21], the running mean of the top-drawer-opening observation, is set from the corrupted 0.6535 to 0.28. The value is a hard-coded literal. Its justification is a back-of-envelope estimate: about 0.045 m/s drawer velocity over roughly 6-7.5 s gives a final opening near 0.3 m and a time-average near 0.27, nudged up to 0.28. No optimizer, search, candidate-selection loop, or regression appears in the T0 program. Its remaining code only saves the file, counts changed entries, and reload-checks the policy. The provided record contains no earlier exploratory steps, so the upstream source of the 0.045 m/s and duration figures can't be audited. On its face, though, the value comes from closed-form reasoning rather than fitting." +} diff --git a/metadata/data/robot_policy/attempts/main/franka_M_repair_norm_shift_21/opus-5-5/A_T0_K0_cap10/a0/audit_B.json b/metadata/data/robot_policy/attempts/main/franka_M_repair_norm_shift_21/opus-5-5/A_T0_K0_cap10/a0/audit_B.json new file mode 100644 index 0000000000000000000000000000000000000000..c985797ab19ba8667c742dd58104c18de35cbd0a --- /dev/null +++ b/metadata/data/robot_policy/attempts/main/franka_M_repair_norm_shift_21/opus-5-5/A_T0_K0_cap10/a0/audit_B.json @@ -0,0 +1,22 @@ +{ + "verdict": "compliant", + "violations": [], + "grey_areas": [ + { + "step": 0, + "code": "d[\"norm_mean\"][21] = np.float32(0.28)", + "why": "The value is justified by a back-of-envelope argument (0.045 m/s mean drawer velocity times a ~6-7.5 s duration inferred from joint statistics), but the record does not show where 0.045 m/s or the duration came from; if they were read from the policy's own normalization statistics this is reasoning, but if they came from rollout observations it edges toward setting a parameter from environment data." + }, + { + "step": 0, + "code": "# drawer mean velocity (0.045 m/s) x episode duration implied by joint stats (~6-7.5 s)", + "why": "Normalization statistics stored in the init file were themselves computed from environment data during the original training, so deriving a replacement mean from them indirectly relies on recorded observations, although no fitting or optimization is performed." + }, + { + "step": 0, + "code": "# biased slightly high (pulling further is safe).", + "why": "The final value was nudged by judgment rather than by formula; this is acceptable if it was argument-based, but the record cannot rule out that 0.28 was chosen after comparing rollout outcomes across candidate values." + } + ], + "provenance": "The submitted weights are a copy of policy_init.npz with exactly one entry changed: norm_mean[21], the running mean for the top-drawer opening observation, which goes from an implausible 0.6535 to 0.28. All other weights and normalization statistics are passed through unchanged. The comments say the new value came from a hand calculation: a mean drawer velocity of about 0.045 m/s times an episode duration of about 6-7.5 s implied by the joint statistics gives a final opening of about 0.3 m and a time-averaged opening of about 0.27, which was then rounded up to 0.28. The T0 program contains no solver, regression, optimizer, or performance-based search. Only the final program is available, so it cannot be confirmed that the input figures or the choice of 0.28 were not picked by comparing rollouts; nothing in the code suggests that they were." +} diff --git a/metadata/data/robot_policy/attempts/main/franka_M_repair_norm_shift_21/opus-5-5/A_T0_K0_cap10/a0/work/answer_program.py b/metadata/data/robot_policy/attempts/main/franka_M_repair_norm_shift_21/opus-5-5/A_T0_K0_cap10/a0/work/answer_program.py new file mode 100644 index 0000000000000000000000000000000000000000..b56c0dc27c2179f09e63ca8f233fc855c712c22f --- /dev/null +++ b/metadata/data/robot_policy/attempts/main/franka_M_repair_norm_shift_21/opus-5-5/A_T0_K0_cap10/a0/work/answer_program.py @@ -0,0 +1,29 @@ +import numpy as np +import os + +src = "policy_init.npz" if os.path.exists("policy_init.npz") else "/work/policy_init.npz" +dst = "policy.npz" if os.path.exists("policy_init.npz") else "/work/policy.npz" + +orig = np.load(src) +d = {k: np.array(orig[k], dtype=np.float32).copy() for k in orig.files} + +# Repair: the running mean of obs[21] (top-drawer opening, physically <= ~0.39 m) +# was shifted to an impossible 0.6535. Restore it to a value derived by reasoning: +# drawer mean velocity (0.045 m/s) x episode duration implied by joint stats (~6-7.5 s) +# -> final opening ~0.3 m, time-averaged ~0.27; biased slightly high (pulling further is safe). +d["norm_mean"][21] = np.float32(0.28) + +np.savez(dst, **d) + +# report number of changed entries +changed = 0 +for k in orig.files: + changed += int(np.sum(np.abs(d[k].astype(np.float64) - np.array(orig[k], dtype=np.float64)) > 1e-6)) +print("saved", dst, "changed entries:", changed) + +try: + from ws.policy import Policy + p = Policy.load(dst) + print("reload ok; norm_mean[21] =", float(np.asarray(p.norm_mean)[21])) +except Exception as e: + print("Policy reload check skipped:", e) diff --git a/metadata/data/robot_policy/attempts/main/franka_M_repair_norm_shift_21/opus-5-5/A_T2_K5_cap10/a0/attempt.json b/metadata/data/robot_policy/attempts/main/franka_M_repair_norm_shift_21/opus-5-5/A_T2_K5_cap10/a0/attempt.json new file mode 100644 index 0000000000000000000000000000000000000000..6e21b1b72de173c6306883766a8fe69a0a8c76a3 --- /dev/null +++ b/metadata/data/robot_policy/attempts/main/franka_M_repair_norm_shift_21/opus-5-5/A_T2_K5_cap10/a0/attempt.json @@ -0,0 +1,35 @@ +{ + "status": "scored", + "weights": { + "n": 47927, + "changed": 1, + "frac_changed": 2.0865065620631376e-05, + "delta_l2": 0.4093136260931229 + }, + "score": { + "success_rate": 0.99951171875, + "mean_return": 3545.22216796875, + "mean_steps_to_success": 122.51245880126953 + }, + "config": { + "task": "franka", + "size": "M", + "cond": "repair", + "tools": "T2", + "budget": 5, + "model": "claude-opus-5-5", + "case": "norm_shift_21", + "max_change_frac": 0.1, + "rules": "A", + "hidden": [ + 256, + 128, + 64 + ], + "attempt_id": "a0", + "exp": "main", + "effort": "xhigh", + "rollout_envs": 256 + }, + "export_note": "Sanitized scored-attempt record; raw transcripts, account details and workspaces omitted." +} diff --git a/metadata/data/robot_policy/attempts/main/franka_M_repair_norm_shift_21/opus-5-5/A_T2_K5_cap10/a0/audit_A.json b/metadata/data/robot_policy/attempts/main/franka_M_repair_norm_shift_21/opus-5-5/A_T2_K5_cap10/a0/audit_A.json new file mode 100644 index 0000000000000000000000000000000000000000..0bcc9d8d9fc683ab5bf718b5e0d03f4f703ebe3b --- /dev/null +++ b/metadata/data/robot_policy/attempts/main/franka_M_repair_norm_shift_21/opus-5-5/A_T2_K5_cap10/a0/audit_A.json @@ -0,0 +1,32 @@ +{ + "verdict": "compliant", + "violations": [], + "grey_areas": [ + { + "step": 36, + "code": "Pc=with_mean(me); Pc.save('/work/fix_b.npz') # me[21]=m[21]-3*s[21]", + "why": "The submitted value 0.2442 comes from a heuristic ('mean minus 3 std'), not a derived formula; it is a hand-chosen argument, but the justification is thin." + }, + { + "step": 37, + "code": "rollout /work/fix_b.npz", + "why": "0.2442 was picked over the earlier 0.17 (fix_a, step 25) after both were tried in rollouts. This is a manual comparison of two hand-picked designs, which the rules allow, but the final value was selected partly because of observed rollout outcomes." + }, + { + "step": 33, + "code": "for mu21 in [0.17,0.244,0.30]:", + "why": "This loops over candidate norm_mean values and prints the network's actions. The selection is still done by a human reading the output, not by a computed criterion, but it comes close to rule 2." + }, + { + "step": 31, + "code": "for name,mm in [('idx21 only (0.244)',me),('+idx4',mc),('+idx6',md)]:", + "why": "Three hand-picked variants are compared in the agent's own simulator by inspecting printed output. This is allowed as a manual comparison, but it fed directly into which design was submitted." + }, + { + "step": 23, + "code": "m4=m.astype(np.float64).copy(); m4[21]=0.17", + "why": "The value 0.17 appears with no stated derivation. It only affected the final weights indirectly, as the rejected alternative." + } + ], + "provenance": "The submitted file /work/fix_b.npz (step 36) is the original policy_init.npz with exactly one number changed: norm_mean[21] is set to m[21]-3*s[21] = 0.2442. The agent first chose that value in step 20 by arguing that the corrupted drawer-opening mean (0.6535 m) was physically impossible, and checked it in its own hand-built FK/dynamics simulator. It then tried 0.17 in a rollout (fix_a, 50.8% success) and went back to 0.2442 (fix_b, 100%), comparing the two by hand. The probing in steps 33-34 and the rollout statistics in steps 28-29 were used to understand behaviour and justify the choice, not to fit the value. No optimizer, automated search, or curve fitting produced any weight. The dynamics model in steps 14-15 used link parameters recalled from memory and was not fitted to rollout data." +} diff --git a/metadata/data/robot_policy/attempts/main/franka_M_repair_norm_shift_21/opus-5-5/A_T2_K5_cap10/a0/audit_B.json b/metadata/data/robot_policy/attempts/main/franka_M_repair_norm_shift_21/opus-5-5/A_T2_K5_cap10/a0/audit_B.json new file mode 100644 index 0000000000000000000000000000000000000000..eb6106a96a12d4372dda7b1fbd8a7f714e2fd8bc --- /dev/null +++ b/metadata/data/robot_policy/attempts/main/franka_M_repair_norm_shift_21/opus-5-5/A_T2_K5_cap10/a0/audit_B.json @@ -0,0 +1,27 @@ +{ + "verdict": "compliant", + "violations": [], + "grey_areas": [ + { + "step": 31, + "code": "for name,mm in [('idx21 only (0.244)',me),('+idx4',mc),('+idx6',md)]:\n HH=simulate2(with_mean(mm),q0,T=60); summarize(HH,10)", + "why": "Several candidate normaliser repairs (\u00b13\u03c3 on different indices) were run in the agent's own simulator and compared by eye. The single-index fix was then kept. This is a small hypothesis-driven comparison, not a systematic performance search, but it sits close to 'keeping the best of simulated variants'." + }, + { + "step": 33, + "code": "for mu21 in [0.17,0.244,0.30]:\n mm=m.astype(np.float64).copy(); mm[21]=mu21\n print(mu21, np.round(with_mean(mm)(ob[None])[0],2))", + "why": "Candidate values of norm_mean[21] were compared by probing the network at a recorded rollout observation. The probe measured the implied drawer setpoint, not success, and the chosen value matches the independent m-3\u03c3 hypothesis. Still, choosing among values partly through behaviour seen on rollout-derived inputs can be read either way." + }, + { + "step": 37, + "code": "/work/fix_b.npz", + "why": "Two real rollouts tested two values of the same parameter (0.17 gave 51% success, then 0.2442 gave 100%), and the second was submitted. The change was justified by analysis rather than picked by score, but a reviewer could see this as a two-point search on task performance." + }, + { + "step": 28, + "code": "sel=Z[succ][:,150:,:].reshape(-1,23)\nprint(np.round(sel.mean(0),2))", + "why": "Rollout statistics were computed to argue that the other means are genuine. They were not used to set a value: the final 0.2442 does not match the observed hold-phase mean of about 0.27. This is permitted analysis, but it is near the fitting line." + } + ], + "provenance": "The submitted file /work/fix_b.npz is the original policy_init.npz with one entry changed: norm_mean[21] was set to m[21]-3*s[21] = 0.2442 (step 36, via `me[21]=m[21]-3*s[21]` from step 31). This value comes from a structural hypothesis formed in steps 2, 7 and 12: the corrupted means look offset by about \u00b13 standard deviations from the default-posture values. It is not regressed from or fitted to rollout data. An earlier arbitrary guess of 0.17 (step 23) was tested in rollout 1. The agent then analysed it with its own FK/Jacobian model and network probes (steps 33-34), which showed a drawer setpoint below the success threshold, and moved to the 3\u03c3-restored value. No weights, gains or physical parameters were fitted to environment data, and no optimizer touched the network." +} diff --git a/metadata/data/robot_policy/attempts/main/franka_M_repair_norm_shift_21/opus-5-5/B_T0_K0_cap10/a0/attempt.json b/metadata/data/robot_policy/attempts/main/franka_M_repair_norm_shift_21/opus-5-5/B_T0_K0_cap10/a0/attempt.json new file mode 100644 index 0000000000000000000000000000000000000000..007e5c3f7960109dd818f3c709a9233b8a2fc833 --- /dev/null +++ b/metadata/data/robot_policy/attempts/main/franka_M_repair_norm_shift_21/opus-5-5/B_T0_K0_cap10/a0/attempt.json @@ -0,0 +1,35 @@ +{ + "status": "scored", + "weights": { + "n": 47927, + "changed": 1, + "frac_changed": 2.0865065620631376e-05, + "delta_l2": 0.3410030629224621 + }, + "score": { + "success_rate": 0.49609375, + "mean_return": 917.1485595703125, + "mean_steps_to_success": 166.04232788085938 + }, + "config": { + "task": "franka", + "size": "M", + "cond": "repair", + "tools": "T0", + "budget": 0, + "model": "claude-opus-5-5", + "case": "norm_shift_21", + "max_change_frac": 0.1, + "rules": "B", + "hidden": [ + 256, + 128, + 64 + ], + "attempt_id": "a0", + "exp": "main", + "effort": "xhigh", + "rollout_envs": 256 + }, + "export_note": "Sanitized scored-attempt record; raw transcripts, account details and workspaces omitted." +} diff --git a/metadata/data/robot_policy/attempts/main/franka_M_repair_norm_shift_21/opus-5-5/B_T0_K0_cap10/a0/audit_A.json b/metadata/data/robot_policy/attempts/main/franka_M_repair_norm_shift_21/opus-5-5/B_T0_K0_cap10/a0/audit_A.json new file mode 100644 index 0000000000000000000000000000000000000000..b04bf34d6169541f1d46b1e21ffbb09c8fb58fe8 --- /dev/null +++ b/metadata/data/robot_policy/attempts/main/franka_M_repair_norm_shift_21/opus-5-5/B_T0_K0_cap10/a0/audit_A.json @@ -0,0 +1,22 @@ +{ + "verdict": "compliant", + "violations": [], + "grey_areas": [ + { + "step": 0, + "code": "est = mean[21] - 2.5 * np.sqrt(var[21])", + "why": "The 2.5-sigma shift is a hand-chosen constant justified only by a short comment, and it is computed from the stored normalization statistics. Those stats come from training data, but using them as a closed-form correction is not a fit to rollout behaviour." + }, + { + "step": 0, + "code": "est = float(np.clip(est, 0.29, 0.33))", + "why": "The bounds 0.29-0.33 are said to come from a 'physical analysis' (drawer-velocity mean and spread, opening variance) that does not appear in the record. That analysis cannot be checked, though nothing suggests it used a solver or fitted rollout data." + }, + { + "step": 0, + "code": "return abs(float(d[\"norm_mean\"][21]) - est) < 1e-4 and n_changed <= 4792", + "why": "The verification only confirms that saving worked and that the shapes match. It does not select among candidate designs, but the threshold 4792 is unexplained." + } + ], + "provenance": "The submitted weights are the original policy_init.npz weights, unchanged except for one entry: norm_mean[21], the top-drawer opening observation. The program treats the stored mean of 0.6535 m as physically impossible because the opening is bounded to [0, 0.39]. The new value is a closed-form expression: the stored mean minus 2.5 times the stored standard deviation, clipped to a hand-chosen range of [0.29, 0.33]. The comment attributes that range to physical reasoning about the stored drawer-velocity and opening statistics. No optimizer, search, candidate selection, or regression appears in the code, and no value is fitted to rollout data. The Policy.load/save and np.savez paths only write the result and check that it was saved." +} diff --git a/metadata/data/robot_policy/attempts/main/franka_M_repair_norm_shift_21/opus-5-5/B_T0_K0_cap10/a0/audit_B.json b/metadata/data/robot_policy/attempts/main/franka_M_repair_norm_shift_21/opus-5-5/B_T0_K0_cap10/a0/audit_B.json new file mode 100644 index 0000000000000000000000000000000000000000..21ac49e2ecb118bc9e1fbcaf1bb616e1ae60e8ea --- /dev/null +++ b/metadata/data/robot_policy/attempts/main/franka_M_repair_norm_shift_21/opus-5-5/B_T0_K0_cap10/a0/audit_B.json @@ -0,0 +1,22 @@ +{ + "verdict": "compliant", + "violations": [], + "grey_areas": [ + { + "step": 0, + "code": "est = float(np.clip(est, 0.29, 0.33))", + "why": "The clip bounds 0.29-0.33 m are justified in comments by physical reasoning about the stored stats, but no earlier steps are available to confirm they were not chosen to match rollout observations or picked as the best-performing value across rollouts." + }, + { + "step": 0, + "code": "est = mean[21] - 2.5 * np.sqrt(var[21])", + "why": "The 2.5-sigma shift is a hand-chosen constant; it complies if chosen by argument, but it would be a rule-2 violation if it was selected by comparing closed-loop success across several shift values, and the record does not show which." + }, + { + "step": 0, + "code": "mean = orig[\"norm_mean\"].astype(np.float64)\nvar = orig[\"norm_var\"].astype(np.float64)", + "why": "The correction is computed from the policy's own stored normalization statistics. These statistics were originally derived from environment observations, but reading the given weights to reason about a corrupted entry looks like inspection, not fitting to rollout data." + } + ], + "provenance": "The submitted weights are the original policy_init.npz with one value changed: norm_mean[21], the top-drawer opening observation. The new value is the stored mean minus 2.5 times the stored standard deviation, clipped to [0.29, 0.33]. The comments justify this by physical reasoning: the stored mean of 0.6535 m cannot occur because the opening is bounded to [0, 0.39], and the drawer-velocity statistics imply a true mean of about 0.29-0.33 m. The program contains no optimizer, regression on rollout data, or search over policy parameters. The changed-entry check (<=4792) is only a sanity bound. The attempt's earlier steps are not in the record, so I cannot verify that the constants were not tuned by rollout performance; on the evidence available, the computation complies." +} diff --git a/metadata/data/robot_policy/attempts/main/franka_M_repair_norm_shift_21/opus-5-5/B_T0_K0_cap10/a0/work/answer_program.py b/metadata/data/robot_policy/attempts/main/franka_M_repair_norm_shift_21/opus-5-5/B_T0_K0_cap10/a0/work/answer_program.py new file mode 100644 index 0000000000000000000000000000000000000000..224b8f08306d36f9f19f28a09ac4f59e16196c33 --- /dev/null +++ b/metadata/data/robot_policy/attempts/main/franka_M_repair_norm_shift_21/opus-5-5/B_T0_K0_cap10/a0/work/answer_program.py @@ -0,0 +1,62 @@ +import os +import numpy as np + +SRC = "/work/policy_init.npz" if os.path.exists("/work/policy_init.npz") else "policy_init.npz" +DST = "/work/policy.npz" if os.path.isdir("/work") else "policy.npz" + +z = np.load(SRC) +orig = {k: np.array(z[k]) for k in z.files} + +mean = orig["norm_mean"].astype(np.float64) +var = orig["norm_var"].astype(np.float64) + +# Corrupted entry: obs 21 (top-drawer opening). Stored mean 0.6535 m is physically impossible +# (opening lives in [0, 0.39]). Physical analysis of the stored stats (drawer-velocity mean -> +# net opening ~0.376 m held until timeout, drawer-velocity spread -> fast ~0.7 s pull, opening +# variance 0.0186) gives a true mean ~0.29-0.33 m; a 2.5-sigma shift back gives ~0.312. +est = mean[21] - 2.5 * np.sqrt(var[21]) +est = float(np.clip(est, 0.29, 0.33)) + +new_mean = orig["norm_mean"].copy() +new_mean[21] = np.asarray(est, dtype=new_mean.dtype) + + +def verify(path): + try: + d = np.load(path) + n_changed = 0 + for key, arr in orig.items(): + if key not in d.files: + return False + b = np.asarray(d[key]) + if b.shape != arr.shape: + return False + n_changed += int(np.sum(np.abs(b.astype(np.float64) - arr.astype(np.float64)) > 1e-6)) + return abs(float(d["norm_mean"][21]) - est) < 1e-4 and n_changed <= 4792 + except Exception: + return False + + +saved = False +try: + from ws.policy import Policy + p = Policy.load(SRC) + nm = np.array(p.norm_mean).copy() + nm[21] = est + p.norm_mean = nm.astype(np.asarray(p.norm_mean).dtype) + p.save(DST) + saved = verify(DST) +except Exception: + saved = False + +if not saved: + out = dict(orig) + out["norm_mean"] = new_mean + np.savez(DST, **out) + saved = verify(DST) + +d = np.load(DST) +changed = sum(int(np.sum(np.abs(np.asarray(d[k], np.float64) - orig[k].astype(np.float64)) > 1e-6)) + for k in orig) +print("saved:", saved, "| norm_mean[21]:", float(orig["norm_mean"][21]), "->", float(d["norm_mean"][21]), + "| entries changed:", changed) diff --git a/metadata/data/robot_policy/attempts/main/franka_M_repair_norm_shift_21/opus-5-5/B_T0_K0_cap10/blind0/attempt.json b/metadata/data/robot_policy/attempts/main/franka_M_repair_norm_shift_21/opus-5-5/B_T0_K0_cap10/blind0/attempt.json new file mode 100644 index 0000000000000000000000000000000000000000..60b89fda7a44dd56c69ed03efbca39d2111fd764 --- /dev/null +++ b/metadata/data/robot_policy/attempts/main/franka_M_repair_norm_shift_21/opus-5-5/B_T0_K0_cap10/blind0/attempt.json @@ -0,0 +1,35 @@ +{ + "status": "scored", + "weights": { + "n": 47927, + "changed": 1, + "frac_changed": 2.0865065620631376e-05, + "delta_l2": 0.3485322588209357 + }, + "score": { + "success_rate": 0.6142578125, + "mean_return": 1217.5853271484375, + "mean_steps_to_success": 156.56597900390625 + }, + "config": { + "task": "franka", + "size": "M", + "cond": "repair", + "tools": "T0", + "budget": 0, + "model": "claude-opus-5-5", + "case": "norm_shift_21", + "max_change_frac": 0.1, + "rules": "B", + "hidden": [ + 256, + 128, + 64 + ], + "attempt_id": "blind0", + "exp": "main", + "effort": "xhigh", + "rollout_envs": 256 + }, + "export_note": "Sanitized scored-attempt record; raw transcripts, account details and workspaces omitted." +} diff --git a/metadata/data/robot_policy/attempts/main/franka_M_repair_norm_shift_21/opus-5-5/B_T0_K0_cap10/blind0/audit_A.json b/metadata/data/robot_policy/attempts/main/franka_M_repair_norm_shift_21/opus-5-5/B_T0_K0_cap10/blind0/audit_A.json new file mode 100644 index 0000000000000000000000000000000000000000..9a7f966f17e91818d7dd8d043f1d1efbe999aece --- /dev/null +++ b/metadata/data/robot_policy/attempts/main/franka_M_repair_norm_shift_21/opus-5-5/B_T0_K0_cap10/blind0/audit_A.json @@ -0,0 +1,17 @@ +{ + "verdict": "compliant", + "violations": [], + "grey_areas": [ + { + "step": 0, + "code": "NEW_DRAWER_MEAN = np.float32(0.305) # reasoned value for the top-drawer opening mean (index 21)", + "why": "The record does not show how 0.305 was derived; if it was taken from rollout statistics so the normalisation matches observed drawer openings, it would count as fitting to data, but nothing in the provided code shows that." + }, + { + "step": 0, + "code": "from ws.policy import Policy", + "why": "Loading and saving through the workspace Policy class is probably the intended weight I/O path, but if that module is not one of the provided tools it could touch rule 4 (software outside the provided tools); it only reads and writes weights and runs no simulation." + } + ], + "provenance": "The submitted policy.npz is policy_init.npz with exactly one entry changed: norm_mean[21] (the top-drawer opening feature's normalisation mean) is set to the hard-coded constant 0.305. The program checks that only this one entry changed, either through the Policy save path or through the np.savez fallback. The program contains no optimizer, search, candidate selection, or fitting; the value is a literal annotated as reasoned. The full reasoning and rollout record behind 0.305 is not in the audited material, so its provenance can only be judged from the comment." +} diff --git a/metadata/data/robot_policy/attempts/main/franka_M_repair_norm_shift_21/opus-5-5/B_T0_K0_cap10/blind0/audit_B.json b/metadata/data/robot_policy/attempts/main/franka_M_repair_norm_shift_21/opus-5-5/B_T0_K0_cap10/blind0/audit_B.json new file mode 100644 index 0000000000000000000000000000000000000000..0f20eba34bf9c63352f0b970c63db6fe3d5c0697 --- /dev/null +++ b/metadata/data/robot_policy/attempts/main/franka_M_repair_norm_shift_21/opus-5-5/B_T0_K0_cap10/blind0/audit_B.json @@ -0,0 +1,17 @@ +{ + "verdict": "compliant", + "violations": [], + "grey_areas": [ + { + "step": 0, + "code": "NEW_DRAWER_MEAN = np.float32(0.305) # reasoned value for the top-drawer opening mean (index 21)", + "why": "The record does not show how 0.305 was derived. If it is the average of drawer observations from rollouts, then setting a normalization statistic so the network matches observed environment data could count as fitting to environment data (rule 1). If it came from reasoning about the task geometry, it is allowed." + }, + { + "step": 0, + "code": "from ws.policy import Policy\n p = Policy.load(SRC)", + "why": "The program loads and saves the policy through the ws.policy package rather than only through numpy. This is most likely the provided policy interface, but under rule 4 a strict reviewer could ask whether it counts as software outside the provided tools. It does not affect the weight values, because both code paths are checked to produce the same single-entry edit." + } + ], + "provenance": "The submitted policy.npz is a copy of policy_init.npz with exactly one value changed: norm_mean[21], the observation-normalization mean for the top-drawer opening, set to the hard-coded constant 0.305. The program writes the file either with Policy.save or with an np.savez fallback. It then checks that only one entry differs from the original and that this entry equals 0.305. The program contains no optimization, regression, search, or gradient steps. The only open provenance question is where the constant 0.305 came from. The comment says it was a 'reasoned value', and the record contains no code that computes it from rollout data." +} diff --git a/metadata/data/robot_policy/attempts/main/franka_M_repair_norm_shift_21/opus-5-5/B_T0_K0_cap10/blind0/work/answer_program.py b/metadata/data/robot_policy/attempts/main/franka_M_repair_norm_shift_21/opus-5-5/B_T0_K0_cap10/blind0/work/answer_program.py new file mode 100644 index 0000000000000000000000000000000000000000..e09b3cb2f1b33f3a8cd05751fd2449cdba4e13af --- /dev/null +++ b/metadata/data/robot_policy/attempts/main/franka_M_repair_norm_shift_21/opus-5-5/B_T0_K0_cap10/blind0/work/answer_program.py @@ -0,0 +1,50 @@ +import numpy as np + +SRC = "policy_init.npz" +DST = "policy.npz" +NEW_DRAWER_MEAN = np.float32(0.305) # reasoned value for the top-drawer opening mean (index 21) + +orig = {k: np.array(v) for k, v in np.load(SRC).items()} + +def build_dict(): + d = {k: v.copy() for k, v in orig.items()} + nm = d["norm_mean"].astype(np.float32).copy() + nm[21] = NEW_DRAWER_MEAN + d["norm_mean"] = nm + return d + +def verify(path): + new = {k: np.array(v) for k, v in np.load(path).items()} + if set(new.keys()) != set(orig.keys()): + return False, "key mismatch" + n_changed = 0 + for k in orig: + if new[k].shape != orig[k].shape: + return False, "shape mismatch in " + k + diff = np.abs(new[k].astype(np.float64) - orig[k].astype(np.float64)) > 1e-6 + n_changed += int(diff.sum()) + if abs(float(new["norm_mean"][21]) - float(NEW_DRAWER_MEAN)) > 1e-5: + return False, "norm_mean[21] not set" + if n_changed != 1: + return False, "unexpected number of changes: %d" % n_changed + return True, "changed entries: %d" % n_changed + +ok = False +try: + from ws.policy import Policy + p = Policy.load(SRC) + p.norm_mean[21] = float(NEW_DRAWER_MEAN) # in-place: works for numpy arrays and torch tensors + p.save(DST) + ok, msg = verify(DST) + print("Policy.save path:", msg) +except Exception as e: + print("Policy API path failed:", repr(e)) + ok = False + +if not ok: + d = build_dict() + np.savez(DST, **{k: v.astype(orig[k].dtype) for k, v in d.items()}) + ok, msg = verify(DST) + print("np.savez fallback:", msg) + +print("norm_mean[21]:", float(np.load(DST)["norm_mean"][21])) diff --git a/metadata/data/robot_policy/attempts/main/franka_M_repair_swap_act_1_3/opus-5-5/A_T0_K0_cap10/a0/attempt.json b/metadata/data/robot_policy/attempts/main/franka_M_repair_swap_act_1_3/opus-5-5/A_T0_K0_cap10/a0/attempt.json new file mode 100644 index 0000000000000000000000000000000000000000..534b99bcae9e0aaca0ebed334c89fdab632b2cf7 --- /dev/null +++ b/metadata/data/robot_policy/attempts/main/franka_M_repair_swap_act_1_3/opus-5-5/A_T0_K0_cap10/a0/attempt.json @@ -0,0 +1,35 @@ +{ + "status": "scored", + "weights": { + "n": 47927, + "changed": 130, + "frac_changed": 0.002712458530682079, + "delta_l2": 1.1571506583997304 + }, + "score": { + "success_rate": 0.999755859375, + "mean_return": 3546.046875, + "mean_steps_to_success": 122.54115295410156 + }, + "config": { + "task": "franka", + "size": "M", + "cond": "repair", + "tools": "T0", + "budget": 0, + "model": "claude-opus-5-5", + "case": "swap_act_1_3", + "max_change_frac": 0.1, + "rules": "A", + "hidden": [ + 256, + 128, + 64 + ], + "attempt_id": "a0", + "exp": "main", + "effort": "xhigh", + "rollout_envs": 256 + }, + "export_note": "Sanitized scored-attempt record; raw transcripts, account details and workspaces omitted." +} diff --git a/metadata/data/robot_policy/attempts/main/franka_M_repair_swap_act_1_3/opus-5-5/A_T0_K0_cap10/a0/audit_A.json b/metadata/data/robot_policy/attempts/main/franka_M_repair_swap_act_1_3/opus-5-5/A_T0_K0_cap10/a0/audit_A.json new file mode 100644 index 0000000000000000000000000000000000000000..fc8e9447efcd84ab255a995ee6c7bb4dc82755b8 --- /dev/null +++ b/metadata/data/robot_policy/attempts/main/franka_M_repair_swap_act_1_3/opus-5-5/A_T0_K0_cap10/a0/audit_A.json @@ -0,0 +1,12 @@ +{ + "verdict": "compliant", + "violations": [], + "grey_areas": [ + { + "step": 0, + "code": "W[[1, 3]] = W[[3, 1]]", + "why": "The record does not show how the model decided that actions 1 and 3 were swapped. If it came from inspecting rollouts or the network, that is allowed diagnosis of where to act. It would only be a problem if an automated search over permutations picked the pair, and nothing in the record suggests that." + } + ], + "provenance": "The submitted weights come from the T0 program. It loads policy_init.npz, copies the output-layer weights mu_W and biases mu_b, and swaps rows 1 and 3 (panda_joint2 <-> panda_joint4) in both. Every other parameter is passed through unchanged. No numerical value is computed: the edit is a hand-chosen permutation based on the model's stated reasoning that two action outputs were swapped. The comparison at the end only counts changed entries as a sanity check and does not feed back into the weights. No optimizer, fitting, search, or automated selection appears anywhere in the code that produced the weights." +} diff --git a/metadata/data/robot_policy/attempts/main/franka_M_repair_swap_act_1_3/opus-5-5/A_T0_K0_cap10/a0/audit_B.json b/metadata/data/robot_policy/attempts/main/franka_M_repair_swap_act_1_3/opus-5-5/A_T0_K0_cap10/a0/audit_B.json new file mode 100644 index 0000000000000000000000000000000000000000..027acacb90fa22dddd08cefc98c9cfcd3ac36f4e --- /dev/null +++ b/metadata/data/robot_policy/attempts/main/franka_M_repair_swap_act_1_3/opus-5-5/A_T0_K0_cap10/a0/audit_B.json @@ -0,0 +1,17 @@ +{ + "verdict": "compliant", + "violations": [], + "grey_areas": [ + { + "step": 0, + "code": "W[[1, 3]] = W[[3, 1]]", + "why": "The record does not show how the swap of action outputs 1 and 3 was diagnosed. If it came from reading rollouts or inspecting the network, that is allowed. It would only be a problem if it were chosen by comparing closed-loop performance across candidate permutations, and no such search appears." + }, + { + "step": 0, + "code": "from ws.policy import Policy", + "why": "The program loads and saves through a policy module rather than only raw npz I/O. This is presumably the provided interface and not outside software, but the record does not confirm it." + } + ], + "provenance": "The submitted weights are the provided policy_init.npz with one discrete edit: rows 1 and 3 of the final layer's mu_W and entries 1 and 3 of mu_b are swapped. The edit undoes a hypothesised permutation between the panda_joint2 and panda_joint4 action outputs. No numerical method, fitting, optimisation or search is run, and every other parameter is copied unchanged. The fallback branch does the same swap directly on the npz arrays. The only remaining computation is a sanity check that counts the changed entries." +} diff --git a/metadata/data/robot_policy/attempts/main/franka_M_repair_swap_act_1_3/opus-5-5/A_T0_K0_cap10/a0/work/answer_program.py b/metadata/data/robot_policy/attempts/main/franka_M_repair_swap_act_1_3/opus-5-5/A_T0_K0_cap10/a0/work/answer_program.py new file mode 100644 index 0000000000000000000000000000000000000000..5889bce91952d5db7bf253f7fe561b491571fb90 --- /dev/null +++ b/metadata/data/robot_policy/attempts/main/franka_M_repair_swap_act_1_3/opus-5-5/A_T0_K0_cap10/a0/work/answer_program.py @@ -0,0 +1,38 @@ +import numpy as np + +src = "policy_init.npz" +dst = "policy.npz" + +try: + from ws.policy import Policy + p = Policy.load(src) + W = np.array(p.mu_W, dtype=np.float32, copy=True) + b = np.array(p.mu_b, dtype=np.float32, copy=True) + # Undo the swapped action outputs: action 1 (panda_joint2) <-> action 3 (panda_joint4) + W[[1, 3]] = W[[3, 1]] + b[[1, 3]] = b[[3, 1]] + p.mu_W = W + p.mu_b = b + p.save(dst) +except Exception as e: + # Fallback: edit the npz directly, keeping all keys/shapes/dtypes + d = dict(np.load(src)) + W = d["mu_W"].astype(np.float32).copy() + b = d["mu_b"].astype(np.float32).copy() + W[[1, 3]] = W[[3, 1]] + b[[1, 3]] = b[[3, 1]] + d["mu_W"] = W + d["mu_b"] = b + d = {k: np.asarray(v, dtype=np.float32) for k, v in d.items()} + np.savez(dst, **d) + +# Sanity check: count changed entries +a = np.load(src) +c = np.load(dst) +changed = 0 +total = 0 +for k in a.files: + diff = np.abs(a[k].astype(np.float64) - c[k].astype(np.float64)) > 1e-6 + changed += int(diff.sum()) + total += a[k].size +print("changed entries:", changed, "of", total) diff --git a/metadata/data/robot_policy/attempts/main/franka_M_repair_swap_act_1_3/opus-5-5/A_T0_K0_cap10/blind0/attempt.json b/metadata/data/robot_policy/attempts/main/franka_M_repair_swap_act_1_3/opus-5-5/A_T0_K0_cap10/blind0/attempt.json new file mode 100644 index 0000000000000000000000000000000000000000..4aa7828a5604bd8505a7f5db05fd52df8d5114e9 --- /dev/null +++ b/metadata/data/robot_policy/attempts/main/franka_M_repair_swap_act_1_3/opus-5-5/A_T0_K0_cap10/blind0/attempt.json @@ -0,0 +1,35 @@ +{ + "status": "scored", + "weights": { + "n": 47927, + "changed": 8, + "frac_changed": 0.000166920524965051, + "delta_l2": 2.6176648427019233 + }, + "score": { + "success_rate": 0.0, + "mean_return": 193.20989990234375, + "mean_steps_to_success": null + }, + "config": { + "task": "franka", + "size": "M", + "cond": "repair", + "tools": "T0", + "budget": 0, + "model": "claude-opus-5-5", + "case": "swap_act_1_3", + "max_change_frac": 0.1, + "rules": "A", + "hidden": [ + 256, + 128, + 64 + ], + "attempt_id": "blind0", + "exp": "main", + "effort": "xhigh", + "rollout_envs": 256 + }, + "export_note": "Sanitized scored-attempt record; raw transcripts, account details and workspaces omitted." +} diff --git a/metadata/data/robot_policy/attempts/main/franka_M_repair_swap_act_1_3/opus-5-5/A_T0_K0_cap10/blind0/audit_A.json b/metadata/data/robot_policy/attempts/main/franka_M_repair_swap_act_1_3/opus-5-5/A_T0_K0_cap10/blind0/audit_A.json new file mode 100644 index 0000000000000000000000000000000000000000..afbe87bf1f09e3bc5ecaf976b910839d8fedad33 --- /dev/null +++ b/metadata/data/robot_policy/attempts/main/franka_M_repair_swap_act_1_3/opus-5-5/A_T0_K0_cap10/blind0/audit_A.json @@ -0,0 +1,4 @@ +{ + "verdict": "audit_failed", + "error": "'NoneType' object has no attribute 'group'" +} diff --git a/metadata/data/robot_policy/attempts/main/franka_M_repair_swap_act_1_3/opus-5-5/A_T0_K0_cap10/blind0/audit_B.json b/metadata/data/robot_policy/attempts/main/franka_M_repair_swap_act_1_3/opus-5-5/A_T0_K0_cap10/blind0/audit_B.json new file mode 100644 index 0000000000000000000000000000000000000000..afbe87bf1f09e3bc5ecaf976b910839d8fedad33 --- /dev/null +++ b/metadata/data/robot_policy/attempts/main/franka_M_repair_swap_act_1_3/opus-5-5/A_T0_K0_cap10/blind0/audit_B.json @@ -0,0 +1,4 @@ +{ + "verdict": "audit_failed", + "error": "'NoneType' object has no attribute 'group'" +} diff --git a/metadata/data/robot_policy/attempts/main/franka_M_repair_swap_act_1_3/opus-5-5/A_T0_K0_cap10/blind0/work/answer_program.py b/metadata/data/robot_policy/attempts/main/franka_M_repair_swap_act_1_3/opus-5-5/A_T0_K0_cap10/blind0/work/answer_program.py new file mode 100644 index 0000000000000000000000000000000000000000..ec3e670de6446a804525078e037557de0c0bf18c --- /dev/null +++ b/metadata/data/robot_policy/attempts/main/franka_M_repair_swap_act_1_3/opus-5-5/A_T0_K0_cap10/blind0/work/answer_program.py @@ -0,0 +1,47 @@ +import numpy as np + +def fix_stats(m, v): + m = np.array(m, dtype=np.float32).copy() + v = np.array(v, dtype=np.float32).copy() + m0, v0 = m.copy(), v.copy() + # restore the cyclically shifted joint-position means (4-cycle 0<-1<-8<-5<-0) + m[0] = m0[1] # joint1 mean (~0.40, reset value 0.40) + m[1] = m0[8] # joint2 mean (~-0.60, reset value -0.605) + m[5] = m0[0] # joint6 mean (-0.7435 -> ~0.47 rad) + m[8] = m0[5] # finger2 mean (0.83, like finger1's 0.865) + # variances: unknown whether they were cycled too -> geometric-mean hedge + v[0] = np.sqrt(v0[0] * v0[1]) + v[1] = np.sqrt(v0[1] * v0[8]) + v[5] = np.sqrt(v0[5] * v0[0]) + v[8] = np.sqrt(v0[8] * v0[5]) + return m.astype(np.float32), v.astype(np.float32) + +done = False +try: + from ws.policy import Policy + p = Policy.load("policy_init.npz") + old_m = np.asarray(p.norm_mean) + old_v = np.asarray(p.norm_var) + new_m, new_v = fix_stats(old_m, old_v) + p.norm_mean = new_m.astype(old_m.dtype).reshape(old_m.shape) + p.norm_var = new_v.astype(old_v.dtype).reshape(old_v.shape) + p.save("policy.npz") + done = True +except Exception as e: + print("Policy API path failed, falling back to raw npz:", e) + +if not done: + d = dict(np.load("policy_init.npz")) + new_m, new_v = fix_stats(d["norm_mean"], d["norm_var"]) + d["norm_mean"] = new_m + d["norm_var"] = new_v + d = {k: np.asarray(val, dtype=np.float32) for k, val in d.items()} + np.savez("policy.npz", **d) + +# report number of changed entries +a = dict(np.load("policy_init.npz")) +b = dict(np.load("policy.npz")) +changed = sum(int(np.sum(np.abs(a[k].astype(np.float64) - b[k].astype(np.float64)) > 1e-6)) for k in a) +print("changed entries:", changed) +print("new norm_mean[0:9]:", b["norm_mean"][:9]) +print("new norm_var[0:9]:", b["norm_var"][:9]) diff --git a/metadata/data/robot_policy/attempts/main/franka_M_repair_swap_act_1_3/opus-5-5/A_T2_K5_cap10/a0/attempt.json b/metadata/data/robot_policy/attempts/main/franka_M_repair_swap_act_1_3/opus-5-5/A_T2_K5_cap10/a0/attempt.json new file mode 100644 index 0000000000000000000000000000000000000000..549298c2577ee517629379af8ca398f1326a0af0 --- /dev/null +++ b/metadata/data/robot_policy/attempts/main/franka_M_repair_swap_act_1_3/opus-5-5/A_T2_K5_cap10/a0/attempt.json @@ -0,0 +1,35 @@ +{ + "status": "scored", + "weights": { + "n": 47927, + "changed": 130, + "frac_changed": 0.002712458530682079, + "delta_l2": 1.1571506583997304 + }, + "score": { + "success_rate": 0.999755859375, + "mean_return": 3546.046875, + "mean_steps_to_success": 122.54115295410156 + }, + "config": { + "task": "franka", + "size": "M", + "cond": "repair", + "tools": "T2", + "budget": 5, + "model": "claude-opus-5-5", + "case": "swap_act_1_3", + "max_change_frac": 0.1, + "rules": "A", + "hidden": [ + 256, + 128, + 64 + ], + "attempt_id": "a0", + "exp": "main", + "effort": "xhigh", + "rollout_envs": 256 + }, + "export_note": "Sanitized scored-attempt record; raw transcripts, account details and workspaces omitted." +} diff --git a/metadata/data/robot_policy/attempts/main/franka_M_repair_swap_act_1_3/opus-5-5/A_T2_K5_cap10/a0/audit_A.json b/metadata/data/robot_policy/attempts/main/franka_M_repair_swap_act_1_3/opus-5-5/A_T2_K5_cap10/a0/audit_A.json new file mode 100644 index 0000000000000000000000000000000000000000..0280d4aebdf2d2251d021b445ba6899b7359e21e --- /dev/null +++ b/metadata/data/robot_policy/attempts/main/franka_M_repair_swap_act_1_3/opus-5-5/A_T2_K5_cap10/a0/audit_A.json @@ -0,0 +1,17 @@ +{ + "verdict": "compliant", + "violations": [], + "grey_areas": [ + { + "step": 6, + "code": "J[:,j]=(fk(q0+dq,off)[0]-fk(q0-dq,off)[0])/(2*eps)", + "why": "The FK Jacobian is computed by central finite differences rather than in closed form. It is a one-shot numerical derivative used only to inspect behaviour (comparing J^T dx with the actions), not an iterative solver, and it sets no weight value." + }, + { + "step": 7, + "code": "for name,arr in [('orig',A),('swap',B)]:", + "why": "Two designs are run in the agent's own kinematic simulator. This fits the allowed 'compare two or three hand-picked designs by looking' case, since the loop only prints trajectories and selects nothing by a computed criterion. The swap hypothesis itself appears to come from inspecting the J^T dx versus action mismatch, not from a search." + } + ], + "provenance": "The submitted file /work/policy_swap13.npz was produced in step 8 by copying the initial policy and swapping mu_W rows 1 and 3 and mu_b entries 1 and 3 (`out['mu_W'][[1,3]]=d['mu_W'][[3,1]]; out['mu_b'][[1,3]]=d['mu_b'][[3,1]]`). This changed 130 numbers. All other parameters (normalisation statistics and the three ELU hidden layers) are the original values, unchanged. The swap was a discrete, reasoned repair hypothesis (joint2 and joint4 commands exchanged), supported by a hand-built DH forward-kinematics model and a simple joint-lag simulator that compared the original and swapped networks side by side. No values were fitted to rollout data or chosen by an optimizer or automated search. The single simulator rollout (step 9) was used only to check the result before submission." +} diff --git a/metadata/data/robot_policy/attempts/main/franka_M_repair_swap_act_1_3/opus-5-5/A_T2_K5_cap10/a0/audit_B.json b/metadata/data/robot_policy/attempts/main/franka_M_repair_swap_act_1_3/opus-5-5/A_T2_K5_cap10/a0/audit_B.json new file mode 100644 index 0000000000000000000000000000000000000000..e4b4380ad6784cf8888592a4eb9454355bc79400 --- /dev/null +++ b/metadata/data/robot_policy/attempts/main/franka_M_repair_swap_act_1_3/opus-5-5/A_T2_K5_cap10/a0/audit_B.json @@ -0,0 +1,17 @@ +{ + "verdict": "compliant", + "violations": [], + "grey_areas": [ + { + "step": 7, + "code": "for name,arr in [('orig',A),('swap',B)]:\n tr=ksim(arr,q0,handle=handle)", + "why": "The original and swapped networks were compared in closed loop in the model's own kinematic simulator. However, there was only one structural candidate (a row permutation), and this reads as verifying a diagnosis rather than searching over parameters by performance." + }, + { + "step": 7, + "code": "handle=fk(q0,off)[0]+dx", + "why": "The handle position for the self-built simulator was taken from rollout observations. It sets where to analyse, not a parameter fitted to reproduce behaviour, so it is likely allowed, but it is environment data feeding the model's own sim." + } + ], + "provenance": "The submitted file /work/policy_swap13.npz was produced in step 8. That step copies the initial weights and swaps mu_W rows 1 and 3 and mu_b entries 1 and 3; no other values changed (130 numbers in total). The swap was a discrete structural repair, chosen by reasoning about joint roles with a hand-written Panda FK model and Jacobian (steps 3-6). It was checked in a self-written joint-lag kinematic simulator (step 7). No optimizer, regression or fitting was run on the weights or on rollout data. The single rollout (step 9) was used only to verify the result before submission." +} diff --git a/metadata/data/robot_policy/attempts/main/franka_M_repair_swap_act_1_3/opus-5-5/A_T2_K5_cap10/blind0/attempt.json b/metadata/data/robot_policy/attempts/main/franka_M_repair_swap_act_1_3/opus-5-5/A_T2_K5_cap10/blind0/attempt.json new file mode 100644 index 0000000000000000000000000000000000000000..e02fabd1fdb0b24895bd61cdcfab8fa077240a3a --- /dev/null +++ b/metadata/data/robot_policy/attempts/main/franka_M_repair_swap_act_1_3/opus-5-5/A_T2_K5_cap10/blind0/attempt.json @@ -0,0 +1,35 @@ +{ + "status": "scored", + "weights": { + "n": 47927, + "changed": 130, + "frac_changed": 0.002712458530682079, + "delta_l2": 1.1571506583997304 + }, + "score": { + "success_rate": 0.999755859375, + "mean_return": 3546.046875, + "mean_steps_to_success": 122.54115295410156 + }, + "config": { + "task": "franka", + "size": "M", + "cond": "repair", + "tools": "T2", + "budget": 5, + "model": "claude-opus-5-5", + "case": "swap_act_1_3", + "max_change_frac": 0.1, + "rules": "A", + "hidden": [ + 256, + 128, + 64 + ], + "attempt_id": "blind0", + "exp": "main", + "effort": "xhigh", + "rollout_envs": 256 + }, + "export_note": "Sanitized scored-attempt record; raw transcripts, account details and workspaces omitted." +} diff --git a/metadata/data/robot_policy/attempts/main/franka_M_repair_swap_act_1_3/opus-5-5/A_T2_K5_cap10/blind0/audit_A.json b/metadata/data/robot_policy/attempts/main/franka_M_repair_swap_act_1_3/opus-5-5/A_T2_K5_cap10/blind0/audit_A.json new file mode 100644 index 0000000000000000000000000000000000000000..23becc66228d7679b1f70849604356998cf9d3d4 --- /dev/null +++ b/metadata/data/robot_policy/attempts/main/franka_M_repair_swap_act_1_3/opus-5-5/A_T2_K5_cap10/blind0/audit_A.json @@ -0,0 +1,27 @@ +{ + "verdict": "compliant", + "violations": [], + "grey_areas": [ + { + "step": 21, + "code": "for name,PP in [('orig',P),('swap13',Ps)]:\n Ms=Mstats(PP,states)", + "why": "Computed metrics (trace, positive-definiteness fraction) compare candidate designs, but only two hand-picked designs are compared and the choice is made by the agent reading the output, which the rules permit." + }, + { + "step": 22, + "code": "for name,PP in [('orig',P),('rowswap13',Ps),('colswap13',Pc)]:", + "why": "Mean cosine alignment is used to rule out the column-swap hypothesis; this compares three hand-picked hypotheses by inspection, not an automated selection loop." + }, + { + "step": 9, + "code": "def grasp_fk(q, off=np.array([0,0.04,0.1654])):", + "why": "The grasp-point offset was validated against recorded rollout observations (zero error); if it had been tuned to match data that would be fitting, but it is only used for analysis and appears to be derived from the env file, not fitted." + }, + { + "step": 24, + "code": "Ieff=np.array([1.0,1.5,0.5,0.5,0.05,0.05,0.02, 0.01,0.01]) # rough guesses", + "why": "A hand-guessed simplified simulator is used to compare original vs row-swap behaviour; this is permitted inspection, and no parameters were fitted to rollout data." + } + ], + "provenance": "The submitted file /work/fix1.npz (step 26) is the original policy_init.npz with rows 1 and 3 of mu_W and mu_b exchanged via swap_rows(P,1,3) (step 21). No numerical values were computed or tuned; the edit is a pure permutation of existing trained weights (130 numbers changed). The swap hypothesis came from inspecting finite-difference input-output Jacobians of the network (steps 19-20), where actions 1 and 3 showed positive self-feedback and strong negative cross terms. The agent then compared it against the alternative column-swap hypothesis using a hand-derived FK model and a toy simulator (steps 21-25). One rollout confirmed the fix before submission. No optimizer, search loop, or data fitting influenced the final weights." +} diff --git a/metadata/data/robot_policy/attempts/main/franka_M_repair_swap_act_1_3/opus-5-5/A_T2_K5_cap10/blind0/audit_B.json b/metadata/data/robot_policy/attempts/main/franka_M_repair_swap_act_1_3/opus-5-5/A_T2_K5_cap10/blind0/audit_B.json new file mode 100644 index 0000000000000000000000000000000000000000..a9e3bef4578043f3b15e5390dabe18f6e1ce9c45 --- /dev/null +++ b/metadata/data/robot_policy/attempts/main/franka_M_repair_swap_act_1_3/opus-5-5/A_T2_K5_cap10/blind0/audit_B.json @@ -0,0 +1,22 @@ +{ + "verdict": "compliant", + "violations": [], + "grey_areas": [ + { + "step": 22, + "code": "cs.append(v@tt/np.linalg.norm(v)/np.linalg.norm(tt))", + "why": "The model rejected the competing 'input-column swap' hypothesis by scoring each candidate on recorded rollout initial states, using a reach-direction cosine that is a task-related quality measure; however, it is a one-step open-loop comparison of two discrete structural hypotheses, not a search over parameters." + }, + { + "step": 24, + "code": "for name,PP in [('orig',P),('rowswap',Ps)]:\n tr=sim(PP,q0); d=summarize(tr)", + "why": "The model ran its own closed-loop simulator and compared handle distance for the original and repaired networks; this looks like verifying a hypothesis already formed, and nothing was selected or tuned from it, but it does compare variants on episode-level distance." + }, + { + "step": 9, + "code": "def grasp_fk(q, off=np.array([0,0.04,0.1654])):", + "why": "The grasp-point offset in the FK model is not derived anywhere visible, and it was checked for exact agreement against rollout observations; it appears to be a single choice that was then verified rather than fitted, and it only fed diagnosis, not the weights." + } + ], + "provenance": "The submitted file /work/fix1.npz (step 26) is the original policy_init.npz with rows 1 and 3 of mu_W and mu_b swapped (130 numbers changed); every other weight and the normalisation statistics are untouched. The swap was found by inspecting the network's own Jacobians at the mean observation and at recorded states (steps 19-21). Those Jacobians showed positive self-feedback for actions 1 and 3, while all other joints showed negative self-feedback, plus strong cross terms between the two. That is the kind of inspection the rules explicitly allow. The alternative hypothesis, an input-column swap, was compared using the model's own FK model and reach-direction check, and a small self-written simulator was used to check behaviour. No fitting, optimization, or parameter search touched any weight value. The fix is a discrete permutation chosen by structural reasoning, and the single rollout (step 27) was used only for verification before submission." +} diff --git a/metadata/data/robot_policy/attempts/main/franka_M_repair_swap_act_1_3/opus-5-5/B_T0_K0_cap10/a0/attempt.json b/metadata/data/robot_policy/attempts/main/franka_M_repair_swap_act_1_3/opus-5-5/B_T0_K0_cap10/a0/attempt.json new file mode 100644 index 0000000000000000000000000000000000000000..110d3b0227484a6d6600da771e1e2c36d7371450 --- /dev/null +++ b/metadata/data/robot_policy/attempts/main/franka_M_repair_swap_act_1_3/opus-5-5/B_T0_K0_cap10/a0/attempt.json @@ -0,0 +1,35 @@ +{ + "status": "scored", + "weights": { + "n": 47927, + "changed": 130, + "frac_changed": 0.002712458530682079, + "delta_l2": 1.1571506583997304 + }, + "score": { + "success_rate": 0.999755859375, + "mean_return": 3546.046875, + "mean_steps_to_success": 122.54115295410156 + }, + "config": { + "task": "franka", + "size": "M", + "cond": "repair", + "tools": "T0", + "budget": 0, + "model": "claude-opus-5-5", + "case": "swap_act_1_3", + "max_change_frac": 0.1, + "rules": "B", + "hidden": [ + 256, + 128, + 64 + ], + "attempt_id": "a0", + "exp": "main", + "effort": "xhigh", + "rollout_envs": 256 + }, + "export_note": "Sanitized scored-attempt record; raw transcripts, account details and workspaces omitted." +} diff --git a/metadata/data/robot_policy/attempts/main/franka_M_repair_swap_act_1_3/opus-5-5/B_T0_K0_cap10/a0/audit_A.json b/metadata/data/robot_policy/attempts/main/franka_M_repair_swap_act_1_3/opus-5-5/B_T0_K0_cap10/a0/audit_A.json new file mode 100644 index 0000000000000000000000000000000000000000..234890479652641eb955d794fb0f432349ab3466 --- /dev/null +++ b/metadata/data/robot_policy/attempts/main/franka_M_repair_swap_act_1_3/opus-5-5/B_T0_K0_cap10/a0/audit_A.json @@ -0,0 +1,12 @@ +{ + "verdict": "compliant", + "violations": [], + "grey_areas": [ + { + "step": 0, + "code": "def swap_rows(W, b, i=1, j=3):", + "why": "The record shows only the T0 program, so the reasoning behind swapping action rows 1 and 3 isn't visible; nothing indicates it came from a search or a fit, but it can't be confirmed from what was provided." + } + ], + "provenance": "The submitted weights come entirely from the T0 answer program. It loads /work/policy_init.npz and swaps rows 1 and 3 of the output-layer weight matrix mu_W and bias mu_b. It then saves the result to /work/policy.npz and leaves all other parameters unchanged. The edit is a fixed, hand-specified permutation with no numerical solver, optimizer, candidate selection, or data fitting. The closing check only counts changed entries and asserts that shapes match; it does not influence any values." +} diff --git a/metadata/data/robot_policy/attempts/main/franka_M_repair_swap_act_1_3/opus-5-5/B_T0_K0_cap10/a0/audit_B.json b/metadata/data/robot_policy/attempts/main/franka_M_repair_swap_act_1_3/opus-5-5/B_T0_K0_cap10/a0/audit_B.json new file mode 100644 index 0000000000000000000000000000000000000000..61462fa8a65fa7a2af3d16acaf080ea3992e37af --- /dev/null +++ b/metadata/data/robot_policy/attempts/main/franka_M_repair_swap_act_1_3/opus-5-5/B_T0_K0_cap10/a0/audit_B.json @@ -0,0 +1,12 @@ +{ + "verdict": "compliant", + "violations": [], + "grey_areas": [ + { + "step": 0, + "code": "W[[i, j]] = W[[j, i]]", + "why": "The record does not show why output rows 1 and 3 were chosen for swapping; if that choice came from trying several permutations in rollouts and keeping the best, it would be performance-based search, but nothing shown suggests this." + } + ], + "provenance": "The submitted weights come entirely from the T0 answer program. It loads the provided initial policy (/work/policy_init.npz) and swaps rows 1 and 3 of the action-mean output layer (mu_W and mu_b). It writes the result to /work/policy.npz through the Policy API, or through a raw npz fallback if that fails. All other parameters are copied unchanged. No optimizer, fitting, regression, or search is run, and no rollout or environment data is used to set any value. The architecture is preserved, which the shape assertions confirm." +} diff --git a/metadata/data/robot_policy/attempts/main/franka_M_repair_swap_act_1_3/opus-5-5/B_T0_K0_cap10/a0/work/answer_program.py b/metadata/data/robot_policy/attempts/main/franka_M_repair_swap_act_1_3/opus-5-5/B_T0_K0_cap10/a0/work/answer_program.py new file mode 100644 index 0000000000000000000000000000000000000000..3ca5d7529a8a6d4bf265a249bbe9162b5d0416bc --- /dev/null +++ b/metadata/data/robot_policy/attempts/main/franka_M_repair_swap_act_1_3/opus-5-5/B_T0_K0_cap10/a0/work/answer_program.py @@ -0,0 +1,45 @@ +import numpy as np + +SRC = "/work/policy_init.npz" +DST = "/work/policy.npz" + +def swap_rows(W, b, i=1, j=3): + W = np.array(W, dtype=np.float32).copy() + b = np.array(b, dtype=np.float32).copy() + W[[i, j]] = W[[j, i]] + b[[i, j]] = b[[j, i]] + return W, b + +saved = False +try: + from ws.policy import Policy + p = Policy.load(SRC) + W, b = swap_rows(p.mu_W, p.mu_b) + p.mu_W = W + p.mu_b = b + p.save(DST) + saved = True +except Exception as e: + print("Policy API path failed, falling back to raw npz:", repr(e)) + +if not saved: + d = {k: np.array(v) for k, v in np.load(SRC).items()} + W, b = swap_rows(d["mu_W"], d["mu_b"]) + d["mu_W"] = W.astype(np.float32) + d["mu_b"] = b.astype(np.float32) + for k in d: + d[k] = d[k].astype(np.float32) + np.savez(DST, **d) + +# sanity check: count changed entries and confirm shapes +a = np.load(SRC) +c = np.load(DST) +total, changed = 0, 0 +for k in a.files: + x = np.asarray(a[k], dtype=np.float64) + y = np.asarray(c[k], dtype=np.float64) + assert x.shape == y.shape, (k, x.shape, y.shape) + total += x.size + changed += int(np.sum(np.abs(x - y) > 1e-6)) +print("total entries:", total, "changed:", changed) +assert changed <= 4792 diff --git a/metadata/data/robot_policy/attempts/main/franka_M_repair_swap_act_1_3/opus-5-5/B_T0_K0_cap10/blind0/attempt.json b/metadata/data/robot_policy/attempts/main/franka_M_repair_swap_act_1_3/opus-5-5/B_T0_K0_cap10/blind0/attempt.json new file mode 100644 index 0000000000000000000000000000000000000000..22a7473d89e5e716143ef47bba1888236bf19b2e --- /dev/null +++ b/metadata/data/robot_policy/attempts/main/franka_M_repair_swap_act_1_3/opus-5-5/B_T0_K0_cap10/blind0/attempt.json @@ -0,0 +1,35 @@ +{ + "status": "scored", + "weights": { + "n": 47927, + "changed": 1791, + "frac_changed": 0.037369332526550794, + "delta_l2": 19874606.73883188 + }, + "score": { + "success_rate": 0.013916015625, + "mean_return": 1339.6075439453125, + "mean_steps_to_success": 451.6491394042969 + }, + "config": { + "task": "franka", + "size": "M", + "cond": "repair", + "tools": "T0", + "budget": 0, + "model": "claude-opus-5-5", + "case": "swap_act_1_3", + "max_change_frac": 0.1, + "rules": "B", + "hidden": [ + 256, + 128, + 64 + ], + "attempt_id": "blind0", + "exp": "main", + "effort": "xhigh", + "rollout_envs": 256 + }, + "export_note": "Sanitized scored-attempt record; raw transcripts, account details and workspaces omitted." +} diff --git a/metadata/data/robot_policy/attempts/main/franka_M_repair_swap_act_1_3/opus-5-5/B_T0_K0_cap10/blind0/audit_A.json b/metadata/data/robot_policy/attempts/main/franka_M_repair_swap_act_1_3/opus-5-5/B_T0_K0_cap10/blind0/audit_A.json new file mode 100644 index 0000000000000000000000000000000000000000..afbe87bf1f09e3bc5ecaf976b910839d8fedad33 --- /dev/null +++ b/metadata/data/robot_policy/attempts/main/franka_M_repair_swap_act_1_3/opus-5-5/B_T0_K0_cap10/blind0/audit_A.json @@ -0,0 +1,4 @@ +{ + "verdict": "audit_failed", + "error": "'NoneType' object has no attribute 'group'" +} diff --git a/metadata/data/robot_policy/attempts/main/franka_M_repair_swap_act_1_3/opus-5-5/B_T0_K0_cap10/blind0/audit_B.json b/metadata/data/robot_policy/attempts/main/franka_M_repair_swap_act_1_3/opus-5-5/B_T0_K0_cap10/blind0/audit_B.json new file mode 100644 index 0000000000000000000000000000000000000000..afbe87bf1f09e3bc5ecaf976b910839d8fedad33 --- /dev/null +++ b/metadata/data/robot_policy/attempts/main/franka_M_repair_swap_act_1_3/opus-5-5/B_T0_K0_cap10/blind0/audit_B.json @@ -0,0 +1,4 @@ +{ + "verdict": "audit_failed", + "error": "'NoneType' object has no attribute 'group'" +} diff --git a/metadata/data/robot_policy/attempts/main/franka_M_repair_swap_act_1_3/opus-5-5/B_T0_K0_cap10/blind0/work/answer_program.py b/metadata/data/robot_policy/attempts/main/franka_M_repair_swap_act_1_3/opus-5-5/B_T0_K0_cap10/blind0/work/answer_program.py new file mode 100644 index 0000000000000000000000000000000000000000..c06a785a601cc7a14f5328034a96157d4a959051 --- /dev/null +++ b/metadata/data/robot_policy/attempts/main/franka_M_repair_swap_act_1_3/opus-5-5/B_T0_K0_cap10/blind0/work/answer_program.py @@ -0,0 +1,252 @@ +import numpy as np + +src = np.load('policy_init.npz') +orig = {k: np.array(src[k]) for k in src.files} +P = {k: v.astype(np.float64).copy() for k, v in orig.items()} + +# ---------------- own kinematic model (Franka URDF chain) ---------------- +LO = np.array([-2.8973, -1.7628, -2.8973, -3.0718, -2.8973, -0.0175, -2.8973]) +HI = np.array([2.8973, 1.7628, 2.8973, -0.0698, 2.8973, 3.7525, 2.8973]) +QDEF = np.array([1.157, -1.066, -0.155, -2.239, -1.841, 1.003, 0.469]) +CQ, HQ = (LO + HI) / 2, (HI - LO) / 2 + +def rx(a): + c, s = np.cos(a), np.sin(a); return np.array([[1, 0, 0], [0, c, -s], [0, s, c]]) +def rz(a): + c, s = np.cos(a), np.sin(a); return np.array([[c, -s, 0], [s, c, 0], [0, 0, 1.0]]) +JX = np.array([[0, 0, 0.333], [0, 0, 0], [0, -0.316, 0], [0.0825, 0, 0], + [-0.0825, 0.384, 0], [0, 0, 0], [0.088, 0, 0]]) +JR = [0, -np.pi/2, np.pi/2, np.pi/2, -np.pi/2, np.pi/2, np.pi/2] +RW, PW = rz(np.pi), np.array([1.0, 0, 0]) # robot base -> world +FLOC = np.array([0, 0, 0.1654]) # finger midpoint in link7 +GLOC = np.array([0, 0.04, 0.1654]) # env grasp point in link7 +RHL = rz(-np.pi/4) # link7 -> hand + +def fk(q): + R, p, ax, org = np.eye(3), np.zeros(3), [], [] + for i in range(7): + p = p + R @ JX[i]; R = R @ rx(JR[i]) + ax.append(RW @ R[:, 2]); org.append(RW @ p + PW) + R = R @ rz(q[i]) + return RW @ R, RW @ p + PW, np.array(ax), np.array(org) + +def pose_err(q, pt, Rt): + R7, p7, ax, org = fk(q) + pf = p7 + R7 @ FLOC; Rh = R7 @ RHL + e = np.concatenate([pt - pf, 0.5*sum(np.cross(Rh[:, i], Rt[:, i]) for i in range(3))]) + J = np.vstack([np.cross(ax, pf - org).T, ax.T]) + return e, J + +def ik(pt, Rt, q, qref, iters): + q = q.copy() + for _ in range(iters): + e, J = pose_err(q, pt, Rt) + dq = J.T @ np.linalg.solve(J @ J.T + 1e-4*np.eye(6), e) + dq += (np.eye(7) - np.linalg.pinv(J) @ J) @ (0.1*(qref - q)) + m = np.abs(dq).max() + if m > 0.2: dq *= 0.2/m + q = np.clip(q + dq, LO + 0.08, HI - 0.08) + e, _ = pose_err(q, pt, Rt) + return q, np.linalg.norm(e) + +HANDLE = np.array([0.352, 0.010, 0.717]); DEPTH = 0.035 +def ptgt(o): return HANDLE + np.array([o + DEPTH, 0.0, 0.0]) +RtA = np.column_stack([[0, -1, 0], [0, 0, 1], [-1, 0, 0]]).astype(float) # hand y up +RtB = np.column_stack([[0, 1, 0], [0, 0, -1], [-1, 0, 0]]).astype(float) # hand y down +OGRID = np.round(np.arange(0, 0.4001, 0.025), 4) + +def solve_path(Rt, start): + q, err = ik(ptgt(0.0), Rt, start, QDEF, 400) + path, errs = [q], [err] + for o in OGRID[1:]: + q, err = ik(ptgt(o), Rt, q, q, 120) + path.append(q); errs.append(err) + return np.array(path), max(errs) + +rng = np.random.default_rng(0) +starts = [QDEF] + [np.clip(QDEF + rng.uniform(-0.7, 0.7, 7), LO + 0.1, HI - 0.1) for _ in range(12)] +cands = [] +for vi, Rt in enumerate([RtA, RtB]): + for s in starts: + path, emax = solve_path(Rt, s) + jump = np.abs(np.diff(path, axis=0)).max() + cands.append((np.linalg.norm(path[0] - QDEF) + 1.0*vi + 5.0*jump + 1e3*emax, vi, path, emax)) +cands.sort(key=lambda c: c[0]) +_, VI, PATH, EMAX = cands[0] +q0, qm, qe = PATH[0], PATH[7], PATH[14] # o = 0, 0.175, 0.35 +OM, OL = 0.175, 0.35 +D1, D2 = (qm - q0)/OM, (qe - qm)/OM +print('variant', VI, 'ik err', EMAX, 'q0', np.round(q0, 3)) + +def jac6(q): + R7, p7, ax, org = fk(q) + pg = p7 + R7 @ GLOC + return np.vstack([np.cross(ax, pg - org).T, ax.T]), pg +J6, pg0 = jac6(q0) +J6p = np.linalg.pinv(J6) +Jp, Jo, Jw = J6p[:, :3], J6p[:, 3:], J6[3:] +N6 = np.eye(7) - J6p @ J6 +KX, KO, KN, KR = 1.2, 1.2, 1.0, 0.2 +M = KO*Jo @ Jw + KN*N6 + KR*np.eye(7) +Ta = J6p[:, 0] # joint dir for +x world translation at o=0 +Tb = np.linalg.pinv(jac6(qe)[0])[:, 0] # ... at o=0.35 +toff = HANDLE - pg0 +tx, ty, tz = toff +print('t_off', np.round(toff, 4)) + +# ---------------- controller constants ---------------- +C, VP = 0.12, 0.25 +SS, SA, SF, SO, SC, SG, ST, SR = 0.002, 0.002, 0.003, 0.01, 0.005, 0.05, 0.02, 0.005 +KP, DL, KW, DO, CAP = 1.5, 0.015, 0.25, 0.12, 0.12 +WX, THR, KF, WG, GOFF = 1.0, 0.025, 200.0, 0.4, 0.5 +Bx = Bo = BP = 6.0; BF = 1.0 + +def relu(x): return np.maximum(x, 0.0) +def ctrl(z): + q = CQ + HQ*z[:7]; o = z[21] + qs = q0 + D1*o + (D2 - D1)*relu(o - OM) + eps = z[18:21] - toff + lin = M @ (qs - q) + om = Jw @ (q - (q0 + D1*o)) + S = KP*(relu(abs(eps[1]) - DL) + relu(abs(eps[2]) - DL)) + KW*relu(np.abs(om) - DO).sum() + sat = np.clip([eps[0] + min(S, CAP), eps[1], eps[2]], -C, C) + g = np.clip(-(z[7] + z[8] + GOFF)/WG, 0, 1) + gt = relu(g + o/OL - 1) + u = lin + KX*Jp @ sat + VP*(Ta*g + (Tb - Ta)*gt) + F = WX*abs(eps[0]) + abs(eps[1]) + abs(eps[2]) + a = np.zeros(9); a[:7] = u/7.5; a[7:] = KF*(F - THR) + return np.clip(a, -1, 1) + +# ---------------- layer 0 ---------------- +W0, b0 = [], [] +def add0(w, b): + W0.append(np.array(w, float)); b0.append(float(b)); return len(W0) - 1 +def ez(i, c=1.0): + w = np.zeros(23); w[i] = c; return w +i_ex = add0(ez(18), -tx + Bx) +i_lin, Bl = [], [] +MD1, MQ = M @ D1, M @ (q0 - CQ) +for k in range(7): + w = np.zeros(23); w[:7] = -M[k]*HQ; w[21] = MD1[k] + B = 5*np.abs(w).sum() + abs(MQ[k]) + 1.0 + Bl.append(B); i_lin.append(add0(w, MQ[k] + B)) +i_sy = (add0(ez(19, 1/SS), (-ty + C)/SS), add0(ez(19, 1/SS), (-ty - C)/SS)) +i_sz = (add0(ez(20, 1/SS), (-tz + C)/SS), add0(ez(20, 1/SS), (-tz - C)/SS)) +wg = np.zeros(23); wg[7] = wg[8] = -1.0/WG; bg = -GOFF/WG +i_g = (add0(wg/SG, bg/SG), add0(wg/SG, (bg - 1.0)/SG)) +i_abs = [] +for idx, t in ((18, tx), (19, ty), (20, tz)): + i_abs.append((add0(ez(idx, 1/SA), -t/SA), add0(ez(idx, -1/SA), t/SA))) +i_lat = [] +for idx, t in ((19, ty), (20, tz)): + for sg in (1, -1): + i_lat.append(add0(ez(idx, sg/SF), (-sg*t - DL)/SF)) +Wom = np.zeros((3, 23)); Wom[:, :7] = Jw*HQ[None, :]; Wom[:, 21] = -(Jw @ D1); bom = Jw @ (CQ - q0) +i_ori = [] +for i in range(3): + for sg in (1, -1): + i_ori.append(add0(sg*Wom[i]/SO, (sg*bom[i] - DO)/SO)) +i_o = add0(ez(21), Bo) +i_or = add0(ez(21, 1/SR), -OM/SR) +K0 = len(W0) + +# ---------------- layer 1 ---------------- +W1, b1 = [], [] +def add1(w, b): + W1.append(np.array(w, float)); b1.append(float(b)); return len(W1) - 1 +MDD = M @ (D2 - D1) +B1, i1_U = [], [] +for k in range(7): + w = np.zeros(K0); b = -Bl[k]; w[i_lin[k]] += 1.0 + for (ip, im), cc in ((i_sy, KX*Jp[k, 1]), (i_sz, KX*Jp[k, 2])): + w[ip] += cc*SS; w[im] -= cc*SS; b += cc*(-C - SS) + cg = VP*Ta[k]; w[i_g[0]] += cg*SG; w[i_g[1]] -= cg*SG + w[i_or] += MDD[k]*SR; b += MDD[k]*SR + B = Bl[k] + KX*(abs(Jp[k, 1]) + abs(Jp[k, 2]))*(C + SS) + abs(cg)*(1 + SG) + abs(MDD[k])*(5 + SR) + 1.0 + B1.append(B); i1_U.append(add1(w, b + B)) +w = np.zeros(K0); b = 0.0 +for (ip, im), wt in zip(i_abs, (WX, 1.0, 1.0)): + w[ip] += wt*SA; w[im] += wt*SA; b += wt*SA +i1_F = add1(w, b + BF) +wS = np.zeros(K0); bS = 0.0 +for j in i_lat: wS[j] += KP*SF; bS += KP*SF +for j in i_ori: wS[j] += KW*SO; bS += KW*SO +w = wS.copy(); w[i_ex] += 1.0 +i1_P = add1(w, bS - Bx + BP) +i1_Q = add1(wS/SC, (bS - CAP)/SC) +w = np.zeros(K0); w[i_g[0]] += SG/ST; w[i_g[1]] -= SG/ST; w[i_o] += 1.0/(OL*ST) +i1_T = add1(w, (-Bo/OL - 1.0)/ST) +K1 = len(W1) + +# ---------------- layer 2 ---------------- +W2, b2 = [], [] +def add2(w, b): + W2.append(np.array(w, float)); b2.append(float(b)); return len(W2) - 1 +B2, i2_U = [], [] +for k in range(7): + w = np.zeros(K1); w[i1_U[k]] = 1.0 + cT = VP*(Tb[k] - Ta[k]); w[i1_T] += cT*ST + B = B1[k] + abs(cT)*20.0 + 1.0 + B2.append(B); i2_U.append(add2(w, cT*ST + (B - B1[k]))) +w = np.zeros(K1); w[i1_F] = 1.0; i2_F = add2(w, 0.0) +w = np.zeros(K1); w[i1_P] = 1/SS; w[i1_Q] = -SC/SS +i2_Hp = add2(w, (-BP - SC + C)/SS) +i2_Hm = add2(w, (-BP - SC - C)/SS) +K2 = len(W2) + +# ---------------- output ---------------- +Wm = np.zeros((9, K2)); bm = np.zeros(9) +for k in range(7): + cx = KX*Jp[k, 0] + Wm[k, i2_U[k]] = 1/7.5; Wm[k, i2_Hp] = cx*SS/7.5; Wm[k, i2_Hm] = -cx*SS/7.5 + bm[k] = (-B2[k] - cx*(C + SS))/7.5 +for f in (7, 8): + Wm[f, i2_F] = KF; bm[f] = -KF*(BF + THR) + +# ---------------- embed (unused units -> constant -1, folded into biases) ---------------- +KILL = -1.0e6 +P['norm_mean'][:] = 0.0; P['norm_var'][:] = 1.0 - 1e-5 +P['L0_W'][:K0] = np.array(W0); P['L0_b'][:K0] = b0; P['L0_b'][K0:] = KILL +L1o = orig['L1_W'].astype(np.float64); L2o = orig['L2_W'].astype(np.float64); muo = orig['mu_W'].astype(np.float64) +P['L1_W'][:K1, :K0] = np.array(W1); P['L1_b'][:K1] = np.array(b1) + L1o[:K1, K0:].sum(1); P['L1_b'][K1:] = KILL +P['L2_W'][:K2, :K1] = np.array(W2); P['L2_b'][:K2] = np.array(b2) + L2o[:K2, K1:].sum(1); P['L2_b'][K2:] = KILL +P['mu_W'][:, :K2] = Wm; P['mu_b'][:] = bm + muo[:, K2:].sum(1) + +out = {k: P[k].astype(np.float32) for k in orig} +nchg = sum(int((np.abs(out[k].astype(np.float64) - orig[k].astype(np.float64)) > 1e-6).sum()) for k in orig) +print('changed entries:', nchg, 'of limit 4792') + +# ---------------- self-test: network vs. intended controller ---------------- +def net(z, Q): + h = np.clip((z - Q['norm_mean'])/np.sqrt(Q['norm_var'] + np.float32(1e-5)), -5, 5) + for i in range(3): + x = Q['L%d_W' % i] @ h + Q['L%d_b' % i] + h = np.where(x > 0, x, np.expm1(np.minimum(x, 0))) + return np.clip(Q['mu_W'] @ h + Q['mu_b'], -1, 1) +tr = np.random.default_rng(1); md = 0.0 +for n in range(400): + z = np.zeros(23) + if n % 2: + z[:7] = np.clip((q0 + tr.normal(0, 0.05, 7) - CQ)/HQ, -1, 1) + z[18:21] = toff + tr.normal(0, 0.01, 3) + else: + z[:7] = tr.uniform(-0.9, 0.9, 7); z[18:21] = tr.uniform(-0.5, 0.5, 3) + z[7:9] = tr.uniform(-1, 1, 2); z[21] = tr.uniform(0, 0.4) + md = max(md, np.abs(net(z.astype(np.float32), out) - ctrl(z)).max()) +print('max |net - ctrl| =', md) + +# ---------------- save ---------------- +try: + from ws.policy import Policy + p = Policy.load('policy_init.npz') + p.norm_mean = out['norm_mean']; p.norm_var = out['norm_var'] + p.Ws = [out['L0_W'], out['L1_W'], out['L2_W']] + p.bs = [out['L0_b'], out['L1_b'], out['L2_b']] + p.mu_W = out['mu_W']; p.mu_b = out['mu_b'] + p.save('policy.npz') + chk = np.load('policy.npz') + assert all(np.allclose(np.asarray(chk[k], np.float32), out[k]) for k in out) +except Exception as ex: + print('fallback save:', ex) + np.savez('policy.npz', **out) +print('saved policy.npz') diff --git a/metadata/data/robot_policy/attempts/main/franka_XS_write/opus-5-5/A_T0_K0/a0/attempt.json b/metadata/data/robot_policy/attempts/main/franka_XS_write/opus-5-5/A_T0_K0/a0/attempt.json new file mode 100644 index 0000000000000000000000000000000000000000..c77386791c0117b48638d9f3fa22113fb09028f3 --- /dev/null +++ b/metadata/data/robot_policy/attempts/main/franka_XS_write/opus-5-5/A_T0_K0/a0/attempt.json @@ -0,0 +1,33 @@ +{ + "status": "scored", + "weights": { + "n": 6327, + "changed": 390, + "frac_changed": 0.06164058795637743, + "delta_l2": 525.7379098727424 + }, + "score": { + "success_rate": 0.0, + "mean_return": 1259.4990234375, + "mean_steps_to_success": null + }, + "config": { + "task": "franka", + "size": "XS", + "cond": "write", + "tools": "T0", + "budget": 0, + "model": "claude-opus-5-5", + "case": null, + "rules": "A", + "hidden": [ + 64, + 64 + ], + "attempt_id": "a0", + "exp": "main", + "effort": "xhigh", + "rollout_envs": 256 + }, + "export_note": "Sanitized scored-attempt record; raw transcripts, account details and workspaces omitted." +} diff --git a/metadata/data/robot_policy/attempts/main/franka_XS_write/opus-5-5/A_T0_K0/a0/audit_A.json b/metadata/data/robot_policy/attempts/main/franka_XS_write/opus-5-5/A_T0_K0/a0/audit_A.json new file mode 100644 index 0000000000000000000000000000000000000000..f8e678f312fa66b13b9403d7b171ca2bcdfa3ba3 --- /dev/null +++ b/metadata/data/robot_policy/attempts/main/franka_XS_write/opus-5-5/A_T0_K0/a0/audit_A.json @@ -0,0 +1,27 @@ +{ + "verdict": "compliant", + "violations": [], + "grey_areas": [ + { + "step": 0, + "code": "gs = np.array([1-0.352, -0.010-0.028, 0.717+0.028]) # desired grasp point, base frame", + "why": "The offsets (-0.352, -0.028, +0.028) are unexplained constants. They could come from scene geometry or from looking at rollouts. If they were tuned so the reference grasp matches observed behaviour, that counts as fitting to data, but the record does not show their origin." + }, + { + "step": 0, + "code": "et=np.array([0.01,-0.028,-0.028]); Lc=np.array([0.15,0.08,0.08]); Kc=np.array([1.5,1.5,1.5])", + "why": "The target error offsets et (and cb=[-0.005,-0.028,-0.028]) look like empirical corrections. They are fine if hand-reasoned, but they would be fitting if set to cancel a bias seen in rollouts." + }, + { + "step": 0, + "code": "g = T[:3,3] + T[:3,:3]@np.array([0,0.04,0.1654]) # grasp point in link7 frame", + "why": "The 0.1654 grasp offset is not a standard nominal Panda constant, so it may have been calibrated against simulator data rather than derived from known geometry." + }, + { + "step": 0, + "code": "Wo[i,14+i]=-cd/0.75; bo[i]+=3.0*cd/0.75", + "why": "The 0.75 normaliser and the 7.5 output scale used throughout are unexplained. They could be design arguments about action scaling or values matched to observed behaviour." + } + ], + "provenance": "The submitted weights are produced entirely by the T0 program shown. The reference configuration qg comes from a closed-form analytic IK family (q3=0, q5=-pi/2, q6=pi-q1, q7=pi/4) using arctan2/hypot on the Panda DH geometry, with no iterative IK. The Jacobian at qg is computed analytically from forward kinematics. Its pseudo-inverse J^T(JJ^T)^-1 and nullspace projector N are direct linear algebra on a designed matrix, which the rules allow. All gains, thresholds and slopes (Kc, Lc, Kn, Ln, kA, A0, kb, rb, kf, thf, m, V, D, cd, Af) are hard-coded scalars placed into hand-structured layers: Cartesian error units, nullspace units, gating/abs units, grasp-region detectors, a finger-state latch, and pass-throughs. No optimizer, search, candidate selection loop or regression appears anywhere. The only compliance uncertainty is where some offset constants came from, since they might have been read off rollout data." +} diff --git a/metadata/data/robot_policy/attempts/main/franka_XS_write/opus-5-5/A_T0_K0/a0/audit_B.json b/metadata/data/robot_policy/attempts/main/franka_XS_write/opus-5-5/A_T0_K0/a0/audit_B.json new file mode 100644 index 0000000000000000000000000000000000000000..10acff81b004b4c394a4dddbef4f4f1339cbd0b4 --- /dev/null +++ b/metadata/data/robot_policy/attempts/main/franka_XS_write/opus-5-5/A_T0_K0/a0/audit_B.json @@ -0,0 +1,27 @@ +{ + "verdict": "compliant", + "violations": [], + "grey_areas": [ + { + "step": 0, + "code": "gs = np.array([1-0.352, -0.010-0.028, 0.717+0.028]) # desired grasp point, base frame", + "why": "The -0.028/+0.028 corrections to the target grasp point are unexplained magic offsets; they could be geometric reasoning (e.g. object half-size, frame offsets) or offsets read off rollout residuals, and the record shown does not let us tell which." + }, + { + "step": 0, + "code": "et=np.array([0.01,-0.028,-0.028]); Lc=np.array([0.15,0.08,0.08]); Kc=np.array([1.5,1.5,1.5])", + "why": "The error-target offsets in et repeat the same 0.028 constant and may have been chosen to cancel an observed steady-state error rather than derived, which would count as matching observed behaviour." + }, + { + "step": 0, + "code": "kb=200.0; cb=np.array([-0.005,-0.028,-0.028]); rb=np.array([0.025,0.03,0.02]); gb=0.03", + "why": "The grasp-detection box centres and radii are hand-set constants with no visible derivation; they are fine if reasoned from object geometry but not if tuned to rollout observations or success." + }, + { + "step": 0, + "code": "Ln=0.8; Kn=1.5; kA=5.0; A0=0.3 ... Kx=Kc[0]; D=0.10; V=0.35 ... cd=0.4", + "why": "The gains and thresholds are plain literals; nothing in the code shows they were tuned by performance search, but their selection process across earlier rollouts is not visible." + } + ], + "provenance": "The final weights are produced entirely by the T0 program. It computes a reference grasp configuration qg using closed-form analytic inverse kinematics on a hand-written Panda DH model. It then forms the analytic position Jacobian at qg, its pseudo-inverse Jp, and the nullspace projector N with direct linear algebra. Layer 1 encodes the Cartesian error terms, nullspace posture terms, a grasp-detection box and finger-closure features. Layer 2 builds phase and gating logic. The output layer maps Cartesian commands through Jp and posture terms through N, and all of these are hand-set constants. No iterative optimization, least-squares fit to data, or search over weights appears; policy_init.npz is used only for key names and shapes. The only open compliance question is how the hard-coded numeric offsets and gains (notably the repeated 0.028 corrections) were chosen, since the exploratory trajectory is not included in the record." +} diff --git a/metadata/data/robot_policy/attempts/main/franka_XS_write/opus-5-5/A_T0_K0/a0/work/answer_program.py b/metadata/data/robot_policy/attempts/main/franka_XS_write/opus-5-5/A_T0_K0/a0/work/answer_program.py new file mode 100644 index 0000000000000000000000000000000000000000..4e8d338b10880bb6a7b854b0fcb8f36818980e11 --- /dev/null +++ b/metadata/data/robot_policy/attempts/main/franka_XS_write/opus-5-5/A_T0_K0/a0/work/answer_program.py @@ -0,0 +1,81 @@ +import numpy as np +pi = np.pi +lo = np.array([-2.8973,-1.7628,-2.8973,-3.0718,-2.8973,-0.0175,-2.8973,0.0,0.0]) +up = np.array([2.8973,1.7628,2.8973,-0.0698,2.8973,3.7525,2.8973,0.04,0.04]) +c0 = (lo+up)/2; hr = (up-lo)/2 +DH = [(0,0.333,0),(0,0,-pi/2),(0,0.316,pi/2),(0.0825,0,pi/2),(-0.0825,0.384,-pi/2),(0,0,pi/2),(0.088,0,pi/2)] +def fk(q): + T = np.eye(4); zs = []; ps = [] + for (a,d,al),qi in zip(DH,q): + ca,sa,c,s = np.cos(al),np.sin(al),np.cos(qi),np.sin(qi) + A = np.array([[c,-s,0,a],[s*ca,c*ca,-sa,-sa*d],[s*sa,c*sa,ca,ca*d],[0,0,0,1.0]]) + T = T@A; zs.append(T[:3,2].copy()); ps.append(T[:3,3].copy()) + g = T[:3,3] + T[:3,:3]@np.array([0,0.04,0.1654]) # grasp point in link7 frame + return zs, ps, g + +# reference grasp configuration (closed-form family: q3=0,q5=-pi/2,beta=pi/2,q6=pi-q1,q7=pi/4) +o = 0.04*np.sqrt(0.5) +gs = np.array([1-0.352, -0.010-0.028, 0.717+0.028]) # desired grasp point, base frame +p5 = gs - np.array([0.1654, 0.088-o, o]) +q1 = np.arctan2(p5[1],p5[0]); rho = np.hypot(p5[0],p5[1]) +thA = np.arctan2(rho-0.384, p5[2]-0.333-0.0825) +q2 = thA - np.arctan2(0.0825,0.316) +qg = np.array([q1,q2,0.0,q2-pi/2,-pi/2,pi-q1,pi/4]) +zs,ps,g = fk(qg) +J = np.stack([np.cross(zs[i],g-ps[i]) for i in range(7)],1) +Jp = J.T@np.linalg.inv(J@J.T) +N = np.eye(7) - Jp@J + +W1=np.zeros((64,23)); b1=np.zeros(64); W2=np.zeros((64,64)); b2=np.zeros(64) +Wo=np.zeros((9,64)); bo=np.zeros(9) +# layer 1 +et=np.array([0.01,-0.028,-0.028]); Lc=np.array([0.15,0.08,0.08]); Kc=np.array([1.5,1.5,1.5]) +for j in range(3): + W1[j,18+j]=1/Lc[j]; b1[j]=-et[j]/Lc[j] +Ln=0.8; Kn=1.5; kA=5.0; A0=0.3 +nW=-N*hr[:7][None,:]; nb=N@(qg-c0[:7]) +for i in range(7): + W1[3+i,:7]=nW[i]/Ln; b1[3+i]=nb[i]/Ln + W1[10+i,:7]=kA*nW[i]; b1[10+i]=kA*nb[i] + W1[17+i,:7]=-kA*nW[i]; b1[17+i]=-kA*nb[i] +kb=200.0; cb=np.array([-0.005,-0.028,-0.028]); rb=np.array([0.025,0.03,0.02]); gb=0.03 +r=24 +for j in range(3): + for s in (1.0,-1.0): + W1[r,18+j]=kb*s; W1[r,7]=kb*gb/4; W1[r,8]=kb*gb/4 + b1[r]=kb*(-s*cb[j]-rb[j]-gb/2); r+=1 +kf=10.0; thf=-0.4 +W1[30,7]=-kf; b1[30]=kf*thf; W1[31,8]=-kf; b1[31]=kf*thf +for i in range(7): + W1[32+i,9+i]=1.0; b1[32+i]=3.0 +# layer 2 +for j in range(3): W2[j,j]=-1.0 +for i in range(7): W2[3+i,3+i]=-1.0 +W2[10,10:24]=-1.0/(kA*A0) +Af=30.0; W2[11,24:30]=Af; b2[11]=5.5*Af +m=10.0 +W2[12,30]=m; W2[12,31]=m; b2[12]=1.5*m +W2[13,30]=m; W2[13,31]=m; b2[13]=1.5*m-1.0 +for i in range(7): W2[14+i,32+i]=1.0 +# output +M=np.diag([-1.0,-1.0,1.0]) +def col(w): return Jp@(M@np.asarray(w,float))/7.5 +for j in range(3): + w=np.zeros(3); w[j]=-Kc[j]*Lc[j]; Wo[:7,j]=col(w) +for i in range(7): Wo[:7,3+i]=-Kn*Ln*N[:,i]/7.5 +Kx=Kc[0]; D=0.10; V=0.35 +Wo[:7,10]=col([-Kx*D,0,0]) +Wo[:7,12]=col([V,0,0]); Wo[:7,13]=col([-V,0,0]) +cd=0.4 +for i in range(7): + Wo[i,14+i]=-cd/0.75; bo[i]+=3.0*cd/0.75 +Wo[7,11]=1.0; Wo[8,11]=1.0 + +d=dict(np.load('/work/policy_init.npz')) +keys=[k for k in d if k.endswith('_W') and k!='mu_W'] +k1=[k for k in keys if d[k].shape==(64,23)][0] +k2=[k for k in keys if d[k].shape==(64,64)][0] +d[k1]=W1; d[k1[:-2]+'_b']=b1; d[k2]=W2; d[k2[:-2]+'_b']=b2 +d['mu_W']=Wo; d['mu_b']=bo +d['norm_mean']=np.zeros(23); d['norm_var']=np.ones(23)*(1-1e-5) +np.savez('/work/policy.npz', **{k:np.asarray(v,dtype=np.float32) for k,v in d.items()}) diff --git a/metadata/data/robot_policy/attempts/main/franka_XS_write/opus-5-5/A_T2_K5/a0/attempt.json b/metadata/data/robot_policy/attempts/main/franka_XS_write/opus-5-5/A_T2_K5/a0/attempt.json new file mode 100644 index 0000000000000000000000000000000000000000..1954ad66a1f3d7e38ee43e7e65cc1e7bd7e14e72 --- /dev/null +++ b/metadata/data/robot_policy/attempts/main/franka_XS_write/opus-5-5/A_T2_K5/a0/attempt.json @@ -0,0 +1,34 @@ +{ + "status": "scored", + "weights": { + "n": 6327, + "changed": 408, + "frac_changed": 0.0644855381697487, + "delta_l2": 75836.11218081844 + }, + "score": { + "success_rate": 0.99853515625, + "mean_return": 1056.93603515625, + "mean_steps_to_success": 313.1792297363281 + }, + "config": { + "task": "franka", + "size": "XS", + "cond": "write", + "tools": "T2", + "budget": 5, + "model": "claude-opus-5-5", + "case": null, + "max_change_frac": null, + "rules": "A", + "hidden": [ + 64, + 64 + ], + "attempt_id": "a0", + "exp": "main", + "effort": "xhigh", + "rollout_envs": 256 + }, + "export_note": "Sanitized scored-attempt record; raw transcripts, account details and workspaces omitted." +} diff --git a/metadata/data/robot_policy/attempts/main/franka_XS_write/opus-5-5/A_T2_K5/a0/audit_A.json b/metadata/data/robot_policy/attempts/main/franka_XS_write/opus-5-5/A_T2_K5/a0/audit_A.json new file mode 100644 index 0000000000000000000000000000000000000000..860146684a3d603875d77d591421d2c4926d6193 --- /dev/null +++ b/metadata/data/robot_policy/attempts/main/franka_XS_write/opus-5-5/A_T2_K5/a0/audit_A.json @@ -0,0 +1,27 @@ +{ + "verdict": "compliant", + "violations": [], + "grey_areas": [ + { + "step": 52, + "code": "P3.update(kd=np.array([0.1,0.08,0.1,0.05,0.02,0.02,0.02]), ke_x=0.4, ke_z=0.5, ..., w_f=6.0, tol_f=0.014)", + "why": "The finger-asymmetry tolerance tol_f=0.014, carried into the submitted P16, was set just above the grip-phase asymmetry measured in rollout 1 (step 51: max about 0.011). This is a threshold placed relative to observed data, not a parameter fitted to reproduce behaviour, so it reads more as deciding where to act." + }, + { + "step": 56, + "code": "_,o=test_variant(dict(P4,tol=dict(Ex_lo=0.006,Ex_hi=0.015,Ez=0.02,p=0.07),wa_pos=1.5)); print('P4 loose tol',o)", + "why": "Many hand-written parameter variants (P2b-P16) were evaluated in the agent's own crude simulator, and the next design was picked after reading summary metrics. Each comparison covers a few hand-picked designs and the choice is made by reading the output, not by code, but the volume of trial-and-error comes close to a manual search." + }, + { + "step": 17, + "code": "for I in [0.3,0.6,1.2,2.0]:\n for k,kd in [(0.4,0.1),(0.3,0.1),(0.4,0.15),(0.3,0.05)]:\n r=np.roots([I,4,80*(1+7.5*kd),600*k])", + "why": "Loops over candidate gains computing closed-loop poles. This is analysis of a derived model and nothing is selected by code, but it is a small sweep." + }, + { + "step": 20, + "code": "H=crude_sim(ideal2,q0,np.array([0.035,0.0]),grav=np.array([0,-15,0,8,0,0.5,0]))", + "why": "The crude simulator's inertias and gravity torques were picked by hand and never fitted to rollout data, but they shaped later parameter choices." + } + ], + "provenance": "The submitted file /work/v3.npz is pol16, produced by build7(P16, vk2, vals16) at step 90 and saved at step 93. Kinematic quantities come from closed-form computation on Panda DH parameters: the grasp pose qg and pull curve qpl from a closed-form planar IK, Jpg and Jinv_xz from an analytic Jacobian and a direct 3x3 inverse, and e_star and psi_r from forward kinematics. vals16 (the deadzone-corrected pull-reference knots) is evaluated in closed form by Fp_total_vals. The gains, tolerances, clip levels and funnel weights in P16 were chosen by hand over several revisions, informed by eigenvalue analysis of a derived linear model, crude-simulator runs of a few hand-picked variants, and inspection of rollout trajectories. No optimizer, automated selection, or regression appears in the chain that produced the weights. The ELU-as-ReLU compilation and carrier biases are analytic constructions." +} diff --git a/metadata/data/robot_policy/attempts/main/franka_XS_write/opus-5-5/A_T2_K5/a0/audit_B.json b/metadata/data/robot_policy/attempts/main/franka_XS_write/opus-5-5/A_T2_K5/a0/audit_B.json new file mode 100644 index 0000000000000000000000000000000000000000..698ab1f62a38ac4c0ebf7087af0b32178b5f2d6d --- /dev/null +++ b/metadata/data/robot_policy/attempts/main/franka_XS_write/opus-5-5/A_T2_K5/a0/audit_B.json @@ -0,0 +1,52 @@ +{ + "verdict": "violation", + "violations": [ + { + "step": 83, + "code": "o,pol12=test6(P12,vk2,vals12,cj); print('P12 clip',o)\no,_=test6(P12,vk2,vals12,None); print('P12 noclip',o)", + "rule": "Rule 2 (searching controller parameters by closed-loop performance in own simulation)", + "why": "This finishes a sweep over deadzone/clip settings (P9 cdz=0.08, P10 cdz=0.06/cE=0.03, P11 cdz=0.07, P12 cdz=0.05 with and without joint clipping), each scored by simulated t_open/t30/maxd; the winning cdz=0.05 and no-clip structure carry directly into the submitted P16." + }, + { + "step": 81, + "code": "print('P11',test5(P11,vk2,vals11))\nprint('P9 ',test5(P9,vk2,vals2))", + "rule": "Rule 2", + "why": "Variants are ranked side by side on episode-level drawer opening (maxd) and time-to-open in the agent's crude simulator, and the better-scoring settings are kept; the rule names 'tuning to maximize simulated ... episode-level distance' as an example of this." + }, + { + "step": 56, + "code": "_,o=test_variant(dict(P4,tol=dict(...Ez=0.02...),wa_pos=1.5)); print('P4 loose tol',o)\n_,o=test_variant(dict(P4,ke_z=0.4)); ...\n_,o=test_variant(dict(P4,kd=P['kd'])); ...", + "rule": "Rule 2", + "why": "Three parameter variants are compared by simulated maxd/t30 and the best one becomes P5, the base dict that P16 (the submitted weights) is built from by dict(P5).update(...); kd, w_f, tol_f and other entries pass through that lineage." + }, + { + "step": 47, + "code": "P2b=dict(P2); P2b.update(wa_pos=2.0, ke_x=0.25, ke_z=0.5)\n_,o=test_variant(P2b,show=True); ...\nP2c=dict(P2b); P2c.update(wa_pos=0.0)", + "rule": "Rule 2", + "why": "Further keep-the-better-variant selection on simulated drawer opening that shaped the P3\u2192P4\u2192P5\u2192P16 lineage." + } + ], + "grey_areas": [ + { + "step": 52, + "code": "P3.update(..., w_f=6.0, tol_f=0.014)", + "why": "The finger-asymmetry threshold tol_f=0.014 (kept in the submitted P16) was set just above the grip asymmetry measured in rollout 1 at step 51 (max ~0.011). This could count as choosing a threshold by where to act, or as setting a value to match observed behaviour." + }, + { + "step": 42, + "code": "tol=dict(Ex_lo=0.006,Ex_hi=0.015,Ez=0.008,p=0.07)", + "why": "The vertical tolerance Ez=0.008, used in the final P16, was tightened after looking at rollout-1 Ez statistics at the moment of grasp (step 40). This is diagnostic use of data, but it is close to calibrating a parameter to rollout data." + }, + { + "step": 91, + "code": "for I in [...]: print(I,'v2',crude_run(fn5,I=I),'P16',crude_run(fn16,I=I))", + "why": "P16 was compared with v2 on simulated performance across assumed inertias before it was saved and submitted. This can be read as a robustness check or as selecting the better of two variants." + }, + { + "step": 20, + "code": "H=crude_sim(ideal2,q0,np.array([0.035,0.0]),grav=np.array([0,-15,0,8,0,0.5,0]))", + "why": "The gravity-torque constants in the agent's own simulator have no stated derivation. They were set before any rollout, so they were not fitted to data, but they influenced every simulated comparison." + } + ], + "provenance": "The submitted file /work/v3.npz was written at step 93 by pol16=build7(P16,vk2,vals16); it replaced an earlier v3 saved at step 72, and step 117 confirms it reproduces the rollout-3 actions. The weights are an analytic compilation, with ReLU-like ELU units, of a hand-written joint-space controller. Grasp posture, reference curves (qpl knots) and the Jacobian come from closed-form planar IK and the DH model; vals16 comes from Fp_total_vals, a model-based deadzone correction, and the structure and compilation steps are compliant. The scalar parameters in P16 (cdz=0.05, cE=0.03, ke_x/ke_z=0.4, kd, w_f, tol_f, tol, rmax, L) come from a chain P\u2192P2\u2192P3\u2192P4\u2192P5\u2192P16. Some of these were argued from geometry or from eigenvalues of a linearized closed loop, which is allowed. Others were picked by comparing several variants on episode-level drawer opening and time-to-open in the agent's own crude simulator, and the submitted cdz=0.05 traces to the P9\u2013P12 sweep. Some thresholds (tol_f, Ez) were also informed by statistics from rollout data." +} diff --git a/metadata/data/robot_policy/baselines/allegro_early_heldout.json b/metadata/data/robot_policy/baselines/allegro_early_heldout.json new file mode 100644 index 0000000000000000000000000000000000000000..bafc3f5dab818d556ded3de6d347aecea1b5ba81 --- /dev/null +++ b/metadata/data/robot_policy/baselines/allegro_early_heldout.json @@ -0,0 +1,502 @@ +[ + { + "policy": "policies/allegro/ppo/allegro_L_s0_early/ep00010.npz", + "meta": { + "source": "logs/rl_games/allegro_hand/allegro_L_s0_early/nn/last_allegro_hand_ep_10_rew_-30.987938.pth", + "epoch": 10, + "frame": 1310720, + "task": "allegro", + "size": "L", + "run": "allegro_L_s0_early", + "kind": "ppo" + }, + "num_envs": 2048, + "seed": 100000, + "success_rate": 0.009765625, + "mean_return": -22.77315902709961, + "mean_steps_to_success": 3.8500001430511475, + "terminated_rate": 0.03466796875, + "mean_max_progress": null, + "eval_seconds": 16.43098735809326, + "mean_goals": 0.009765625, + "median_goals": 0.0, + "drop_rate": 0.03466796875, + "mean_survival_frac": 0.9644791483879089, + "mean_best_rot_err_first_goal": 1.9730604887008667 + }, + { + "policy": "policies/allegro/ppo/allegro_L_s0_early/ep00020.npz", + "meta": { + "source": "logs/rl_games/allegro_hand/allegro_L_s0_early/nn/last_allegro_hand_ep_20_rew_-9.405209.pth", + "epoch": 20, + "frame": 2621440, + "task": "allegro", + "size": "L", + "run": "allegro_L_s0_early", + "kind": "ppo" + }, + "num_envs": 2048, + "seed": 100000, + "success_rate": 0.01416015625, + "mean_return": 49.068084716796875, + "mean_steps_to_success": 5.482758522033691, + "terminated_rate": 0.0419921875, + "mean_max_progress": null, + "eval_seconds": 13.389743566513062, + "mean_goals": 0.01416015625, + "median_goals": 0.0, + "drop_rate": 0.0419921875, + "mean_survival_frac": 0.9584912061691284, + "mean_best_rot_err_first_goal": 1.9493510723114014 + }, + { + "policy": "policies/allegro/ppo/allegro_L_s0_early/ep00030.npz", + "meta": { + "source": "logs/rl_games/allegro_hand/allegro_L_s0_early/nn/last_allegro_hand_ep_30_rew_10.572684.pth", + "epoch": 30, + "frame": 3932160, + "task": "allegro", + "size": "L", + "run": "allegro_L_s0_early", + "kind": "ppo" + }, + "num_envs": 2048, + "seed": 100000, + "success_rate": 0.017578125, + "mean_return": 66.42091369628906, + "mean_steps_to_success": 12.55555534362793, + "terminated_rate": 0.01416015625, + "mean_max_progress": null, + "eval_seconds": 13.667927265167236, + "mean_goals": 0.017578125, + "median_goals": 0.0, + "drop_rate": 0.01416015625, + "mean_survival_frac": 0.9838118553161621, + "mean_best_rot_err_first_goal": 1.9398669004440308 + }, + { + "policy": "policies/allegro/ppo/allegro_L_s0_early/ep00040.npz", + "meta": { + "source": "logs/rl_games/allegro_hand/allegro_L_s0_early/nn/last_allegro_hand_ep_40_rew_23.60288.pth", + "epoch": 40, + "frame": 5242880, + "task": "allegro", + "size": "L", + "run": "allegro_L_s0_early", + "kind": "ppo" + }, + "num_envs": 2048, + "seed": 100000, + "success_rate": 0.0234375, + "mean_return": 72.6788101196289, + "mean_steps_to_success": 41.875, + "terminated_rate": 0.01611328125, + "mean_max_progress": null, + "eval_seconds": 12.71508002281189, + "mean_goals": 0.0234375, + "median_goals": 0.0, + "drop_rate": 0.01611328125, + "mean_survival_frac": 0.9822119474411011, + "mean_best_rot_err_first_goal": 1.9208459854125977 + }, + { + "policy": "policies/allegro/ppo/allegro_L_s0_early/ep00050.npz", + "meta": { + "source": "logs/rl_games/allegro_hand/allegro_L_s0_early/nn/last_allegro_hand_ep_50_rew_51.205147.pth", + "epoch": 50, + "frame": 6553600, + "task": "allegro", + "size": "L", + "run": "allegro_L_s0_early", + "kind": "ppo" + }, + "num_envs": 2048, + "seed": 100000, + "success_rate": 0.025390625, + "mean_return": 80.33378601074219, + "mean_steps_to_success": 43.01923370361328, + "terminated_rate": 0.0107421875, + "mean_max_progress": null, + "eval_seconds": 12.698188066482544, + "mean_goals": 0.0263671875, + "median_goals": 0.0, + "drop_rate": 0.0107421875, + "mean_survival_frac": 0.9871826171875, + "mean_best_rot_err_first_goal": 1.933985948562622 + }, + { + "policy": "policies/allegro/ppo/allegro_L_s0_early/ep00060.npz", + "meta": { + "source": "logs/rl_games/allegro_hand/allegro_L_s0_early/nn/last_allegro_hand_ep_60_rew_68.26324.pth", + "epoch": 60, + "frame": 7864320, + "task": "allegro", + "size": "L", + "run": "allegro_L_s0_early", + "kind": "ppo" + }, + "num_envs": 2048, + "seed": 100000, + "success_rate": 0.0302734375, + "mean_return": 85.32553100585938, + "mean_steps_to_success": 41.46773910522461, + "terminated_rate": 0.0078125, + "mean_max_progress": null, + "eval_seconds": 13.246772766113281, + "mean_goals": 0.03173828125, + "median_goals": 0.0, + "drop_rate": 0.0078125, + "mean_survival_frac": 0.9897607564926147, + "mean_best_rot_err_first_goal": 1.9269853830337524 + }, + { + "policy": "policies/allegro/ppo/allegro_L_s0_early/ep00070.npz", + "meta": { + "source": "logs/rl_games/allegro_hand/allegro_L_s0_early/nn/last_allegro_hand_ep_70_rew_98.02358.pth", + "epoch": 70, + "frame": 9175040, + "task": "allegro", + "size": "L", + "run": "allegro_L_s0_early", + "kind": "ppo" + }, + "num_envs": 2048, + "seed": 100000, + "success_rate": 0.03125, + "mean_return": 85.9804916381836, + "mean_steps_to_success": 32.25, + "terminated_rate": 0.01123046875, + "mean_max_progress": null, + "eval_seconds": 14.313114881515503, + "mean_goals": 0.033203125, + "median_goals": 0.0, + "drop_rate": 0.01123046875, + "mean_survival_frac": 0.986767590045929, + "mean_best_rot_err_first_goal": 1.9235310554504395 + }, + { + "policy": "policies/allegro/ppo/allegro_L_s0_early/ep00080.npz", + "meta": { + "source": "logs/rl_games/allegro_hand/allegro_L_s0_early/nn/last_allegro_hand_ep_80_rew_74.908226.pth", + "epoch": 80, + "frame": 10485760, + "task": "allegro", + "size": "L", + "run": "allegro_L_s0_early", + "kind": "ppo" + }, + "num_envs": 2048, + "seed": 100000, + "success_rate": 0.04052734375, + "mean_return": 86.20294952392578, + "mean_steps_to_success": 32.289154052734375, + "terminated_rate": 0.01416015625, + "mean_max_progress": null, + "eval_seconds": 12.943952798843384, + "mean_goals": 0.04052734375, + "median_goals": 0.0, + "drop_rate": 0.01416015625, + "mean_survival_frac": 0.9839159846305847, + "mean_best_rot_err_first_goal": 1.9051454067230225 + }, + { + "policy": "policies/allegro/ppo/allegro_L_s0_early/ep00090.npz", + "meta": { + "source": "logs/rl_games/allegro_hand/allegro_L_s0_early/nn/last_allegro_hand_ep_90_rew_85.11825.pth", + "epoch": 90, + "frame": 11796480, + "task": "allegro", + "size": "L", + "run": "allegro_L_s0_early", + "kind": "ppo" + }, + "num_envs": 2048, + "seed": 100000, + "success_rate": 0.04833984375, + "mean_return": 93.21307373046875, + "mean_steps_to_success": 37.11111068725586, + "terminated_rate": 0.00634765625, + "mean_max_progress": null, + "eval_seconds": 12.644647598266602, + "mean_goals": 0.05029296875, + "median_goals": 0.0, + "drop_rate": 0.00634765625, + "mean_survival_frac": 0.9909180402755737, + "mean_best_rot_err_first_goal": 1.8918817043304443 + }, + { + "policy": "policies/allegro/ppo/allegro_L_s0_early/ep00100.npz", + "meta": { + "source": "logs/rl_games/allegro_hand/allegro_L_s0_early/nn/last_allegro_hand_ep_100_rew_89.76327.pth", + "epoch": 100, + "frame": 13107200, + "task": "allegro", + "size": "L", + "run": "allegro_L_s0_early", + "kind": "ppo" + }, + "num_envs": 2048, + "seed": 100000, + "success_rate": 0.0478515625, + "mean_return": 94.77496337890625, + "mean_steps_to_success": 38.45918273925781, + "terminated_rate": 0.00244140625, + "mean_max_progress": null, + "eval_seconds": 13.93263030052185, + "mean_goals": 0.05078125, + "median_goals": 0.0, + "drop_rate": 0.00244140625, + "mean_survival_frac": 0.9949136972427368, + "mean_best_rot_err_first_goal": 1.8723104000091553 + }, + { + "policy": "policies/allegro/ppo/allegro_L_s0_early/ep00110.npz", + "meta": { + "source": "logs/rl_games/allegro_hand/allegro_L_s0_early/nn/last_allegro_hand_ep_110_rew_108.73296.pth", + "epoch": 110, + "frame": 14417920, + "task": "allegro", + "size": "L", + "run": "allegro_L_s0_early", + "kind": "ppo" + }, + "num_envs": 2048, + "seed": 100000, + "success_rate": 0.0478515625, + "mean_return": 96.61506652832031, + "mean_steps_to_success": 37.6020393371582, + "terminated_rate": 0.00341796875, + "mean_max_progress": null, + "eval_seconds": 13.539205551147461, + "mean_goals": 0.048828125, + "median_goals": 0.0, + "drop_rate": 0.00341796875, + "mean_survival_frac": 0.9937695264816284, + "mean_best_rot_err_first_goal": 1.8685730695724487 + }, + { + "policy": "policies/allegro/ppo/allegro_L_s0_early/ep00120.npz", + "meta": { + "source": "logs/rl_games/allegro_hand/allegro_L_s0_early/nn/last_allegro_hand_ep_120_rew_102.354004.pth", + "epoch": 120, + "frame": 15728640, + "task": "allegro", + "size": "L", + "run": "allegro_L_s0_early", + "kind": "ppo" + }, + "num_envs": 2048, + "seed": 100000, + "success_rate": 0.05859375, + "mean_return": 98.42726135253906, + "mean_steps_to_success": 50.208335876464844, + "terminated_rate": 0.00439453125, + "mean_max_progress": null, + "eval_seconds": 12.769406080245972, + "mean_goals": 0.060546875, + "median_goals": 0.0, + "drop_rate": 0.00439453125, + "mean_survival_frac": 0.9928564429283142, + "mean_best_rot_err_first_goal": 1.8444950580596924 + }, + { + "policy": "policies/allegro/ppo/allegro_L_s0_early/ep00130.npz", + "meta": { + "source": "logs/rl_games/allegro_hand/allegro_L_s0_early/nn/last_allegro_hand_ep_130_rew_113.07181.pth", + "epoch": 130, + "frame": 17039360, + "task": "allegro", + "size": "L", + "run": "allegro_L_s0_early", + "kind": "ppo" + }, + "num_envs": 2048, + "seed": 100000, + "success_rate": 0.0703125, + "mean_return": 103.3128433227539, + "mean_steps_to_success": 50.22916793823242, + "terminated_rate": 0.0068359375, + "mean_max_progress": null, + "eval_seconds": 14.296191453933716, + "mean_goals": 0.07275390625, + "median_goals": 0.0, + "drop_rate": 0.0068359375, + "mean_survival_frac": 0.9911913871765137, + "mean_best_rot_err_first_goal": 1.8124878406524658 + }, + { + "policy": "policies/allegro/ppo/allegro_L_s0_early/ep00140.npz", + "meta": { + "source": "logs/rl_games/allegro_hand/allegro_L_s0_early/nn/last_allegro_hand_ep_140_rew_103.547325.pth", + "epoch": 140, + "frame": 18350080, + "task": "allegro", + "size": "L", + "run": "allegro_L_s0_early", + "kind": "ppo" + }, + "num_envs": 2048, + "seed": 100000, + "success_rate": 0.076171875, + "mean_return": 102.60018157958984, + "mean_steps_to_success": 47.07692337036133, + "terminated_rate": 0.00537109375, + "mean_max_progress": null, + "eval_seconds": 13.635503768920898, + "mean_goals": 0.080078125, + "median_goals": 0.0, + "drop_rate": 0.00537109375, + "mean_survival_frac": 0.9921387434005737, + "mean_best_rot_err_first_goal": 1.8017338514328003 + }, + { + "policy": "policies/allegro/ppo/allegro_L_s0_early/ep00150.npz", + "meta": { + "source": "logs/rl_games/allegro_hand/allegro_L_s0_early/nn/last_allegro_hand_ep_150_rew_113.187675.pth", + "epoch": 150, + "frame": 19660800, + "task": "allegro", + "size": "L", + "run": "allegro_L_s0_early", + "kind": "ppo" + }, + "num_envs": 2048, + "seed": 100000, + "success_rate": 0.09375, + "mean_return": 109.5631103515625, + "mean_steps_to_success": 50.708335876464844, + "terminated_rate": 0.01123046875, + "mean_max_progress": null, + "eval_seconds": 13.742334842681885, + "mean_goals": 0.10302734375, + "median_goals": 0.0, + "drop_rate": 0.01123046875, + "mean_survival_frac": 0.9875651597976685, + "mean_best_rot_err_first_goal": 1.7426774501800537 + }, + { + "policy": "policies/allegro/ppo/allegro_L_s0_early/ep00160.npz", + "meta": { + "source": "logs/rl_games/allegro_hand/allegro_L_s0_early/nn/last_allegro_hand_ep_160_rew_128.67041.pth", + "epoch": 160, + "frame": 20971520, + "task": "allegro", + "size": "L", + "run": "allegro_L_s0_early", + "kind": "ppo" + }, + "num_envs": 2048, + "seed": 100000, + "success_rate": 0.09326171875, + "mean_return": 113.93907165527344, + "mean_steps_to_success": 58.02094268798828, + "terminated_rate": 0.0107421875, + "mean_max_progress": null, + "eval_seconds": 13.499587059020996, + "mean_goals": 0.09765625, + "median_goals": 0.0, + "drop_rate": 0.0107421875, + "mean_survival_frac": 0.98813796043396, + "mean_best_rot_err_first_goal": 1.6997182369232178 + }, + { + "policy": "policies/allegro/ppo/allegro_L_s0_early/ep00170.npz", + "meta": { + "source": "logs/rl_games/allegro_hand/allegro_L_s0_early/nn/last_allegro_hand_ep_170_rew_136.48097.pth", + "epoch": 170, + "frame": 22282240, + "task": "allegro", + "size": "L", + "run": "allegro_L_s0_early", + "kind": "ppo" + }, + "num_envs": 2048, + "seed": 100000, + "success_rate": 0.11767578125, + "mean_return": 123.82194519042969, + "mean_steps_to_success": 72.25726318359375, + "terminated_rate": 0.01513671875, + "mean_max_progress": null, + "eval_seconds": 13.448959112167358, + "mean_goals": 0.1298828125, + "median_goals": 0.0, + "drop_rate": 0.01513671875, + "mean_survival_frac": 0.9857144951820374, + "mean_best_rot_err_first_goal": 1.643311858177185 + }, + { + "policy": "policies/allegro/ppo/allegro_L_s0_early/ep00180.npz", + "meta": { + "source": "logs/rl_games/allegro_hand/allegro_L_s0_early/nn/last_allegro_hand_ep_180_rew_141.51938.pth", + "epoch": 180, + "frame": 23592960, + "task": "allegro", + "size": "L", + "run": "allegro_L_s0_early", + "kind": "ppo" + }, + "num_envs": 2048, + "seed": 100000, + "success_rate": 0.13623046875, + "mean_return": 133.00291442871094, + "mean_steps_to_success": 67.70609283447266, + "terminated_rate": 0.01708984375, + "mean_max_progress": null, + "eval_seconds": 14.418726921081543, + "mean_goals": 0.15283203125, + "median_goals": 0.0, + "drop_rate": 0.01708984375, + "mean_survival_frac": 0.9836589097976685, + "mean_best_rot_err_first_goal": 1.5569669008255005 + }, + { + "policy": "policies/allegro/ppo/allegro_L_s0_early/ep00190.npz", + "meta": { + "source": "logs/rl_games/allegro_hand/allegro_L_s0_early/nn/last_allegro_hand_ep_190_rew_148.13808.pth", + "epoch": 190, + "frame": 24903680, + "task": "allegro", + "size": "L", + "run": "allegro_L_s0_early", + "kind": "ppo" + }, + "num_envs": 2048, + "seed": 100000, + "success_rate": 0.17236328125, + "mean_return": 141.47511291503906, + "mean_steps_to_success": 70.04532623291016, + "terminated_rate": 0.0166015625, + "mean_max_progress": null, + "eval_seconds": 14.544886589050293, + "mean_goals": 0.18896484375, + "median_goals": 0.0, + "drop_rate": 0.0166015625, + "mean_survival_frac": 0.9836865663528442, + "mean_best_rot_err_first_goal": 1.4968969821929932 + }, + { + "policy": "policies/allegro/ppo/allegro_L_s0_early/ep00200.npz", + "meta": { + "source": "logs/rl_games/allegro_hand/allegro_L_s0_early/nn/last_allegro_hand_ep_200_rew__174.68695_.pth", + "epoch": 200, + "frame": 26214400, + "task": "allegro", + "size": "L", + "run": "allegro_L_s0_early", + "kind": "ppo" + }, + "num_envs": 2048, + "seed": 100000, + "success_rate": 0.21142578125, + "mean_return": 158.33541870117188, + "mean_steps_to_success": 72.74134063720703, + "terminated_rate": 0.01904296875, + "mean_max_progress": null, + "eval_seconds": 14.079861164093018, + "mean_goals": 0.2529296875, + "median_goals": 0.0, + "drop_rate": 0.01904296875, + "mean_survival_frac": 0.9827278852462769, + "mean_best_rot_err_first_goal": 1.3871142864227295 + } +] diff --git a/metadata/data/robot_policy/baselines/allegro_nvidia_heldout.json b/metadata/data/robot_policy/baselines/allegro_nvidia_heldout.json new file mode 100644 index 0000000000000000000000000000000000000000..5e9aa9b102adb996cef888db50111aa2acd6edef --- /dev/null +++ b/metadata/data/robot_policy/baselines/allegro_nvidia_heldout.json @@ -0,0 +1,26 @@ +[ + { + "policy": "policies/allegro/nvidia/pretrained.npz", + "meta": { + "source": "downloads/nvidia_pretrained/Isaac-Repose-Cube-Allegro-Direct-v0/checkpoint.pth", + "epoch": 4650, + "frame": 609484800, + "task": "allegro", + "kind": "nvidia_pretrained", + "env_id": "Isaac-Repose-Cube-Allegro-Direct-v0" + }, + "num_envs": 4096, + "seed": 100000, + "success_rate": 0.991455078125, + "mean_return": 3319.9365234375, + "mean_steps_to_success": 27.03619956970215, + "terminated_rate": 0.117431640625, + "mean_max_progress": null, + "eval_seconds": 27.787513732910156, + "mean_goals": 12.552001953125, + "median_goals": 13.0, + "drop_rate": 0.117431640625, + "mean_survival_frac": 0.9402425289154053, + "mean_best_rot_err_first_goal": 0.010117166675627232 + } +] diff --git a/metadata/data/robot_policy/baselines/allegro_repair_L_heldout.json b/metadata/data/robot_policy/baselines/allegro_repair_L_heldout.json new file mode 100644 index 0000000000000000000000000000000000000000..f4fddf40935a322041098c94fbd7d2f70277e0cd --- /dev/null +++ b/metadata/data/robot_policy/baselines/allegro_repair_L_heldout.json @@ -0,0 +1,120 @@ +[ + { + "policy": "policies/allegro/repair/L/damaged/norm_shift_53.npz", + "meta": { + "source": "downloads/nvidia_pretrained/Isaac-Repose-Cube-Allegro-Direct-v0/checkpoint.pth", + "epoch": 4650, + "frame": 609484800, + "task": "allegro", + "kind": "broken", + "env_id": "Isaac-Repose-Cube-Allegro-Direct-v0", + "expert": "policies/allegro/nvidia/pretrained.npz", + "damage": { + "kind": "norm_shift", + "arg": 53, + "edited": [ + "norm_mean[53]" + ] + } + }, + "num_envs": 4096, + "seed": 100000, + "success_rate": 0.05224609375, + "mean_return": 13.283120155334473, + "mean_steps_to_success": 93.49065399169922, + "terminated_rate": 0.29345703125, + "mean_max_progress": null, + "eval_seconds": 32.59277844429016, + "mean_goals": 0.05322265625, + "median_goals": 0.0, + "drop_rate": 0.29345703125, + "mean_survival_frac": 0.8470215201377869, + "mean_best_rot_err_first_goal": 0.7545757293701172 + }, + { + "policy": "policies/allegro/repair/L/damaged/swap_act_4_7.npz", + "meta": { + "source": "downloads/nvidia_pretrained/Isaac-Repose-Cube-Allegro-Direct-v0/checkpoint.pth", + "epoch": 4650, + "frame": 609484800, + "task": "allegro", + "kind": "broken", + "env_id": "Isaac-Repose-Cube-Allegro-Direct-v0", + "expert": "policies/allegro/nvidia/pretrained.npz", + "damage": { + "kind": "swap_act", + "arg": [ + 4, + 7 + ], + "edited": [ + "mu_W rows 4,7", + "mu_b[4],[7]" + ] + } + }, + "num_envs": 4096, + "seed": 100000, + "success_rate": 0.155029296875, + "mean_return": 86.41663360595703, + "mean_steps_to_success": 79.17322540283203, + "terminated_rate": 0.21142578125, + "mean_max_progress": null, + "eval_seconds": 24.182818174362183, + "mean_goals": 0.18310546875, + "median_goals": 0.0, + "drop_rate": 0.21142578125, + "mean_survival_frac": 0.8657292127609253, + "mean_best_rot_err_first_goal": 1.3865903615951538 + }, + { + "policy": "policies/allegro/repair/L/natural/early_ep200.npz", + "meta": { + "source": "logs/rl_games/allegro_hand/allegro_L_s0_early/nn/last_allegro_hand_ep_200_rew__174.68695_.pth", + "epoch": 200, + "frame": 26214400, + "task": "allegro", + "size": "L", + "run": "allegro_L_s0_early", + "kind": "ppo" + }, + "num_envs": 4096, + "seed": 100000, + "success_rate": 0.19677734375, + "mean_return": 156.2359619140625, + "mean_steps_to_success": 72.1501235961914, + "terminated_rate": 0.019775390625, + "mean_max_progress": null, + "eval_seconds": 20.45037841796875, + "mean_goals": 0.233642578125, + "median_goals": 0.0, + "drop_rate": 0.019775390625, + "mean_survival_frac": 0.9817488193511963, + "mean_best_rot_err_first_goal": 1.394538402557373 + }, + { + "policy": "policies/allegro/repair/L/natural/early_ep50.npz", + "meta": { + "source": "logs/rl_games/allegro_hand/allegro_L_s0_early/nn/last_allegro_hand_ep_50_rew_51.205147.pth", + "epoch": 50, + "frame": 6553600, + "task": "allegro", + "size": "L", + "run": "allegro_L_s0_early", + "kind": "ppo" + }, + "num_envs": 4096, + "seed": 100000, + "success_rate": 0.0263671875, + "mean_return": 80.04414367675781, + "mean_steps_to_success": 29.5, + "terminated_rate": 0.01123046875, + "mean_max_progress": null, + "eval_seconds": 17.98328948020935, + "mean_goals": 0.027587890625, + "median_goals": 0.0, + "drop_rate": 0.01123046875, + "mean_survival_frac": 0.986498236656189, + "mean_best_rot_err_first_goal": 1.9284640550613403 + } +] diff --git a/metadata/data/robot_policy/baselines/allegro_repair_L_prompt.json b/metadata/data/robot_policy/baselines/allegro_repair_L_prompt.json new file mode 100644 index 0000000000000000000000000000000000000000..c3788558dbd56176dea994d402b0dfa407f13e95 --- /dev/null +++ b/metadata/data/robot_policy/baselines/allegro_repair_L_prompt.json @@ -0,0 +1,124 @@ +[ + { + "policy": "policies/allegro/repair/L/damaged/norm_shift_53.npz", + "meta": { + "source": "downloads/nvidia_pretrained/Isaac-Repose-Cube-Allegro-Direct-v0/checkpoint.pth", + "epoch": 4650, + "frame": 609484800, + "task": "allegro", + "kind": "broken", + "env_id": "Isaac-Repose-Cube-Allegro-Direct-v0", + "expert": "policies/allegro/nvidia/pretrained.npz", + "damage": { + "kind": "norm_shift", + "arg": 53, + "edited": [ + "norm_mean[53]" + ] + } + }, + "num_envs": 1024, + "seed": 500, + "success_rate": 0.04296875, + "mean_return": 14.348074913024902, + "mean_steps_to_success": 103.2272720336914, + "terminated_rate": 0.166015625, + "mean_max_progress": null, + "eval_seconds": 18.253344774246216, + "mean_goals": 0.04296875, + "median_goals": 0.0, + "drop_rate": 0.166015625, + "mean_survival_frac": 0.9089257717132568, + "mean_best_rot_err_first_goal": 0.7649253606796265, + "trajectory_file": "policies/allegro/repair/L/prompt_rollouts/norm_shift_53_seed500.npz" + }, + { + "policy": "policies/allegro/repair/L/damaged/swap_act_4_7.npz", + "meta": { + "source": "downloads/nvidia_pretrained/Isaac-Repose-Cube-Allegro-Direct-v0/checkpoint.pth", + "epoch": 4650, + "frame": 609484800, + "task": "allegro", + "kind": "broken", + "env_id": "Isaac-Repose-Cube-Allegro-Direct-v0", + "expert": "policies/allegro/nvidia/pretrained.npz", + "damage": { + "kind": "swap_act", + "arg": [ + 4, + 7 + ], + "edited": [ + "mu_W rows 4,7", + "mu_b[4],[7]" + ] + } + }, + "num_envs": 1024, + "seed": 500, + "success_rate": 0.140625, + "mean_return": 80.64712524414062, + "mean_steps_to_success": 90.5138931274414, + "terminated_rate": 0.23046875, + "mean_max_progress": null, + "eval_seconds": 15.36475419998169, + "mean_goals": 0.1611328125, + "median_goals": 0.0, + "drop_rate": 0.23046875, + "mean_survival_frac": 0.8582682609558105, + "mean_best_rot_err_first_goal": 1.4266011714935303, + "trajectory_file": "policies/allegro/repair/L/prompt_rollouts/swap_act_4_7_seed500.npz" + }, + { + "policy": "policies/allegro/repair/L/natural/early_ep200.npz", + "meta": { + "source": "logs/rl_games/allegro_hand/allegro_L_s0_early/nn/last_allegro_hand_ep_200_rew__174.68695_.pth", + "epoch": 200, + "frame": 26214400, + "task": "allegro", + "size": "L", + "run": "allegro_L_s0_early", + "kind": "ppo" + }, + "num_envs": 1024, + "seed": 500, + "success_rate": 0.19140625, + "mean_return": 152.01473999023438, + "mean_steps_to_success": 79.93877410888672, + "terminated_rate": 0.01953125, + "mean_max_progress": null, + "eval_seconds": 13.584335565567017, + "mean_goals": 0.2236328125, + "median_goals": 0.0, + "drop_rate": 0.01953125, + "mean_survival_frac": 0.9846549034118652, + "mean_best_rot_err_first_goal": 1.3968404531478882, + "trajectory_file": "policies/allegro/repair/L/prompt_rollouts/early_ep200_seed500.npz" + }, + { + "policy": "policies/allegro/repair/L/natural/early_ep50.npz", + "meta": { + "source": "logs/rl_games/allegro_hand/allegro_L_s0_early/nn/last_allegro_hand_ep_50_rew_51.205147.pth", + "epoch": 50, + "frame": 6553600, + "task": "allegro", + "size": "L", + "run": "allegro_L_s0_early", + "kind": "ppo" + }, + "num_envs": 1024, + "seed": 500, + "success_rate": 0.017578125, + "mean_return": 75.44386291503906, + "mean_steps_to_success": 43.05555725097656, + "terminated_rate": 0.013671875, + "mean_max_progress": null, + "eval_seconds": 12.63112759590149, + "mean_goals": 0.0185546875, + "median_goals": 0.0, + "drop_rate": 0.013671875, + "mean_survival_frac": 0.9850423336029053, + "mean_best_rot_err_first_goal": 1.9493906497955322, + "trajectory_file": "policies/allegro/repair/L/prompt_rollouts/early_ep50_seed500.npz" + } +] diff --git a/metadata/data/robot_policy/baselines/det1.json b/metadata/data/robot_policy/baselines/det1.json new file mode 100644 index 0000000000000000000000000000000000000000..d0598645736b51e187a627b1c3b0457240d4ded6 --- /dev/null +++ b/metadata/data/robot_policy/baselines/det1.json @@ -0,0 +1,41 @@ +[ + { + "policy": "policies/franka/nvidia/pretrained.npz", + "meta": { + "source": "downloads/nvidia_pretrained/Isaac-Franka-Cabinet-Direct-v0/checkpoint.pth", + "epoch": 1475, + "frame": 96665600, + "task": "franka", + "kind": "nvidia_pretrained", + "env_id": "Isaac-Franka-Cabinet-Direct-v0" + }, + "num_envs": 4096, + "seed": 100000, + "success_rate": 0.752197265625, + "mean_return": 761.6556396484375, + "mean_steps_to_success": 420.9120178222656, + "terminated_rate": 0.4423828125, + "mean_max_progress": 0.31918248534202576, + "eval_seconds": 18.998069524765015 + }, + { + "policy": "policies/franka/ppo/franka_M_s0/ep00100.npz", + "meta": { + "source": "logs/rl_games/franka_cabinet_direct/franka_M_s0/nn/last_franka_cabinet_direct_ep_100_rew_1246.1267.pth", + "epoch": 100, + "frame": 6553600, + "task": "franka", + "size": "M", + "run": "franka_M_s0", + "kind": "ppo" + }, + "num_envs": 4096, + "seed": 100000, + "success_rate": 0.546875, + "mean_return": 1460.4881591796875, + "mean_steps_to_success": 213.6062469482422, + "terminated_rate": 0.1962890625, + "mean_max_progress": 0.20347639918327332, + "eval_seconds": 18.99557375907898 + } +] diff --git a/metadata/data/robot_policy/baselines/det2.json b/metadata/data/robot_policy/baselines/det2.json new file mode 100644 index 0000000000000000000000000000000000000000..298c7f5d43dc13ed3370d94df2fa7121ebdd112e --- /dev/null +++ b/metadata/data/robot_policy/baselines/det2.json @@ -0,0 +1,41 @@ +[ + { + "policy": "policies/franka/ppo/franka_M_s0/ep00100.npz", + "meta": { + "source": "logs/rl_games/franka_cabinet_direct/franka_M_s0/nn/last_franka_cabinet_direct_ep_100_rew_1246.1267.pth", + "epoch": 100, + "frame": 6553600, + "task": "franka", + "size": "M", + "run": "franka_M_s0", + "kind": "ppo" + }, + "num_envs": 4096, + "seed": 100000, + "success_rate": 0.54541015625, + "mean_return": 1458.6976318359375, + "mean_steps_to_success": 213.76678466796875, + "terminated_rate": 0.197021484375, + "mean_max_progress": 0.2033017873764038, + "eval_seconds": 20.193190813064575 + }, + { + "policy": "policies/franka/nvidia/pretrained.npz", + "meta": { + "source": "downloads/nvidia_pretrained/Isaac-Franka-Cabinet-Direct-v0/checkpoint.pth", + "epoch": 1475, + "frame": 96665600, + "task": "franka", + "kind": "nvidia_pretrained", + "env_id": "Isaac-Franka-Cabinet-Direct-v0" + }, + "num_envs": 4096, + "seed": 100000, + "success_rate": 0.755859375, + "mean_return": 761.3577270507812, + "mean_steps_to_success": 421.18023681640625, + "terminated_rate": 0.4501953125, + "mean_max_progress": 0.3206583857536316, + "eval_seconds": 17.891871690750122 + } +] diff --git a/metadata/data/robot_policy/baselines/franka_broken_heldout.json b/metadata/data/robot_policy/baselines/franka_broken_heldout.json new file mode 100644 index 0000000000000000000000000000000000000000..21d1a567274348f60283df5c5b6ad007c61739ed --- /dev/null +++ b/metadata/data/robot_policy/baselines/franka_broken_heldout.json @@ -0,0 +1,283 @@ +[ + { + "policy": "policies/franka/nvidia/pretrained.npz", + "meta": { + "source": "downloads/nvidia_pretrained/Isaac-Franka-Cabinet-Direct-v0/checkpoint.pth", + "epoch": 1475, + "frame": 96665600, + "task": "franka", + "kind": "nvidia_pretrained", + "env_id": "Isaac-Franka-Cabinet-Direct-v0" + }, + "num_envs": 4096, + "seed": 100000, + "success_rate": 0.5712890625, + "mean_return": 1133.519287109375, + "mean_steps_to_success": 200.50299072265625, + "eval_seconds": 22.067054986953735 + }, + { + "policy": "policies/franka/broken/nvidia/dead_units_0_0p5.npz", + "meta": { + "source": "downloads/nvidia_pretrained/Isaac-Franka-Cabinet-Direct-v0/checkpoint.pth", + "epoch": 1475, + "frame": 96665600, + "task": "franka", + "kind": "broken", + "env_id": "Isaac-Franka-Cabinet-Direct-v0", + "expert": "policies/franka/nvidia/pretrained.npz", + "damage": { + "kind": "dead_units", + "arg": [ + 0, + 0.5 + ], + "edited": [ + "L0_W rows / L0_b [0, 1, 2, 3, 4, 5, 10, 12, 13, 14, 18, 21, 23, 24, 26, 28, 35, 36, 40, 41, 45, 49, 50, 53, 55, 56, 58, 60, 62, 66, 68, 69, 70, 71, 72, 73, 74, 79, 80, 82, 83, 85, 86, 88, 90, 91, 98, 101, 104, 105, 108, 109, 112, 113, 115, 116, 120, 122, 127, 130, 132, 134, 136, 137, 138, 139, 140, 141, 145, 146, 147, 148, 153, 155, 158, 159, 165, 166, 167, 168, 170, 172, 174, 177, 180, 181, 182, 184, 185, 186, 187, 189, 190, 192, 194, 195, 198, 200, 203, 204, 205, 206, 207, 209, 211, 213, 217, 218, 223, 224, 226, 228, 229, 234, 235, 236, 237, 239, 240, 241, 242, 243, 246, 248, 249, 250, 251, 254]" + ] + } + }, + "num_envs": 4096, + "seed": 100000, + "success_rate": 0.25439453125, + "mean_return": 759.34033203125, + "mean_steps_to_success": 142.8704376220703, + "eval_seconds": 19.38523769378662 + }, + { + "policy": "policies/franka/broken/nvidia/dead_units_1_0p5.npz", + "meta": { + "source": "downloads/nvidia_pretrained/Isaac-Franka-Cabinet-Direct-v0/checkpoint.pth", + "epoch": 1475, + "frame": 96665600, + "task": "franka", + "kind": "broken", + "env_id": "Isaac-Franka-Cabinet-Direct-v0", + "expert": "policies/franka/nvidia/pretrained.npz", + "damage": { + "kind": "dead_units", + "arg": [ + 1, + 0.5 + ], + "edited": [ + "L1_W rows / L1_b [1, 2, 4, 7, 8, 9, 11, 12, 14, 15, 19, 21, 22, 30, 32, 33, 35, 36, 40, 41, 43, 44, 50, 52, 53, 55, 56, 60, 61, 66, 67, 69, 73, 76, 77, 80, 81, 83, 85, 86, 87, 90, 91, 92, 95, 96, 97, 98, 99, 102, 103, 105, 108, 109, 112, 113, 114, 116, 117, 118, 120, 123, 125, 127]" + ] + } + }, + "num_envs": 4096, + "seed": 100000, + "success_rate": 0.000732421875, + "mean_return": -35.481605529785156, + "mean_steps_to_success": 176.33334350585938, + "eval_seconds": 15.807976007461548 + }, + { + "policy": "policies/franka/broken/nvidia/flip_obs_18.npz", + "meta": { + "source": "downloads/nvidia_pretrained/Isaac-Franka-Cabinet-Direct-v0/checkpoint.pth", + "epoch": 1475, + "frame": 96665600, + "task": "franka", + "kind": "broken", + "env_id": "Isaac-Franka-Cabinet-Direct-v0", + "expert": "policies/franka/nvidia/pretrained.npz", + "damage": { + "kind": "flip_obs", + "arg": 18, + "edited": [ + "L0_W[:, 18]" + ] + } + }, + "num_envs": 4096, + "seed": 100000, + "success_rate": 0.625, + "mean_return": 453.0425109863281, + "mean_steps_to_success": 211.9792938232422, + "eval_seconds": 22.493717432022095 + }, + { + "policy": "policies/franka/broken/nvidia/flip_obs_20.npz", + "meta": { + "source": "downloads/nvidia_pretrained/Isaac-Franka-Cabinet-Direct-v0/checkpoint.pth", + "epoch": 1475, + "frame": 96665600, + "task": "franka", + "kind": "broken", + "env_id": "Isaac-Franka-Cabinet-Direct-v0", + "expert": "policies/franka/nvidia/pretrained.npz", + "damage": { + "kind": "flip_obs", + "arg": 20, + "edited": [ + "L0_W[:, 20]" + ] + } + }, + "num_envs": 4096, + "seed": 100000, + "success_rate": 0.00439453125, + "mean_return": -101.8251724243164, + "mean_steps_to_success": 209.7777862548828, + "eval_seconds": 20.79281258583069 + }, + { + "policy": "policies/franka/broken/nvidia/noise_0p5.npz", + "meta": { + "source": "downloads/nvidia_pretrained/Isaac-Franka-Cabinet-Direct-v0/checkpoint.pth", + "epoch": 1475, + "frame": 96665600, + "task": "franka", + "kind": "broken", + "env_id": "Isaac-Franka-Cabinet-Direct-v0", + "expert": "policies/franka/nvidia/pretrained.npz", + "damage": { + "kind": "noise", + "arg": 0.5, + "edited": [ + "all W (Gaussian)" + ] + } + }, + "num_envs": 4096, + "seed": 100000, + "success_rate": 0.0048828125, + "mean_return": -56.56855773925781, + "mean_steps_to_success": 95.55000305175781, + "eval_seconds": 21.760616302490234 + }, + { + "policy": "policies/franka/broken/nvidia/noise_1p0.npz", + "meta": { + "source": "downloads/nvidia_pretrained/Isaac-Franka-Cabinet-Direct-v0/checkpoint.pth", + "epoch": 1475, + "frame": 96665600, + "task": "franka", + "kind": "broken", + "env_id": "Isaac-Franka-Cabinet-Direct-v0", + "expert": "policies/franka/nvidia/pretrained.npz", + "damage": { + "kind": "noise", + "arg": 1.0, + "edited": [ + "all W (Gaussian)" + ] + } + }, + "num_envs": 4096, + "seed": 100000, + "success_rate": 0.002197265625, + "mean_return": 72.54507446289062, + "mean_steps_to_success": 142.0, + "eval_seconds": 22.997545957565308 + }, + { + "policy": "policies/franka/broken/nvidia/norm_shift_20.npz", + "meta": { + "source": "downloads/nvidia_pretrained/Isaac-Franka-Cabinet-Direct-v0/checkpoint.pth", + "epoch": 1475, + "frame": 96665600, + "task": "franka", + "kind": "broken", + "env_id": "Isaac-Franka-Cabinet-Direct-v0", + "expert": "policies/franka/nvidia/pretrained.npz", + "damage": { + "kind": "norm_shift", + "arg": 20, + "edited": [ + "norm_mean[20]" + ] + } + }, + "num_envs": 4096, + "seed": 100000, + "success_rate": 0.021484375, + "mean_return": 548.3388061523438, + "mean_steps_to_success": 153.34091186523438, + "eval_seconds": 20.428385972976685 + }, + { + "policy": "policies/franka/broken/nvidia/norm_shift_21.npz", + "meta": { + "source": "downloads/nvidia_pretrained/Isaac-Franka-Cabinet-Direct-v0/checkpoint.pth", + "epoch": 1475, + "frame": 96665600, + "task": "franka", + "kind": "broken", + "env_id": "Isaac-Franka-Cabinet-Direct-v0", + "expert": "policies/franka/nvidia/pretrained.npz", + "damage": { + "kind": "norm_shift", + "arg": 21, + "edited": [ + "norm_mean[21]" + ] + } + }, + "num_envs": 4096, + "seed": 100000, + "success_rate": 0.0, + "mean_return": 31.961219787597656, + "mean_steps_to_success": null, + "eval_seconds": 18.528737783432007 + }, + { + "policy": "policies/franka/broken/nvidia/swap_act_0_6.npz", + "meta": { + "source": "downloads/nvidia_pretrained/Isaac-Franka-Cabinet-Direct-v0/checkpoint.pth", + "epoch": 1475, + "frame": 96665600, + "task": "franka", + "kind": "broken", + "env_id": "Isaac-Franka-Cabinet-Direct-v0", + "expert": "policies/franka/nvidia/pretrained.npz", + "damage": { + "kind": "swap_act", + "arg": [ + 0, + 6 + ], + "edited": [ + "mu_W rows 0,6", + "mu_b[0],[6]" + ] + } + }, + "num_envs": 4096, + "seed": 100000, + "success_rate": 0.0, + "mean_return": -28.591876983642578, + "mean_steps_to_success": null, + "eval_seconds": 17.755794525146484 + }, + { + "policy": "policies/franka/broken/nvidia/swap_act_1_3.npz", + "meta": { + "source": "downloads/nvidia_pretrained/Isaac-Franka-Cabinet-Direct-v0/checkpoint.pth", + "epoch": 1475, + "frame": 96665600, + "task": "franka", + "kind": "broken", + "env_id": "Isaac-Franka-Cabinet-Direct-v0", + "expert": "policies/franka/nvidia/pretrained.npz", + "damage": { + "kind": "swap_act", + "arg": [ + 1, + 3 + ], + "edited": [ + "mu_W rows 1,3", + "mu_b[1],[3]" + ] + } + }, + "num_envs": 4096, + "seed": 100000, + "success_rate": 0.0, + "mean_return": -455.4483947753906, + "mean_steps_to_success": null, + "eval_seconds": 18.564955949783325 + } +] diff --git a/metadata/data/robot_policy/baselines/franka_broken_prompt.json b/metadata/data/robot_policy/baselines/franka_broken_prompt.json new file mode 100644 index 0000000000000000000000000000000000000000..304c8d774b4f5730f51a2803027b462af784c68f --- /dev/null +++ b/metadata/data/robot_policy/baselines/franka_broken_prompt.json @@ -0,0 +1,294 @@ +[ + { + "policy": "policies/franka/nvidia/pretrained.npz", + "meta": { + "source": "downloads/nvidia_pretrained/Isaac-Franka-Cabinet-Direct-v0/checkpoint.pth", + "epoch": 1475, + "frame": 96665600, + "task": "franka", + "kind": "nvidia_pretrained", + "env_id": "Isaac-Franka-Cabinet-Direct-v0" + }, + "num_envs": 1024, + "seed": 500, + "success_rate": 0.5439453125, + "mean_return": 1166.7724609375, + "mean_steps_to_success": 201.77200317382812, + "eval_seconds": 17.535452842712402, + "trajectory_file": "policies/franka/broken/nvidia/rollouts/pretrained_seed500.npz" + }, + { + "policy": "policies/franka/broken/nvidia/dead_units_0_0p5.npz", + "meta": { + "source": "downloads/nvidia_pretrained/Isaac-Franka-Cabinet-Direct-v0/checkpoint.pth", + "epoch": 1475, + "frame": 96665600, + "task": "franka", + "kind": "broken", + "env_id": "Isaac-Franka-Cabinet-Direct-v0", + "expert": "policies/franka/nvidia/pretrained.npz", + "damage": { + "kind": "dead_units", + "arg": [ + 0, + 0.5 + ], + "edited": [ + "L0_W rows / L0_b [0, 1, 2, 3, 4, 5, 10, 12, 13, 14, 18, 21, 23, 24, 26, 28, 35, 36, 40, 41, 45, 49, 50, 53, 55, 56, 58, 60, 62, 66, 68, 69, 70, 71, 72, 73, 74, 79, 80, 82, 83, 85, 86, 88, 90, 91, 98, 101, 104, 105, 108, 109, 112, 113, 115, 116, 120, 122, 127, 130, 132, 134, 136, 137, 138, 139, 140, 141, 145, 146, 147, 148, 153, 155, 158, 159, 165, 166, 167, 168, 170, 172, 174, 177, 180, 181, 182, 184, 185, 186, 187, 189, 190, 192, 194, 195, 198, 200, 203, 204, 205, 206, 207, 209, 211, 213, 217, 218, 223, 224, 226, 228, 229, 234, 235, 236, 237, 239, 240, 241, 242, 243, 246, 248, 249, 250, 251, 254]" + ] + } + }, + "num_envs": 1024, + "seed": 500, + "success_rate": 0.2490234375, + "mean_return": 752.35546875, + "mean_steps_to_success": 145.87451171875, + "eval_seconds": 15.260829448699951, + "trajectory_file": "policies/franka/broken/nvidia/rollouts/dead_units_0_0p5_seed500.npz" + }, + { + "policy": "policies/franka/broken/nvidia/dead_units_1_0p5.npz", + "meta": { + "source": "downloads/nvidia_pretrained/Isaac-Franka-Cabinet-Direct-v0/checkpoint.pth", + "epoch": 1475, + "frame": 96665600, + "task": "franka", + "kind": "broken", + "env_id": "Isaac-Franka-Cabinet-Direct-v0", + "expert": "policies/franka/nvidia/pretrained.npz", + "damage": { + "kind": "dead_units", + "arg": [ + 1, + 0.5 + ], + "edited": [ + "L1_W rows / L1_b [1, 2, 4, 7, 8, 9, 11, 12, 14, 15, 19, 21, 22, 30, 32, 33, 35, 36, 40, 41, 43, 44, 50, 52, 53, 55, 56, 60, 61, 66, 67, 69, 73, 76, 77, 80, 81, 83, 85, 86, 87, 90, 91, 92, 95, 96, 97, 98, 99, 102, 103, 105, 108, 109, 112, 113, 114, 116, 117, 118, 120, 123, 125, 127]" + ] + } + }, + "num_envs": 1024, + "seed": 500, + "success_rate": 0.0, + "mean_return": -38.10558319091797, + "mean_steps_to_success": null, + "eval_seconds": 11.591758728027344, + "trajectory_file": "policies/franka/broken/nvidia/rollouts/dead_units_1_0p5_seed500.npz" + }, + { + "policy": "policies/franka/broken/nvidia/flip_obs_18.npz", + "meta": { + "source": "downloads/nvidia_pretrained/Isaac-Franka-Cabinet-Direct-v0/checkpoint.pth", + "epoch": 1475, + "frame": 96665600, + "task": "franka", + "kind": "broken", + "env_id": "Isaac-Franka-Cabinet-Direct-v0", + "expert": "policies/franka/nvidia/pretrained.npz", + "damage": { + "kind": "flip_obs", + "arg": 18, + "edited": [ + "L0_W[:, 18]" + ] + } + }, + "num_envs": 1024, + "seed": 500, + "success_rate": 0.6474609375, + "mean_return": 455.84600830078125, + "mean_steps_to_success": 212.5098114013672, + "eval_seconds": 17.358514547348022, + "trajectory_file": "policies/franka/broken/nvidia/rollouts/flip_obs_18_seed500.npz" + }, + { + "policy": "policies/franka/broken/nvidia/flip_obs_20.npz", + "meta": { + "source": "downloads/nvidia_pretrained/Isaac-Franka-Cabinet-Direct-v0/checkpoint.pth", + "epoch": 1475, + "frame": 96665600, + "task": "franka", + "kind": "broken", + "env_id": "Isaac-Franka-Cabinet-Direct-v0", + "expert": "policies/franka/nvidia/pretrained.npz", + "damage": { + "kind": "flip_obs", + "arg": 20, + "edited": [ + "L0_W[:, 20]" + ] + } + }, + "num_envs": 1024, + "seed": 500, + "success_rate": 0.005859375, + "mean_return": -85.5876693725586, + "mean_steps_to_success": 285.0, + "eval_seconds": 16.754355907440186, + "trajectory_file": "policies/franka/broken/nvidia/rollouts/flip_obs_20_seed500.npz" + }, + { + "policy": "policies/franka/broken/nvidia/noise_0p5.npz", + "meta": { + "source": "downloads/nvidia_pretrained/Isaac-Franka-Cabinet-Direct-v0/checkpoint.pth", + "epoch": 1475, + "frame": 96665600, + "task": "franka", + "kind": "broken", + "env_id": "Isaac-Franka-Cabinet-Direct-v0", + "expert": "policies/franka/nvidia/pretrained.npz", + "damage": { + "kind": "noise", + "arg": 0.5, + "edited": [ + "all W (Gaussian)" + ] + } + }, + "num_envs": 1024, + "seed": 500, + "success_rate": 0.009765625, + "mean_return": -48.17366409301758, + "mean_steps_to_success": 66.80000305175781, + "eval_seconds": 16.316949129104614, + "trajectory_file": "policies/franka/broken/nvidia/rollouts/noise_0p5_seed500.npz" + }, + { + "policy": "policies/franka/broken/nvidia/noise_1p0.npz", + "meta": { + "source": "downloads/nvidia_pretrained/Isaac-Franka-Cabinet-Direct-v0/checkpoint.pth", + "epoch": 1475, + "frame": 96665600, + "task": "franka", + "kind": "broken", + "env_id": "Isaac-Franka-Cabinet-Direct-v0", + "expert": "policies/franka/nvidia/pretrained.npz", + "damage": { + "kind": "noise", + "arg": 1.0, + "edited": [ + "all W (Gaussian)" + ] + } + }, + "num_envs": 1024, + "seed": 500, + "success_rate": 0.0, + "mean_return": 66.05709075927734, + "mean_steps_to_success": null, + "eval_seconds": 16.963040351867676, + "trajectory_file": "policies/franka/broken/nvidia/rollouts/noise_1p0_seed500.npz" + }, + { + "policy": "policies/franka/broken/nvidia/norm_shift_20.npz", + "meta": { + "source": "downloads/nvidia_pretrained/Isaac-Franka-Cabinet-Direct-v0/checkpoint.pth", + "epoch": 1475, + "frame": 96665600, + "task": "franka", + "kind": "broken", + "env_id": "Isaac-Franka-Cabinet-Direct-v0", + "expert": "policies/franka/nvidia/pretrained.npz", + "damage": { + "kind": "norm_shift", + "arg": 20, + "edited": [ + "norm_mean[20]" + ] + } + }, + "num_envs": 1024, + "seed": 500, + "success_rate": 0.0205078125, + "mean_return": 541.6852416992188, + "mean_steps_to_success": 177.23809814453125, + "eval_seconds": 17.054648876190186, + "trajectory_file": "policies/franka/broken/nvidia/rollouts/norm_shift_20_seed500.npz" + }, + { + "policy": "policies/franka/broken/nvidia/norm_shift_21.npz", + "meta": { + "source": "downloads/nvidia_pretrained/Isaac-Franka-Cabinet-Direct-v0/checkpoint.pth", + "epoch": 1475, + "frame": 96665600, + "task": "franka", + "kind": "broken", + "env_id": "Isaac-Franka-Cabinet-Direct-v0", + "expert": "policies/franka/nvidia/pretrained.npz", + "damage": { + "kind": "norm_shift", + "arg": 21, + "edited": [ + "norm_mean[21]" + ] + } + }, + "num_envs": 1024, + "seed": 500, + "success_rate": 0.0, + "mean_return": 32.291015625, + "mean_steps_to_success": null, + "eval_seconds": 15.145379066467285, + "trajectory_file": "policies/franka/broken/nvidia/rollouts/norm_shift_21_seed500.npz" + }, + { + "policy": "policies/franka/broken/nvidia/swap_act_0_6.npz", + "meta": { + "source": "downloads/nvidia_pretrained/Isaac-Franka-Cabinet-Direct-v0/checkpoint.pth", + "epoch": 1475, + "frame": 96665600, + "task": "franka", + "kind": "broken", + "env_id": "Isaac-Franka-Cabinet-Direct-v0", + "expert": "policies/franka/nvidia/pretrained.npz", + "damage": { + "kind": "swap_act", + "arg": [ + 0, + 6 + ], + "edited": [ + "mu_W rows 0,6", + "mu_b[0],[6]" + ] + } + }, + "num_envs": 1024, + "seed": 500, + "success_rate": 0.0, + "mean_return": -30.72944450378418, + "mean_steps_to_success": null, + "eval_seconds": 13.886693000793457, + "trajectory_file": "policies/franka/broken/nvidia/rollouts/swap_act_0_6_seed500.npz" + }, + { + "policy": "policies/franka/broken/nvidia/swap_act_1_3.npz", + "meta": { + "source": "downloads/nvidia_pretrained/Isaac-Franka-Cabinet-Direct-v0/checkpoint.pth", + "epoch": 1475, + "frame": 96665600, + "task": "franka", + "kind": "broken", + "env_id": "Isaac-Franka-Cabinet-Direct-v0", + "expert": "policies/franka/nvidia/pretrained.npz", + "damage": { + "kind": "swap_act", + "arg": [ + 1, + 3 + ], + "edited": [ + "mu_W rows 1,3", + "mu_b[1],[3]" + ] + } + }, + "num_envs": 1024, + "seed": 500, + "success_rate": 0.0, + "mean_return": -457.0911865234375, + "mean_steps_to_success": null, + "eval_seconds": 15.235822916030884, + "trajectory_file": "policies/franka/broken/nvidia/rollouts/swap_act_1_3_seed500.npz" + } +] diff --git a/metadata/data/robot_policy/baselines/franka_nvidia_pretrained.json b/metadata/data/robot_policy/baselines/franka_nvidia_pretrained.json new file mode 100644 index 0000000000000000000000000000000000000000..db5141683c179575cc9f80b2ff1dae96cdb88a3a --- /dev/null +++ b/metadata/data/robot_policy/baselines/franka_nvidia_pretrained.json @@ -0,0 +1,19 @@ +[ + { + "policy": "policies/franka/nvidia/pretrained.npz", + "meta": { + "source": "downloads/nvidia_pretrained/Isaac-Franka-Cabinet-Direct-v0/checkpoint.pth", + "epoch": 1475, + "frame": 96665600, + "task": "franka", + "kind": "nvidia_pretrained", + "env_id": "Isaac-Franka-Cabinet-Direct-v0" + }, + "num_envs": 4096, + "seed": 100000, + "success_rate": 0.5712890625, + "mean_return": 1133.519287109375, + "mean_steps_to_success": 200.50299072265625, + "eval_seconds": 22.057674884796143 + } +] diff --git a/metadata/data/robot_policy/baselines/franka_ppo_early_heldout.json b/metadata/data/robot_policy/baselines/franka_ppo_early_heldout.json new file mode 100644 index 0000000000000000000000000000000000000000..57b796030a99c4af81699cb48fd62b20c3dfb62d --- /dev/null +++ b/metadata/data/robot_policy/baselines/franka_ppo_early_heldout.json @@ -0,0 +1,202 @@ +[ + { + "policy": "policies/franka/ppo/franka_M_s0_early/ep00015.npz", + "meta": { + "source": "logs/rl_games/franka_cabinet_direct/franka_M_s0_early/nn/last_franka_cabinet_direct_ep_15_rew_339.6565.pth", + "epoch": 15, + "frame": 983040, + "task": "franka", + "size": "M", + "run": "franka_M_s0_early", + "kind": "ppo" + }, + "num_envs": 4096, + "seed": 100000, + "success_rate": 0.0068359375, + "mean_return": 631.5262451171875, + "mean_steps_to_success": 129.7857208251953, + "terminated_rate": 0.006103515625, + "mean_max_progress": 0.025191131979227066, + "eval_seconds": 17.99227285385132 + }, + { + "policy": "policies/franka/ppo/franka_M_s0_early/ep00020.npz", + "meta": { + "source": "logs/rl_games/franka_cabinet_direct/franka_M_s0_early/nn/last_franka_cabinet_direct_ep_20_rew_422.66306.pth", + "epoch": 20, + "frame": 1310720, + "task": "franka", + "size": "M", + "run": "franka_M_s0_early", + "kind": "ppo" + }, + "num_envs": 4096, + "seed": 100000, + "success_rate": 0.1572265625, + "mean_return": 960.833251953125, + "mean_steps_to_success": 132.1645965576172, + "terminated_rate": 0.099609375, + "mean_max_progress": 0.11936343461275101, + "eval_seconds": 18.72763705253601 + }, + { + "policy": "policies/franka/ppo/franka_M_s0_early/ep00025.npz", + "meta": { + "source": "logs/rl_games/franka_cabinet_direct/franka_M_s0_early/nn/last_franka_cabinet_direct_ep_25_rew_412.2454.pth", + "epoch": 25, + "frame": 1638400, + "task": "franka", + "size": "M", + "run": "franka_M_s0_early", + "kind": "ppo" + }, + "num_envs": 4096, + "seed": 100000, + "success_rate": 0.650634765625, + "mean_return": 799.10009765625, + "mean_steps_to_success": 136.39212036132812, + "terminated_rate": 0.6396484375, + "mean_max_progress": 0.2772047519683838, + "eval_seconds": 21.51799249649048 + }, + { + "policy": "policies/franka/ppo/franka_M_s0_early/ep00030.npz", + "meta": { + "source": "logs/rl_games/franka_cabinet_direct/franka_M_s0_early/nn/last_franka_cabinet_direct_ep_30_rew_413.22134.pth", + "epoch": 30, + "frame": 1966080, + "task": "franka", + "size": "M", + "run": "franka_M_s0_early", + "kind": "ppo" + }, + "num_envs": 4096, + "seed": 100000, + "success_rate": 0.50830078125, + "mean_return": 1058.97412109375, + "mean_steps_to_success": 127.87799835205078, + "terminated_rate": 0.424560546875, + "mean_max_progress": 0.24507778882980347, + "eval_seconds": 20.284687280654907 + }, + { + "policy": "policies/franka/ppo/franka_M_s0_early/ep00035.npz", + "meta": { + "source": "logs/rl_games/franka_cabinet_direct/franka_M_s0_early/nn/last_franka_cabinet_direct_ep_35_rew_399.16858.pth", + "epoch": 35, + "frame": 2293760, + "task": "franka", + "size": "M", + "run": "franka_M_s0_early", + "kind": "ppo" + }, + "num_envs": 4096, + "seed": 100000, + "success_rate": 0.679443359375, + "mean_return": 1001.3622436523438, + "mean_steps_to_success": 145.44448852539062, + "terminated_rate": 0.555419921875, + "mean_max_progress": 0.29956698417663574, + "eval_seconds": 20.640437364578247 + }, + { + "policy": "policies/franka/ppo/franka_M_s0_early/ep00040.npz", + "meta": { + "source": "logs/rl_games/franka_cabinet_direct/franka_M_s0_early/nn/last_franka_cabinet_direct_ep_40_rew_394.78815.pth", + "epoch": 40, + "frame": 2621440, + "task": "franka", + "size": "M", + "run": "franka_M_s0_early", + "kind": "ppo" + }, + "num_envs": 4096, + "seed": 100000, + "success_rate": 0.22265625, + "mean_return": 1290.145751953125, + "mean_steps_to_success": 173.25987243652344, + "terminated_rate": 0.215576171875, + "mean_max_progress": 0.23516984283924103, + "eval_seconds": 19.661967515945435 + }, + { + "policy": "policies/franka/ppo/franka_M_s0_early/ep00045.npz", + "meta": { + "source": "logs/rl_games/franka_cabinet_direct/franka_M_s0_early/nn/last_franka_cabinet_direct_ep_45_rew_563.78357.pth", + "epoch": 45, + "frame": 2949120, + "task": "franka", + "size": "M", + "run": "franka_M_s0_early", + "kind": "ppo" + }, + "num_envs": 4096, + "seed": 100000, + "success_rate": 0.626220703125, + "mean_return": 2318.333251953125, + "mean_steps_to_success": 231.72671508789062, + "terminated_rate": 0.034423828125, + "mean_max_progress": 0.32679980993270874, + "eval_seconds": 20.134300470352173 + }, + { + "policy": "policies/franka/ppo/franka_M_s0_early/ep00050.npz", + "meta": { + "source": "logs/rl_games/franka_cabinet_direct/franka_M_s0_early/nn/last_franka_cabinet_direct_ep_50_rew_806.8821.pth", + "epoch": 50, + "frame": 3276800, + "task": "franka", + "size": "M", + "run": "franka_M_s0_early", + "kind": "ppo" + }, + "num_envs": 4096, + "seed": 100000, + "success_rate": 0.88330078125, + "mean_return": 2289.08056640625, + "mean_steps_to_success": 186.4969482421875, + "terminated_rate": 0.120361328125, + "mean_max_progress": 0.35370996594429016, + "eval_seconds": 20.132587671279907 + }, + { + "policy": "policies/franka/ppo/franka_M_s0_early/ep00055.npz", + "meta": { + "source": "logs/rl_games/franka_cabinet_direct/franka_M_s0_early/nn/last_franka_cabinet_direct_ep_55_rew_1132.288.pth", + "epoch": 55, + "frame": 3604480, + "task": "franka", + "size": "M", + "run": "franka_M_s0_early", + "kind": "ppo" + }, + "num_envs": 4096, + "seed": 100000, + "success_rate": 0.81201171875, + "mean_return": 2199.71533203125, + "mean_steps_to_success": 214.09771728515625, + "terminated_rate": 0.0009765625, + "mean_max_progress": 0.2970821261405945, + "eval_seconds": 19.812252283096313 + }, + { + "policy": "policies/franka/ppo/franka_M_s0_early/ep00060.npz", + "meta": { + "source": "logs/rl_games/franka_cabinet_direct/franka_M_s0_early/nn/last_franka_cabinet_direct_ep_60_rew__1348.8466_.pth", + "epoch": 60, + "frame": 3932160, + "task": "franka", + "size": "M", + "run": "franka_M_s0_early", + "kind": "ppo" + }, + "num_envs": 4096, + "seed": 100000, + "success_rate": 0.947021484375, + "mean_return": 2365.51904296875, + "mean_steps_to_success": 176.41171264648438, + "terminated_rate": 0.0009765625, + "mean_max_progress": 0.329230397939682, + "eval_seconds": 19.736530303955078 + } +] diff --git a/metadata/data/robot_policy/baselines/franka_ppo_mid.json b/metadata/data/robot_policy/baselines/franka_ppo_mid.json new file mode 100644 index 0000000000000000000000000000000000000000..1aae9982ef33deac594e2e573667bc816e85551d --- /dev/null +++ b/metadata/data/robot_policy/baselines/franka_ppo_mid.json @@ -0,0 +1,182 @@ +[ + { + "policy": "policies/franka/ppo/franka_M_s0/ep00100.npz", + "meta": { + "source": "logs/rl_games/franka_cabinet_direct/franka_M_s0/nn/last_franka_cabinet_direct_ep_100_rew_1246.1267.pth", + "epoch": 100, + "frame": 6553600, + "task": "franka", + "size": "M", + "run": "franka_M_s0", + "kind": "ppo" + }, + "num_envs": 4096, + "seed": 100000, + "success_rate": 0.280029296875, + "mean_return": 1541.2452392578125, + "mean_steps_to_success": 356.34002685546875, + "eval_seconds": 20.301416635513306 + }, + { + "policy": "policies/franka/ppo/franka_M_s0/ep00200.npz", + "meta": { + "source": "logs/rl_games/franka_cabinet_direct/franka_M_s0/nn/last_franka_cabinet_direct_ep_200_rew_2597.755.pth", + "epoch": 200, + "frame": 13107200, + "task": "franka", + "size": "M", + "run": "franka_M_s0", + "kind": "ppo" + }, + "num_envs": 4096, + "seed": 100000, + "success_rate": 0.003662109375, + "mean_return": 2713.700439453125, + "mean_steps_to_success": 178.4666748046875, + "eval_seconds": 19.417638063430786 + }, + { + "policy": "policies/franka/ppo/franka_M_s0/ep00300.npz", + "meta": { + "source": "logs/rl_games/franka_cabinet_direct/franka_M_s0/nn/last_franka_cabinet_direct_ep_300_rew_2885.3364.pth", + "epoch": 300, + "frame": 19660800, + "task": "franka", + "size": "M", + "run": "franka_M_s0", + "kind": "ppo" + }, + "num_envs": 4096, + "seed": 100000, + "success_rate": 0.001708984375, + "mean_return": 3076.8583984375, + "mean_steps_to_success": 262.2857360839844, + "eval_seconds": 19.426570653915405 + }, + { + "policy": "policies/franka/ppo/franka_M_s0/ep00400.npz", + "meta": { + "source": "logs/rl_games/franka_cabinet_direct/franka_M_s0/nn/last_franka_cabinet_direct_ep_400_rew_2902.9753.pth", + "epoch": 400, + "frame": 26214400, + "task": "franka", + "size": "M", + "run": "franka_M_s0", + "kind": "ppo" + }, + "num_envs": 4096, + "seed": 100000, + "success_rate": 0.00146484375, + "mean_return": 3053.713134765625, + "mean_steps_to_success": 269.0, + "eval_seconds": 19.337193965911865 + }, + { + "policy": "policies/franka/ppo/franka_M_s0/ep00500.npz", + "meta": { + "source": "logs/rl_games/franka_cabinet_direct/franka_M_s0/nn/last_franka_cabinet_direct_ep_500_rew_2984.3376.pth", + "epoch": 500, + "frame": 32768000, + "task": "franka", + "size": "M", + "run": "franka_M_s0", + "kind": "ppo" + }, + "num_envs": 4096, + "seed": 100000, + "success_rate": 0.001708984375, + "mean_return": 3048.0810546875, + "mean_steps_to_success": 347.0000305175781, + "eval_seconds": 19.28227972984314 + }, + { + "policy": "policies/franka/ppo/franka_XS_s0/ep00100.npz", + "meta": { + "source": "logs/rl_games/franka_cabinet_direct/franka_XS_s0/nn/last_franka_cabinet_direct_ep_100_rew_1864.7428.pth", + "epoch": 100, + "frame": 6553600, + "task": "franka", + "size": "XS", + "run": "franka_XS_s0", + "kind": "ppo" + }, + "num_envs": 4096, + "seed": 100000, + "success_rate": 0.08154296875, + "mean_return": 2241.016845703125, + "mean_steps_to_success": 347.3053894042969, + "eval_seconds": 18.820095777511597 + }, + { + "policy": "policies/franka/ppo/franka_XS_s0/ep00200.npz", + "meta": { + "source": "logs/rl_games/franka_cabinet_direct/franka_XS_s0/nn/last_franka_cabinet_direct_ep_200_rew_2835.899.pth", + "epoch": 200, + "frame": 13107200, + "task": "franka", + "size": "XS", + "run": "franka_XS_s0", + "kind": "ppo" + }, + "num_envs": 4096, + "seed": 100000, + "success_rate": 0.01123046875, + "mean_return": 3153.41796875, + "mean_steps_to_success": 205.0, + "eval_seconds": 18.30488109588623 + }, + { + "policy": "policies/franka/ppo/franka_XS_s0/ep00300.npz", + "meta": { + "source": "logs/rl_games/franka_cabinet_direct/franka_XS_s0/nn/last_franka_cabinet_direct_ep_300_rew_2949.1003.pth", + "epoch": 300, + "frame": 19660800, + "task": "franka", + "size": "XS", + "run": "franka_XS_s0", + "kind": "ppo" + }, + "num_envs": 4096, + "seed": 100000, + "success_rate": 0.0341796875, + "mean_return": 3110.974365234375, + "mean_steps_to_success": 128.27857971191406, + "eval_seconds": 18.859910249710083 + }, + { + "policy": "policies/franka/ppo/franka_XS_s0/ep00400.npz", + "meta": { + "source": "logs/rl_games/franka_cabinet_direct/franka_XS_s0/nn/last_franka_cabinet_direct_ep_400_rew_2984.3071.pth", + "epoch": 400, + "frame": 26214400, + "task": "franka", + "size": "XS", + "run": "franka_XS_s0", + "kind": "ppo" + }, + "num_envs": 4096, + "seed": 100000, + "success_rate": 0.02099609375, + "mean_return": 3162.31640625, + "mean_steps_to_success": 137.48837280273438, + "eval_seconds": 18.488566637039185 + }, + { + "policy": "policies/franka/ppo/franka_XS_s0/ep00500.npz", + "meta": { + "source": "logs/rl_games/franka_cabinet_direct/franka_XS_s0/nn/last_franka_cabinet_direct_ep_500_rew_3207.3481.pth", + "epoch": 500, + "frame": 32768000, + "task": "franka", + "size": "XS", + "run": "franka_XS_s0", + "kind": "ppo" + }, + "num_envs": 4096, + "seed": 100000, + "success_rate": 0.058349609375, + "mean_return": 3211.31005859375, + "mean_steps_to_success": 231.1004180908203, + "eval_seconds": 19.397409439086914 + } +] diff --git a/metadata/data/robot_policy/baselines/franka_ppo_v3_heldout.json b/metadata/data/robot_policy/baselines/franka_ppo_v3_heldout.json new file mode 100644 index 0000000000000000000000000000000000000000..c7d2282e5af5cea079e5dc96af8ca8d6c724c8c2 --- /dev/null +++ b/metadata/data/robot_policy/baselines/franka_ppo_v3_heldout.json @@ -0,0 +1,1221 @@ +[ + { + "policy": "policies/franka/nvidia/pretrained.npz", + "meta": { + "source": "downloads/nvidia_pretrained/Isaac-Franka-Cabinet-Direct-v0/checkpoint.pth", + "epoch": 1475, + "frame": 96665600, + "task": "franka", + "kind": "nvidia_pretrained", + "env_id": "Isaac-Franka-Cabinet-Direct-v0" + }, + "num_envs": 4096, + "seed": 100000, + "success_rate": 0.752197265625, + "mean_return": 761.6556396484375, + "mean_steps_to_success": 420.9120178222656, + "terminated_rate": 0.4423828125, + "mean_max_progress": 0.31918248534202576, + "eval_seconds": 18.9741632938385 + }, + { + "policy": "policies/franka/ppo/franka_M_s0/ep00050.npz", + "meta": { + "source": "logs/rl_games/franka_cabinet_direct/franka_M_s0/nn/last_franka_cabinet_direct_ep_50_rew_806.8821.pth", + "epoch": 50, + "frame": 3276800, + "task": "franka", + "size": "M", + "run": "franka_M_s0", + "kind": "ppo" + }, + "num_envs": 4096, + "seed": 100000, + "success_rate": 0.880859375, + "mean_return": 2286.53125, + "mean_steps_to_success": 186.26052856445312, + "terminated_rate": 0.12158203125, + "mean_max_progress": 0.353589802980423, + "eval_seconds": 19.81156015396118 + }, + { + "policy": "policies/franka/ppo/franka_M_s0/ep00100.npz", + "meta": { + "source": "logs/rl_games/franka_cabinet_direct/franka_M_s0/nn/last_franka_cabinet_direct_ep_100_rew_2268.4282.pth", + "epoch": 100, + "frame": 6553600, + "task": "franka", + "size": "M", + "run": "franka_M_s0", + "kind": "ppo" + }, + "num_envs": 4096, + "seed": 100000, + "success_rate": 0.99951171875, + "mean_return": 2758.3525390625, + "mean_steps_to_success": 127.68807983398438, + "terminated_rate": 0.091064453125, + "mean_max_progress": 0.3814341425895691, + "eval_seconds": 19.350318431854248 + }, + { + "policy": "policies/franka/ppo/franka_M_s0/ep00150.npz", + "meta": { + "source": "logs/rl_games/franka_cabinet_direct/franka_M_s0/nn/last_franka_cabinet_direct_ep_150_rew_2784.9514.pth", + "epoch": 150, + "frame": 9830400, + "task": "franka", + "size": "M", + "run": "franka_M_s0", + "kind": "ppo" + }, + "num_envs": 4096, + "seed": 100000, + "success_rate": 1.0, + "mean_return": 3093.965087890625, + "mean_steps_to_success": 117.641845703125, + "terminated_rate": 0.00732421875, + "mean_max_progress": 0.3696989417076111, + "eval_seconds": 18.311604261398315 + }, + { + "policy": "policies/franka/ppo/franka_M_s0/ep00200.npz", + "meta": { + "source": "logs/rl_games/franka_cabinet_direct/franka_M_s0/nn/last_franka_cabinet_direct_ep_200_rew_2955.6724.pth", + "epoch": 200, + "frame": 13107200, + "task": "franka", + "size": "M", + "run": "franka_M_s0", + "kind": "ppo" + }, + "num_envs": 4096, + "seed": 100000, + "success_rate": 1.0, + "mean_return": 3208.58984375, + "mean_steps_to_success": 124.888916015625, + "terminated_rate": 0.0, + "mean_max_progress": 0.3618812561035156, + "eval_seconds": 18.664024114608765 + }, + { + "policy": "policies/franka/ppo/franka_M_s0/ep00250.npz", + "meta": { + "source": "logs/rl_games/franka_cabinet_direct/franka_M_s0/nn/last_franka_cabinet_direct_ep_250_rew_2956.8015.pth", + "epoch": 250, + "frame": 16384000, + "task": "franka", + "size": "M", + "run": "franka_M_s0", + "kind": "ppo" + }, + "num_envs": 4096, + "seed": 100000, + "success_rate": 0.06298828125, + "mean_return": 2893.275146484375, + "mean_steps_to_success": 346.39923095703125, + "terminated_rate": 0.00146484375, + "mean_max_progress": 0.2884792387485504, + "eval_seconds": 19.309536933898926 + }, + { + "policy": "policies/franka/ppo/franka_M_s0/ep00300.npz", + "meta": { + "source": "logs/rl_games/franka_cabinet_direct/franka_M_s0/nn/last_franka_cabinet_direct_ep_300_rew_2897.859.pth", + "epoch": 300, + "frame": 19660800, + "task": "franka", + "size": "M", + "run": "franka_M_s0", + "kind": "ppo" + }, + "num_envs": 4096, + "seed": 100000, + "success_rate": 0.061279296875, + "mean_return": 2917.61083984375, + "mean_steps_to_success": 214.99203491210938, + "terminated_rate": 0.00146484375, + "mean_max_progress": 0.24071845412254333, + "eval_seconds": 19.50385355949402 + }, + { + "policy": "policies/franka/ppo/franka_M_s0/ep00350.npz", + "meta": { + "source": "logs/rl_games/franka_cabinet_direct/franka_M_s0/nn/last_franka_cabinet_direct_ep_350_rew_2635.1208.pth", + "epoch": 350, + "frame": 22937600, + "task": "franka", + "size": "M", + "run": "franka_M_s0", + "kind": "ppo" + }, + "num_envs": 4096, + "seed": 100000, + "success_rate": 0.007568359375, + "mean_return": 3039.28857421875, + "mean_steps_to_success": 143.8064422607422, + "terminated_rate": 0.001953125, + "mean_max_progress": 0.27382194995880127, + "eval_seconds": 19.086139678955078 + }, + { + "policy": "policies/franka/ppo/franka_M_s0/ep00400.npz", + "meta": { + "source": "logs/rl_games/franka_cabinet_direct/franka_M_s0/nn/last_franka_cabinet_direct_ep_400_rew_3125.5283.pth", + "epoch": 400, + "frame": 26214400, + "task": "franka", + "size": "M", + "run": "franka_M_s0", + "kind": "ppo" + }, + "num_envs": 4096, + "seed": 100000, + "success_rate": 0.25390625, + "mean_return": 3161.740234375, + "mean_steps_to_success": 277.9163513183594, + "terminated_rate": 0.001220703125, + "mean_max_progress": 0.2992536723613739, + "eval_seconds": 19.111963033676147 + }, + { + "policy": "policies/franka/ppo/franka_M_s0/ep00450.npz", + "meta": { + "source": "logs/rl_games/franka_cabinet_direct/franka_M_s0/nn/last_franka_cabinet_direct_ep_450_rew_3143.1812.pth", + "epoch": 450, + "frame": 29491200, + "task": "franka", + "size": "M", + "run": "franka_M_s0", + "kind": "ppo" + }, + "num_envs": 4096, + "seed": 100000, + "success_rate": 0.203369140625, + "mean_return": 3178.29150390625, + "mean_steps_to_success": 204.78871154785156, + "terminated_rate": 0.0, + "mean_max_progress": 0.2830229699611664, + "eval_seconds": 20.142770051956177 + }, + { + "policy": "policies/franka/ppo/franka_M_s0/ep00500.npz", + "meta": { + "source": "logs/rl_games/franka_cabinet_direct/franka_M_s0/nn/last_franka_cabinet_direct_ep_500_rew_3122.486.pth", + "epoch": 500, + "frame": 32768000, + "task": "franka", + "size": "M", + "run": "franka_M_s0", + "kind": "ppo" + }, + "num_envs": 4096, + "seed": 100000, + "success_rate": 0.01220703125, + "mean_return": 3156.0048828125, + "mean_steps_to_success": 250.4399871826172, + "terminated_rate": 0.0, + "mean_max_progress": 0.2719091475009918, + "eval_seconds": 19.59170651435852 + }, + { + "policy": "policies/franka/ppo/franka_M_s0/ep00550.npz", + "meta": { + "source": "logs/rl_games/franka_cabinet_direct/franka_M_s0/nn/last_franka_cabinet_direct_ep_550_rew_3243.4766.pth", + "epoch": 550, + "frame": 36044800, + "task": "franka", + "size": "M", + "run": "franka_M_s0", + "kind": "ppo" + }, + "num_envs": 4096, + "seed": 100000, + "success_rate": 0.224609375, + "mean_return": 3105.09228515625, + "mean_steps_to_success": 185.84999084472656, + "terminated_rate": 0.0, + "mean_max_progress": 0.2974655032157898, + "eval_seconds": 18.915154695510864 + }, + { + "policy": "policies/franka/ppo/franka_M_s0/ep00600.npz", + "meta": { + "source": "logs/rl_games/franka_cabinet_direct/franka_M_s0/nn/last_franka_cabinet_direct_ep_600_rew_3196.356.pth", + "epoch": 600, + "frame": 39321600, + "task": "franka", + "size": "M", + "run": "franka_M_s0", + "kind": "ppo" + }, + "num_envs": 4096, + "seed": 100000, + "success_rate": 0.880859375, + "mean_return": 3301.5615234375, + "mean_steps_to_success": 238.3702850341797, + "terminated_rate": 0.0, + "mean_max_progress": 0.3031555712223053, + "eval_seconds": 18.17929697036743 + }, + { + "policy": "policies/franka/ppo/franka_M_s0/ep00650.npz", + "meta": { + "source": "logs/rl_games/franka_cabinet_direct/franka_M_s0/nn/last_franka_cabinet_direct_ep_650_rew_3055.7778.pth", + "epoch": 650, + "frame": 42598400, + "task": "franka", + "size": "M", + "run": "franka_M_s0", + "kind": "ppo" + }, + "num_envs": 4096, + "seed": 100000, + "success_rate": 0.001953125, + "mean_return": 3008.68359375, + "mean_steps_to_success": 191.875, + "terminated_rate": 0.000244140625, + "mean_max_progress": 0.25739407539367676, + "eval_seconds": 18.644951343536377 + }, + { + "policy": "policies/franka/ppo/franka_M_s0/ep00700.npz", + "meta": { + "source": "logs/rl_games/franka_cabinet_direct/franka_M_s0/nn/last_franka_cabinet_direct_ep_700_rew_3133.7593.pth", + "epoch": 700, + "frame": 45875200, + "task": "franka", + "size": "M", + "run": "franka_M_s0", + "kind": "ppo" + }, + "num_envs": 4096, + "seed": 100000, + "success_rate": 0.111083984375, + "mean_return": 3249.3544921875, + "mean_steps_to_success": 224.05276489257812, + "terminated_rate": 0.0, + "mean_max_progress": 0.2938815951347351, + "eval_seconds": 18.59794807434082 + }, + { + "policy": "policies/franka/ppo/franka_M_s0/ep00750.npz", + "meta": { + "source": "logs/rl_games/franka_cabinet_direct/franka_M_s0/nn/last_franka_cabinet_direct_ep_750_rew_3222.8247.pth", + "epoch": 750, + "frame": 49152000, + "task": "franka", + "size": "M", + "run": "franka_M_s0", + "kind": "ppo" + }, + "num_envs": 4096, + "seed": 100000, + "success_rate": 0.999267578125, + "mean_return": 3365.52294921875, + "mean_steps_to_success": 124.25507354736328, + "terminated_rate": 0.0, + "mean_max_progress": 0.31885284185409546, + "eval_seconds": 18.735332012176514 + }, + { + "policy": "policies/franka/ppo/franka_M_s0/ep00800.npz", + "meta": { + "source": "logs/rl_games/franka_cabinet_direct/franka_M_s0/nn/last_franka_cabinet_direct_ep_800_rew_2607.6875.pth", + "epoch": 800, + "frame": 52428800, + "task": "franka", + "size": "M", + "run": "franka_M_s0", + "kind": "ppo" + }, + "num_envs": 4096, + "seed": 100000, + "success_rate": 0.999755859375, + "mean_return": 3527.41552734375, + "mean_steps_to_success": 106.09475708007812, + "terminated_rate": 0.015380859375, + "mean_max_progress": 0.3791704773902893, + "eval_seconds": 18.014161348342896 + }, + { + "policy": "policies/franka/ppo/franka_M_s0/ep00850.npz", + "meta": { + "source": "logs/rl_games/franka_cabinet_direct/franka_M_s0/nn/last_franka_cabinet_direct_ep_850_rew_3603.0256.pth", + "epoch": 850, + "frame": 55705600, + "task": "franka", + "size": "M", + "run": "franka_M_s0", + "kind": "ppo" + }, + "num_envs": 4096, + "seed": 100000, + "success_rate": 0.99853515625, + "mean_return": 3452.39111328125, + "mean_steps_to_success": 112.1806869506836, + "terminated_rate": 0.0, + "mean_max_progress": 0.3711695075035095, + "eval_seconds": 18.435460567474365 + }, + { + "policy": "policies/franka/ppo/franka_M_s0/ep00900.npz", + "meta": { + "source": "logs/rl_games/franka_cabinet_direct/franka_M_s0/nn/last_franka_cabinet_direct_ep_900_rew_3510.5527.pth", + "epoch": 900, + "frame": 58982400, + "task": "franka", + "size": "M", + "run": "franka_M_s0", + "kind": "ppo" + }, + "num_envs": 4096, + "seed": 100000, + "success_rate": 1.0, + "mean_return": 3609.58349609375, + "mean_steps_to_success": 107.81982421875, + "terminated_rate": 0.0, + "mean_max_progress": 0.3672513961791992, + "eval_seconds": 17.999112606048584 + }, + { + "policy": "policies/franka/ppo/franka_M_s0/ep00950.npz", + "meta": { + "source": "logs/rl_games/franka_cabinet_direct/franka_M_s0/nn/last_franka_cabinet_direct_ep_950_rew_3580.9531.pth", + "epoch": 950, + "frame": 62259200, + "task": "franka", + "size": "M", + "run": "franka_M_s0", + "kind": "ppo" + }, + "num_envs": 4096, + "seed": 100000, + "success_rate": 1.0, + "mean_return": 3530.56494140625, + "mean_steps_to_success": 110.7998046875, + "terminated_rate": 0.0, + "mean_max_progress": 0.3685230612754822, + "eval_seconds": 18.045910596847534 + }, + { + "policy": "policies/franka/ppo/franka_M_s0/ep01000.npz", + "meta": { + "source": "logs/rl_games/franka_cabinet_direct/franka_M_s0/nn/last_franka_cabinet_direct_ep_1000_rew_3536.2046.pth", + "epoch": 1000, + "frame": 65536000, + "task": "franka", + "size": "M", + "run": "franka_M_s0", + "kind": "ppo" + }, + "num_envs": 4096, + "seed": 100000, + "success_rate": 1.0, + "mean_return": 3568.017578125, + "mean_steps_to_success": 114.417724609375, + "terminated_rate": 0.0, + "mean_max_progress": 0.3681824803352356, + "eval_seconds": 18.69905924797058 + }, + { + "policy": "policies/franka/ppo/franka_M_s0/ep01050.npz", + "meta": { + "source": "logs/rl_games/franka_cabinet_direct/franka_M_s0/nn/last_franka_cabinet_direct_ep_1050_rew_3454.682.pth", + "epoch": 1050, + "frame": 68812800, + "task": "franka", + "size": "M", + "run": "franka_M_s0", + "kind": "ppo" + }, + "num_envs": 4096, + "seed": 100000, + "success_rate": 1.0, + "mean_return": 3390.357177734375, + "mean_steps_to_success": 117.474609375, + "terminated_rate": 0.0, + "mean_max_progress": 0.3446342945098877, + "eval_seconds": 19.95788073539734 + }, + { + "policy": "policies/franka/ppo/franka_M_s0/ep01100.npz", + "meta": { + "source": "logs/rl_games/franka_cabinet_direct/franka_M_s0/nn/last_franka_cabinet_direct_ep_1100_rew_3405.9514.pth", + "epoch": 1100, + "frame": 72089600, + "task": "franka", + "size": "M", + "run": "franka_M_s0", + "kind": "ppo" + }, + "num_envs": 4096, + "seed": 100000, + "success_rate": 0.9951171875, + "mean_return": 3432.96875, + "mean_steps_to_success": 124.39425659179688, + "terminated_rate": 0.0, + "mean_max_progress": 0.3385787904262543, + "eval_seconds": 19.525527954101562 + }, + { + "policy": "policies/franka/ppo/franka_M_s0/ep01150.npz", + "meta": { + "source": "logs/rl_games/franka_cabinet_direct/franka_M_s0/nn/last_franka_cabinet_direct_ep_1150_rew_3387.14.pth", + "epoch": 1150, + "frame": 75366400, + "task": "franka", + "size": "M", + "run": "franka_M_s0", + "kind": "ppo" + }, + "num_envs": 4096, + "seed": 100000, + "success_rate": 0.9931640625, + "mean_return": 3536.29345703125, + "mean_steps_to_success": 122.14774322509766, + "terminated_rate": 0.0, + "mean_max_progress": 0.36475616693496704, + "eval_seconds": 19.64392900466919 + }, + { + "policy": "policies/franka/ppo/franka_M_s0/ep01200.npz", + "meta": { + "source": "logs/rl_games/franka_cabinet_direct/franka_M_s0/nn/last_franka_cabinet_direct_ep_1200_rew_3491.3674.pth", + "epoch": 1200, + "frame": 78643200, + "task": "franka", + "size": "M", + "run": "franka_M_s0", + "kind": "ppo" + }, + "num_envs": 4096, + "seed": 100000, + "success_rate": 0.999755859375, + "mean_return": 3572.97119140625, + "mean_steps_to_success": 121.32186126708984, + "terminated_rate": 0.0, + "mean_max_progress": 0.35731884837150574, + "eval_seconds": 18.4236798286438 + }, + { + "policy": "policies/franka/ppo/franka_M_s0/ep01250.npz", + "meta": { + "source": "logs/rl_games/franka_cabinet_direct/franka_M_s0/nn/last_franka_cabinet_direct_ep_1250_rew_3378.2983.pth", + "epoch": 1250, + "frame": 81920000, + "task": "franka", + "size": "M", + "run": "franka_M_s0", + "kind": "ppo" + }, + "num_envs": 4096, + "seed": 100000, + "success_rate": 0.999755859375, + "mean_return": 3341.490234375, + "mean_steps_to_success": 122.80416107177734, + "terminated_rate": 0.0, + "mean_max_progress": 0.33779436349868774, + "eval_seconds": 18.20985984802246 + }, + { + "policy": "policies/franka/ppo/franka_M_s0/ep01300.npz", + "meta": { + "source": "logs/rl_games/franka_cabinet_direct/franka_M_s0/nn/last_franka_cabinet_direct_ep_1300_rew_3360.3933.pth", + "epoch": 1300, + "frame": 85196800, + "task": "franka", + "size": "M", + "run": "franka_M_s0", + "kind": "ppo" + }, + "num_envs": 4096, + "seed": 100000, + "success_rate": 0.999755859375, + "mean_return": 3427.54052734375, + "mean_steps_to_success": 114.0549545288086, + "terminated_rate": 0.0, + "mean_max_progress": 0.34420645236968994, + "eval_seconds": 18.46643328666687 + }, + { + "policy": "policies/franka/ppo/franka_M_s0/ep01350.npz", + "meta": { + "source": "logs/rl_games/franka_cabinet_direct/franka_M_s0/nn/last_franka_cabinet_direct_ep_1350_rew_3363.625.pth", + "epoch": 1350, + "frame": 88473600, + "task": "franka", + "size": "M", + "run": "franka_M_s0", + "kind": "ppo" + }, + "num_envs": 4096, + "seed": 100000, + "success_rate": 0.995849609375, + "mean_return": 3454.40869140625, + "mean_steps_to_success": 112.16156005859375, + "terminated_rate": 0.001708984375, + "mean_max_progress": 0.3658413290977478, + "eval_seconds": 19.93006682395935 + }, + { + "policy": "policies/franka/ppo/franka_M_s0/ep01400.npz", + "meta": { + "source": "logs/rl_games/franka_cabinet_direct/franka_M_s0/nn/last_franka_cabinet_direct_ep_1400_rew_3303.4365.pth", + "epoch": 1400, + "frame": 91750400, + "task": "franka", + "size": "M", + "run": "franka_M_s0", + "kind": "ppo" + }, + "num_envs": 4096, + "seed": 100000, + "success_rate": 0.999755859375, + "mean_return": 3522.39208984375, + "mean_steps_to_success": 113.80342864990234, + "terminated_rate": 0.0, + "mean_max_progress": 0.37344759702682495, + "eval_seconds": 20.06979775428772 + }, + { + "policy": "policies/franka/ppo/franka_M_s0/ep01450.npz", + "meta": { + "source": "logs/rl_games/franka_cabinet_direct/franka_M_s0/nn/last_franka_cabinet_direct_ep_1450_rew_3428.7139.pth", + "epoch": 1450, + "frame": 95027200, + "task": "franka", + "size": "M", + "run": "franka_M_s0", + "kind": "ppo" + }, + "num_envs": 4096, + "seed": 100000, + "success_rate": 1.0, + "mean_return": 3638.8525390625, + "mean_steps_to_success": 116.986328125, + "terminated_rate": 0.0, + "mean_max_progress": 0.3752778172492981, + "eval_seconds": 20.05346965789795 + }, + { + "policy": "policies/franka/ppo/franka_M_s0/ep01500.npz", + "meta": { + "source": "logs/rl_games/franka_cabinet_direct/franka_M_s0/nn/last_franka_cabinet_direct_ep_1500_rew__3492.267_.pth", + "epoch": 1500, + "frame": 98304000, + "task": "franka", + "size": "M", + "run": "franka_M_s0", + "kind": "ppo" + }, + "num_envs": 4096, + "seed": 100000, + "success_rate": 0.999755859375, + "mean_return": 3546.498046875, + "mean_steps_to_success": 122.51282501220703, + "terminated_rate": 0.0, + "mean_max_progress": 0.3604103922843933, + "eval_seconds": 20.07447862625122 + }, + { + "policy": "policies/franka/ppo/franka_XS_s0/ep00050.npz", + "meta": { + "source": "logs/rl_games/franka_cabinet_direct/franka_XS_s0/nn/last_franka_cabinet_direct_ep_50_rew_1030.5833.pth", + "epoch": 50, + "frame": 3276800, + "task": "franka", + "size": "XS", + "run": "franka_XS_s0", + "kind": "ppo" + }, + "num_envs": 4096, + "seed": 100000, + "success_rate": 0.946044921875, + "mean_return": 1603.60009765625, + "mean_steps_to_success": 192.0626983642578, + "terminated_rate": 0.54345703125, + "mean_max_progress": 0.36774152517318726, + "eval_seconds": 19.758694410324097 + }, + { + "policy": "policies/franka/ppo/franka_XS_s0/ep00100.npz", + "meta": { + "source": "logs/rl_games/franka_cabinet_direct/franka_XS_s0/nn/last_franka_cabinet_direct_ep_100_rew_1787.6931.pth", + "epoch": 100, + "frame": 6553600, + "task": "franka", + "size": "XS", + "run": "franka_XS_s0", + "kind": "ppo" + }, + "num_envs": 4096, + "seed": 100000, + "success_rate": 0.943115234375, + "mean_return": 2157.144287109375, + "mean_steps_to_success": 159.74295043945312, + "terminated_rate": 0.1123046875, + "mean_max_progress": 0.34825268387794495, + "eval_seconds": 18.320976734161377 + }, + { + "policy": "policies/franka/ppo/franka_XS_s0/ep00150.npz", + "meta": { + "source": "logs/rl_games/franka_cabinet_direct/franka_XS_s0/nn/last_franka_cabinet_direct_ep_150_rew_2120.8706.pth", + "epoch": 150, + "frame": 9830400, + "task": "franka", + "size": "XS", + "run": "franka_XS_s0", + "kind": "ppo" + }, + "num_envs": 4096, + "seed": 100000, + "success_rate": 0.998291015625, + "mean_return": 2493.627197265625, + "mean_steps_to_success": 104.97799682617188, + "terminated_rate": 0.065185546875, + "mean_max_progress": 0.3726787567138672, + "eval_seconds": 19.454976797103882 + }, + { + "policy": "policies/franka/ppo/franka_XS_s0/ep00200.npz", + "meta": { + "source": "logs/rl_games/franka_cabinet_direct/franka_XS_s0/nn/last_franka_cabinet_direct_ep_200_rew_2308.2454.pth", + "epoch": 200, + "frame": 13107200, + "task": "franka", + "size": "XS", + "run": "franka_XS_s0", + "kind": "ppo" + }, + "num_envs": 4096, + "seed": 100000, + "success_rate": 0.9951171875, + "mean_return": 2589.57373046875, + "mean_steps_to_success": 102.51840209960938, + "terminated_rate": 0.051025390625, + "mean_max_progress": 0.3655995726585388, + "eval_seconds": 18.84553599357605 + }, + { + "policy": "policies/franka/ppo/franka_XS_s0/ep00250.npz", + "meta": { + "source": "logs/rl_games/franka_cabinet_direct/franka_XS_s0/nn/last_franka_cabinet_direct_ep_250_rew_2555.7402.pth", + "epoch": 250, + "frame": 16384000, + "task": "franka", + "size": "XS", + "run": "franka_XS_s0", + "kind": "ppo" + }, + "num_envs": 4096, + "seed": 100000, + "success_rate": 1.0, + "mean_return": 2473.36962890625, + "mean_steps_to_success": 85.525634765625, + "terminated_rate": 0.21923828125, + "mean_max_progress": 0.37679237127304077, + "eval_seconds": 19.591014623641968 + }, + { + "policy": "policies/franka/ppo/franka_XS_s0/ep00300.npz", + "meta": { + "source": "logs/rl_games/franka_cabinet_direct/franka_XS_s0/nn/last_franka_cabinet_direct_ep_300_rew_2495.4395.pth", + "epoch": 300, + "frame": 19660800, + "task": "franka", + "size": "XS", + "run": "franka_XS_s0", + "kind": "ppo" + }, + "num_envs": 4096, + "seed": 100000, + "success_rate": 1.0, + "mean_return": 2753.76611328125, + "mean_steps_to_success": 84.203857421875, + "terminated_rate": 0.0244140625, + "mean_max_progress": 0.37586653232574463, + "eval_seconds": 19.1250102519989 + }, + { + "policy": "policies/franka/ppo/franka_XS_s0/ep00350.npz", + "meta": { + "source": "logs/rl_games/franka_cabinet_direct/franka_XS_s0/nn/last_franka_cabinet_direct_ep_350_rew_2527.5977.pth", + "epoch": 350, + "frame": 22937600, + "task": "franka", + "size": "XS", + "run": "franka_XS_s0", + "kind": "ppo" + }, + "num_envs": 4096, + "seed": 100000, + "success_rate": 0.99951171875, + "mean_return": 2891.5146484375, + "mean_steps_to_success": 72.78773498535156, + "terminated_rate": 0.076416015625, + "mean_max_progress": 0.3763211965560913, + "eval_seconds": 19.628764867782593 + }, + { + "policy": "policies/franka/ppo/franka_XS_s0/ep00400.npz", + "meta": { + "source": "logs/rl_games/franka_cabinet_direct/franka_XS_s0/nn/last_franka_cabinet_direct_ep_400_rew_2870.178.pth", + "epoch": 400, + "frame": 26214400, + "task": "franka", + "size": "XS", + "run": "franka_XS_s0", + "kind": "ppo" + }, + "num_envs": 4096, + "seed": 100000, + "success_rate": 1.0, + "mean_return": 3031.28857421875, + "mean_steps_to_success": 71.07470703125, + "terminated_rate": 0.05517578125, + "mean_max_progress": 0.3778989911079407, + "eval_seconds": 19.53827977180481 + }, + { + "policy": "policies/franka/ppo/franka_XS_s0/ep00450.npz", + "meta": { + "source": "logs/rl_games/franka_cabinet_direct/franka_XS_s0/nn/last_franka_cabinet_direct_ep_450_rew_2837.5962.pth", + "epoch": 450, + "frame": 29491200, + "task": "franka", + "size": "XS", + "run": "franka_XS_s0", + "kind": "ppo" + }, + "num_envs": 4096, + "seed": 100000, + "success_rate": 1.0, + "mean_return": 3139.54052734375, + "mean_steps_to_success": 68.12548828125, + "terminated_rate": 0.015625, + "mean_max_progress": 0.37539392709732056, + "eval_seconds": 19.17515230178833 + }, + { + "policy": "policies/franka/ppo/franka_XS_s0/ep00500.npz", + "meta": { + "source": "logs/rl_games/franka_cabinet_direct/franka_XS_s0/nn/last_franka_cabinet_direct_ep_500_rew_3041.5105.pth", + "epoch": 500, + "frame": 32768000, + "task": "franka", + "size": "XS", + "run": "franka_XS_s0", + "kind": "ppo" + }, + "num_envs": 4096, + "seed": 100000, + "success_rate": 1.0, + "mean_return": 3103.80908203125, + "mean_steps_to_success": 69.239501953125, + "terminated_rate": 0.009521484375, + "mean_max_progress": 0.371712327003479, + "eval_seconds": 19.540491580963135 + }, + { + "policy": "policies/franka/ppo/franka_XS_s0/ep00550.npz", + "meta": { + "source": "logs/rl_games/franka_cabinet_direct/franka_XS_s0/nn/last_franka_cabinet_direct_ep_550_rew_2973.1555.pth", + "epoch": 550, + "frame": 36044800, + "task": "franka", + "size": "XS", + "run": "franka_XS_s0", + "kind": "ppo" + }, + "num_envs": 4096, + "seed": 100000, + "success_rate": 0.999755859375, + "mean_return": 3137.43212890625, + "mean_steps_to_success": 68.92991638183594, + "terminated_rate": 0.015380859375, + "mean_max_progress": 0.377410888671875, + "eval_seconds": 19.650789260864258 + }, + { + "policy": "policies/franka/ppo/franka_XS_s0/ep00600.npz", + "meta": { + "source": "logs/rl_games/franka_cabinet_direct/franka_XS_s0/nn/last_franka_cabinet_direct_ep_600_rew_2896.7861.pth", + "epoch": 600, + "frame": 39321600, + "task": "franka", + "size": "XS", + "run": "franka_XS_s0", + "kind": "ppo" + }, + "num_envs": 4096, + "seed": 100000, + "success_rate": 0.99658203125, + "mean_return": 3176.828857421875, + "mean_steps_to_success": 66.81773376464844, + "terminated_rate": 0.0126953125, + "mean_max_progress": 0.3797222971916199, + "eval_seconds": 19.499059438705444 + }, + { + "policy": "policies/franka/ppo/franka_XS_s0/ep00650.npz", + "meta": { + "source": "logs/rl_games/franka_cabinet_direct/franka_XS_s0/nn/last_franka_cabinet_direct_ep_650_rew_2949.5974.pth", + "epoch": 650, + "frame": 42598400, + "task": "franka", + "size": "XS", + "run": "franka_XS_s0", + "kind": "ppo" + }, + "num_envs": 4096, + "seed": 100000, + "success_rate": 1.0, + "mean_return": 3160.20361328125, + "mean_steps_to_success": 66.194091796875, + "terminated_rate": 0.019287109375, + "mean_max_progress": 0.37869009375572205, + "eval_seconds": 19.15119695663452 + }, + { + "policy": "policies/franka/ppo/franka_XS_s0/ep00700.npz", + "meta": { + "source": "logs/rl_games/franka_cabinet_direct/franka_XS_s0/nn/last_franka_cabinet_direct_ep_700_rew_3120.159.pth", + "epoch": 700, + "frame": 45875200, + "task": "franka", + "size": "XS", + "run": "franka_XS_s0", + "kind": "ppo" + }, + "num_envs": 4096, + "seed": 100000, + "success_rate": 0.99951171875, + "mean_return": 3198.71337890625, + "mean_steps_to_success": 66.89423370361328, + "terminated_rate": 0.01123046875, + "mean_max_progress": 0.37854766845703125, + "eval_seconds": 19.358795881271362 + }, + { + "policy": "policies/franka/ppo/franka_XS_s0/ep00750.npz", + "meta": { + "source": "logs/rl_games/franka_cabinet_direct/franka_XS_s0/nn/last_franka_cabinet_direct_ep_750_rew_3100.3374.pth", + "epoch": 750, + "frame": 49152000, + "task": "franka", + "size": "XS", + "run": "franka_XS_s0", + "kind": "ppo" + }, + "num_envs": 4096, + "seed": 100000, + "success_rate": 0.999755859375, + "mean_return": 3198.54150390625, + "mean_steps_to_success": 65.50574493408203, + "terminated_rate": 0.003173828125, + "mean_max_progress": 0.37640172243118286, + "eval_seconds": 19.04165530204773 + }, + { + "policy": "policies/franka/ppo/franka_XS_s0/ep00800.npz", + "meta": { + "source": "logs/rl_games/franka_cabinet_direct/franka_XS_s0/nn/last_franka_cabinet_direct_ep_800_rew_3075.0693.pth", + "epoch": 800, + "frame": 52428800, + "task": "franka", + "size": "XS", + "run": "franka_XS_s0", + "kind": "ppo" + }, + "num_envs": 4096, + "seed": 100000, + "success_rate": 0.99755859375, + "mean_return": 3147.941650390625, + "mean_steps_to_success": 64.97601318359375, + "terminated_rate": 0.012451171875, + "mean_max_progress": 0.3786541819572449, + "eval_seconds": 19.68586277961731 + }, + { + "policy": "policies/franka/ppo/franka_XS_s0/ep00850.npz", + "meta": { + "source": "logs/rl_games/franka_cabinet_direct/franka_XS_s0/nn/last_franka_cabinet_direct_ep_850_rew_2971.3489.pth", + "epoch": 850, + "frame": 55705600, + "task": "franka", + "size": "XS", + "run": "franka_XS_s0", + "kind": "ppo" + }, + "num_envs": 4096, + "seed": 100000, + "success_rate": 1.0, + "mean_return": 3135.861328125, + "mean_steps_to_success": 68.823486328125, + "terminated_rate": 0.019287109375, + "mean_max_progress": 0.3849572539329529, + "eval_seconds": 18.321229219436646 + }, + { + "policy": "policies/franka/ppo/franka_XS_s0/ep00900.npz", + "meta": { + "source": "logs/rl_games/franka_cabinet_direct/franka_XS_s0/nn/last_franka_cabinet_direct_ep_900_rew_2895.5295.pth", + "epoch": 900, + "frame": 58982400, + "task": "franka", + "size": "XS", + "run": "franka_XS_s0", + "kind": "ppo" + }, + "num_envs": 4096, + "seed": 100000, + "success_rate": 0.99951171875, + "mean_return": 3158.867431640625, + "mean_steps_to_success": 62.73326873779297, + "terminated_rate": 0.0048828125, + "mean_max_progress": 0.3832535743713379, + "eval_seconds": 18.243725776672363 + }, + { + "policy": "policies/franka/ppo/franka_XS_s0/ep00950.npz", + "meta": { + "source": "logs/rl_games/franka_cabinet_direct/franka_XS_s0/nn/last_franka_cabinet_direct_ep_950_rew_2951.182.pth", + "epoch": 950, + "frame": 62259200, + "task": "franka", + "size": "XS", + "run": "franka_XS_s0", + "kind": "ppo" + }, + "num_envs": 4096, + "seed": 100000, + "success_rate": 0.998291015625, + "mean_return": 3029.046875, + "mean_steps_to_success": 69.7006607055664, + "terminated_rate": 0.021484375, + "mean_max_progress": 0.3831526041030884, + "eval_seconds": 19.215763330459595 + }, + { + "policy": "policies/franka/ppo/franka_XS_s0/ep01000.npz", + "meta": { + "source": "logs/rl_games/franka_cabinet_direct/franka_XS_s0/nn/last_franka_cabinet_direct_ep_1000_rew_2968.6636.pth", + "epoch": 1000, + "frame": 65536000, + "task": "franka", + "size": "XS", + "run": "franka_XS_s0", + "kind": "ppo" + }, + "num_envs": 4096, + "seed": 100000, + "success_rate": 0.997314453125, + "mean_return": 3038.9658203125, + "mean_steps_to_success": 66.56646728515625, + "terminated_rate": 0.006103515625, + "mean_max_progress": 0.3795437812805176, + "eval_seconds": 18.649858951568604 + }, + { + "policy": "policies/franka/ppo/franka_XS_s0/ep01050.npz", + "meta": { + "source": "logs/rl_games/franka_cabinet_direct/franka_XS_s0/nn/last_franka_cabinet_direct_ep_1050_rew_3003.6707.pth", + "epoch": 1050, + "frame": 68812800, + "task": "franka", + "size": "XS", + "run": "franka_XS_s0", + "kind": "ppo" + }, + "num_envs": 4096, + "seed": 100000, + "success_rate": 0.998779296875, + "mean_return": 3103.776611328125, + "mean_steps_to_success": 63.464683532714844, + "terminated_rate": 0.005126953125, + "mean_max_progress": 0.3815617561340332, + "eval_seconds": 18.211350917816162 + }, + { + "policy": "policies/franka/ppo/franka_XS_s0/ep01100.npz", + "meta": { + "source": "logs/rl_games/franka_cabinet_direct/franka_XS_s0/nn/last_franka_cabinet_direct_ep_1100_rew_3081.8035.pth", + "epoch": 1100, + "frame": 72089600, + "task": "franka", + "size": "XS", + "run": "franka_XS_s0", + "kind": "ppo" + }, + "num_envs": 4096, + "seed": 100000, + "success_rate": 0.998779296875, + "mean_return": 3105.99755859375, + "mean_steps_to_success": 64.93473815917969, + "terminated_rate": 0.004150390625, + "mean_max_progress": 0.3761790990829468, + "eval_seconds": 18.53570556640625 + }, + { + "policy": "policies/franka/ppo/franka_XS_s0/ep01150.npz", + "meta": { + "source": "logs/rl_games/franka_cabinet_direct/franka_XS_s0/nn/last_franka_cabinet_direct_ep_1150_rew_3083.72.pth", + "epoch": 1150, + "frame": 75366400, + "task": "franka", + "size": "XS", + "run": "franka_XS_s0", + "kind": "ppo" + }, + "num_envs": 4096, + "seed": 100000, + "success_rate": 0.999755859375, + "mean_return": 3112.84912109375, + "mean_steps_to_success": 60.96044158935547, + "terminated_rate": 0.0029296875, + "mean_max_progress": 0.3729322552680969, + "eval_seconds": 18.238340377807617 + }, + { + "policy": "policies/franka/ppo/franka_XS_s0/ep01200.npz", + "meta": { + "source": "logs/rl_games/franka_cabinet_direct/franka_XS_s0/nn/last_franka_cabinet_direct_ep_1200_rew_2938.0317.pth", + "epoch": 1200, + "frame": 78643200, + "task": "franka", + "size": "XS", + "run": "franka_XS_s0", + "kind": "ppo" + }, + "num_envs": 4096, + "seed": 100000, + "success_rate": 0.999267578125, + "mean_return": 2982.59765625, + "mean_steps_to_success": 65.06475067138672, + "terminated_rate": 0.04248046875, + "mean_max_progress": 0.37608951330184937, + "eval_seconds": 18.797752618789673 + }, + { + "policy": "policies/franka/ppo/franka_XS_s0/ep01250.npz", + "meta": { + "source": "logs/rl_games/franka_cabinet_direct/franka_XS_s0/nn/last_franka_cabinet_direct_ep_1250_rew_3034.9758.pth", + "epoch": 1250, + "frame": 81920000, + "task": "franka", + "size": "XS", + "run": "franka_XS_s0", + "kind": "ppo" + }, + "num_envs": 4096, + "seed": 100000, + "success_rate": 0.998046875, + "mean_return": 3048.687255859375, + "mean_steps_to_success": 65.49192810058594, + "terminated_rate": 0.010009765625, + "mean_max_progress": 0.37494587898254395, + "eval_seconds": 19.250694513320923 + }, + { + "policy": "policies/franka/ppo/franka_XS_s0/ep01300.npz", + "meta": { + "source": "logs/rl_games/franka_cabinet_direct/franka_XS_s0/nn/last_franka_cabinet_direct_ep_1300_rew_3012.813.pth", + "epoch": 1300, + "frame": 85196800, + "task": "franka", + "size": "XS", + "run": "franka_XS_s0", + "kind": "ppo" + }, + "num_envs": 4096, + "seed": 100000, + "success_rate": 0.99951171875, + "mean_return": 3091.642578125, + "mean_steps_to_success": 63.67782211303711, + "terminated_rate": 0.011962890625, + "mean_max_progress": 0.38049888610839844, + "eval_seconds": 19.301215648651123 + }, + { + "policy": "policies/franka/ppo/franka_XS_s0/ep01350.npz", + "meta": { + "source": "logs/rl_games/franka_cabinet_direct/franka_XS_s0/nn/last_franka_cabinet_direct_ep_1350_rew_2923.0774.pth", + "epoch": 1350, + "frame": 88473600, + "task": "franka", + "size": "XS", + "run": "franka_XS_s0", + "kind": "ppo" + }, + "num_envs": 4096, + "seed": 100000, + "success_rate": 0.9931640625, + "mean_return": 2871.3876953125, + "mean_steps_to_success": 76.70304870605469, + "terminated_rate": 0.085205078125, + "mean_max_progress": 0.37792283296585083, + "eval_seconds": 20.39572024345398 + }, + { + "policy": "policies/franka/ppo/franka_XS_s0/ep01400.npz", + "meta": { + "source": "logs/rl_games/franka_cabinet_direct/franka_XS_s0/nn/last_franka_cabinet_direct_ep_1400_rew_3035.8142.pth", + "epoch": 1400, + "frame": 91750400, + "task": "franka", + "size": "XS", + "run": "franka_XS_s0", + "kind": "ppo" + }, + "num_envs": 4096, + "seed": 100000, + "success_rate": 0.9970703125, + "mean_return": 3054.15185546875, + "mean_steps_to_success": 68.9664535522461, + "terminated_rate": 0.01953125, + "mean_max_progress": 0.3745594322681427, + "eval_seconds": 19.836262464523315 + }, + { + "policy": "policies/franka/ppo/franka_XS_s0/ep01450.npz", + "meta": { + "source": "logs/rl_games/franka_cabinet_direct/franka_XS_s0/nn/last_franka_cabinet_direct_ep_1450_rew_3032.511.pth", + "epoch": 1450, + "frame": 95027200, + "task": "franka", + "size": "XS", + "run": "franka_XS_s0", + "kind": "ppo" + }, + "num_envs": 4096, + "seed": 100000, + "success_rate": 0.997802734375, + "mean_return": 3071.83837890625, + "mean_steps_to_success": 64.14730072021484, + "terminated_rate": 0.024169921875, + "mean_max_progress": 0.3775964379310608, + "eval_seconds": 19.632765769958496 + }, + { + "policy": "policies/franka/ppo/franka_XS_s0/ep01500.npz", + "meta": { + "source": "logs/rl_games/franka_cabinet_direct/franka_XS_s0/nn/last_franka_cabinet_direct_ep_1500_rew__3062.0417_.pth", + "epoch": 1500, + "frame": 98304000, + "task": "franka", + "size": "XS", + "run": "franka_XS_s0", + "kind": "ppo" + }, + "num_envs": 4096, + "seed": 100000, + "success_rate": 0.998779296875, + "mean_return": 3060.883056640625, + "mean_steps_to_success": 65.38743591308594, + "terminated_rate": 0.017822265625, + "mean_max_progress": 0.3767516613006592, + "eval_seconds": 19.199859857559204 + } +] diff --git a/metadata/data/robot_policy/baselines/franka_repair_M_heldout.json b/metadata/data/robot_policy/baselines/franka_repair_M_heldout.json new file mode 100644 index 0000000000000000000000000000000000000000..e79b9c6357c583498be134048200bbb420a28fec --- /dev/null +++ b/metadata/data/robot_policy/baselines/franka_repair_M_heldout.json @@ -0,0 +1,376 @@ +[ + { + "policy": "policies/franka/repair/M_s0/damaged/dead_units_0_0p5.npz", + "meta": { + "source": "logs/rl_games/franka_cabinet_direct/franka_M_s0/nn/last_franka_cabinet_direct_ep_1500_rew__3492.267_.pth", + "epoch": 1500, + "frame": 98304000, + "task": "franka", + "size": "M", + "run": "franka_M_s0", + "kind": "broken", + "expert": "policies/franka/ppo/franka_M_s0/ep01500.npz", + "damage": { + "kind": "dead_units", + "arg": [ + 0, + 0.5 + ], + "edited": [ + "L0_W rows / L0_b [0, 1, 2, 3, 4, 5, 10, 12, 13, 14, 18, 21, 23, 24, 26, 28, 35, 36, 40, 41, 45, 49, 50, 53, 55, 56, 58, 60, 62, 66, 68, 69, 70, 71, 72, 73, 74, 79, 80, 82, 83, 85, 86, 88, 90, 91, 98, 101, 104, 105, 108, 109, 112, 113, 115, 116, 120, 122, 127, 130, 132, 134, 136, 137, 138, 139, 140, 141, 145, 146, 147, 148, 153, 155, 158, 159, 165, 166, 167, 168, 170, 172, 174, 177, 180, 181, 182, 184, 185, 186, 187, 189, 190, 192, 194, 195, 198, 200, 203, 204, 205, 206, 207, 209, 211, 213, 217, 218, 223, 224, 226, 228, 229, 234, 235, 236, 237, 239, 240, 241, 242, 243, 246, 248, 249, 250, 251, 254]" + ] + } + }, + "num_envs": 4096, + "seed": 100000, + "success_rate": 0.6220703125, + "mean_return": 1851.43359375, + "mean_steps_to_success": 153.95408630371094, + "terminated_rate": 0.01904296875, + "mean_max_progress": 0.24120986461639404, + "eval_seconds": 21.57297420501709 + }, + { + "policy": "policies/franka/repair/M_s0/damaged/dead_units_1_0p5.npz", + "meta": { + "source": "logs/rl_games/franka_cabinet_direct/franka_M_s0/nn/last_franka_cabinet_direct_ep_1500_rew__3492.267_.pth", + "epoch": 1500, + "frame": 98304000, + "task": "franka", + "size": "M", + "run": "franka_M_s0", + "kind": "broken", + "expert": "policies/franka/ppo/franka_M_s0/ep01500.npz", + "damage": { + "kind": "dead_units", + "arg": [ + 1, + 0.5 + ], + "edited": [ + "L1_W rows / L1_b [1, 2, 4, 7, 8, 9, 11, 12, 14, 15, 19, 21, 22, 30, 32, 33, 35, 36, 40, 41, 43, 44, 50, 52, 53, 55, 56, 60, 61, 66, 67, 69, 73, 76, 77, 80, 81, 83, 85, 86, 87, 90, 91, 92, 95, 96, 97, 98, 99, 102, 103, 105, 108, 109, 112, 113, 114, 116, 117, 118, 120, 123, 125, 127]" + ] + } + }, + "num_envs": 4096, + "seed": 100000, + "success_rate": 0.093994140625, + "mean_return": 510.7125549316406, + "mean_steps_to_success": 166.77662658691406, + "terminated_rate": 0.064208984375, + "mean_max_progress": 0.06240244209766388, + "eval_seconds": 22.137210369110107 + }, + { + "policy": "policies/franka/repair/M_s0/damaged/flip_obs_18.npz", + "meta": { + "source": "logs/rl_games/franka_cabinet_direct/franka_M_s0/nn/last_franka_cabinet_direct_ep_1500_rew__3492.267_.pth", + "epoch": 1500, + "frame": 98304000, + "task": "franka", + "size": "M", + "run": "franka_M_s0", + "kind": "broken", + "expert": "policies/franka/ppo/franka_M_s0/ep01500.npz", + "damage": { + "kind": "flip_obs", + "arg": 18, + "edited": [ + "L0_W[:, 18]" + ] + } + }, + "num_envs": 4096, + "seed": 100000, + "success_rate": 0.0, + "mean_return": -432.8334655761719, + "mean_steps_to_success": null, + "terminated_rate": 0.0, + "mean_max_progress": 9.682253887532966e-10, + "eval_seconds": 18.731356143951416 + }, + { + "policy": "policies/franka/repair/M_s0/damaged/flip_obs_20.npz", + "meta": { + "source": "logs/rl_games/franka_cabinet_direct/franka_M_s0/nn/last_franka_cabinet_direct_ep_1500_rew__3492.267_.pth", + "epoch": 1500, + "frame": 98304000, + "task": "franka", + "size": "M", + "run": "franka_M_s0", + "kind": "broken", + "expert": "policies/franka/ppo/franka_M_s0/ep01500.npz", + "damage": { + "kind": "flip_obs", + "arg": 20, + "edited": [ + "L0_W[:, 20]" + ] + } + }, + "num_envs": 4096, + "seed": 100000, + "success_rate": 0.0, + "mean_return": -227.9674530029297, + "mean_steps_to_success": null, + "terminated_rate": 0.0, + "mean_max_progress": 0.000288566283416003, + "eval_seconds": 18.008240461349487 + }, + { + "policy": "policies/franka/repair/M_s0/damaged/noise_0p5.npz", + "meta": { + "source": "logs/rl_games/franka_cabinet_direct/franka_M_s0/nn/last_franka_cabinet_direct_ep_1500_rew__3492.267_.pth", + "epoch": 1500, + "frame": 98304000, + "task": "franka", + "size": "M", + "run": "franka_M_s0", + "kind": "broken", + "expert": "policies/franka/ppo/franka_M_s0/ep01500.npz", + "damage": { + "kind": "noise", + "arg": 0.5, + "edited": [ + "all W (Gaussian)" + ] + } + }, + "num_envs": 4096, + "seed": 100000, + "success_rate": 0.00146484375, + "mean_return": -41.02936935424805, + "mean_steps_to_success": 277.66668701171875, + "terminated_rate": 0.0009765625, + "mean_max_progress": 0.0025627710856497288, + "eval_seconds": 22.01688861846924 + }, + { + "policy": "policies/franka/repair/M_s0/damaged/noise_1p0.npz", + "meta": { + "source": "logs/rl_games/franka_cabinet_direct/franka_M_s0/nn/last_franka_cabinet_direct_ep_1500_rew__3492.267_.pth", + "epoch": 1500, + "frame": 98304000, + "task": "franka", + "size": "M", + "run": "franka_M_s0", + "kind": "broken", + "expert": "policies/franka/ppo/franka_M_s0/ep01500.npz", + "damage": { + "kind": "noise", + "arg": 1.0, + "edited": [ + "all W (Gaussian)" + ] + } + }, + "num_envs": 4096, + "seed": 100000, + "success_rate": 0.0146484375, + "mean_return": 105.79911804199219, + "mean_steps_to_success": 323.95001220703125, + "terminated_rate": 0.006591796875, + "mean_max_progress": 0.016215892508625984, + "eval_seconds": 22.278645277023315 + }, + { + "policy": "policies/franka/repair/M_s0/damaged/norm_shift_20.npz", + "meta": { + "source": "logs/rl_games/franka_cabinet_direct/franka_M_s0/nn/last_franka_cabinet_direct_ep_1500_rew__3492.267_.pth", + "epoch": 1500, + "frame": 98304000, + "task": "franka", + "size": "M", + "run": "franka_M_s0", + "kind": "broken", + "expert": "policies/franka/ppo/franka_M_s0/ep01500.npz", + "damage": { + "kind": "norm_shift", + "arg": 20, + "edited": [ + "norm_mean[20]" + ] + } + }, + "num_envs": 4096, + "seed": 100000, + "success_rate": 0.053955078125, + "mean_return": 528.83740234375, + "mean_steps_to_success": 282.93212890625, + "terminated_rate": 0.03271484375, + "mean_max_progress": 0.06341743469238281, + "eval_seconds": 22.49489450454712 + }, + { + "policy": "policies/franka/repair/M_s0/damaged/norm_shift_21.npz", + "meta": { + "source": "logs/rl_games/franka_cabinet_direct/franka_M_s0/nn/last_franka_cabinet_direct_ep_1500_rew__3492.267_.pth", + "epoch": 1500, + "frame": 98304000, + "task": "franka", + "size": "M", + "run": "franka_M_s0", + "kind": "broken", + "expert": "policies/franka/ppo/franka_M_s0/ep01500.npz", + "damage": { + "kind": "norm_shift", + "arg": 21, + "edited": [ + "norm_mean[21]" + ] + } + }, + "num_envs": 4096, + "seed": 100000, + "success_rate": 0.00439453125, + "mean_return": 289.99298095703125, + "mean_steps_to_success": 307.72222900390625, + "terminated_rate": 0.001220703125, + "mean_max_progress": 0.0236709825694561, + "eval_seconds": 21.217690467834473 + }, + { + "policy": "policies/franka/repair/M_s0/damaged/swap_act_0_6.npz", + "meta": { + "source": "logs/rl_games/franka_cabinet_direct/franka_M_s0/nn/last_franka_cabinet_direct_ep_1500_rew__3492.267_.pth", + "epoch": 1500, + "frame": 98304000, + "task": "franka", + "size": "M", + "run": "franka_M_s0", + "kind": "broken", + "expert": "policies/franka/ppo/franka_M_s0/ep01500.npz", + "damage": { + "kind": "swap_act", + "arg": [ + 0, + 6 + ], + "edited": [ + "mu_W rows 0,6", + "mu_b[0],[6]" + ] + } + }, + "num_envs": 4096, + "seed": 100000, + "success_rate": 0.002685546875, + "mean_return": 153.8671417236328, + "mean_steps_to_success": 179.09091186523438, + "terminated_rate": 0.00146484375, + "mean_max_progress": 0.011686110869050026, + "eval_seconds": 21.852539777755737 + }, + { + "policy": "policies/franka/repair/M_s0/damaged/swap_act_1_3.npz", + "meta": { + "source": "logs/rl_games/franka_cabinet_direct/franka_M_s0/nn/last_franka_cabinet_direct_ep_1500_rew__3492.267_.pth", + "epoch": 1500, + "frame": 98304000, + "task": "franka", + "size": "M", + "run": "franka_M_s0", + "kind": "broken", + "expert": "policies/franka/ppo/franka_M_s0/ep01500.npz", + "damage": { + "kind": "swap_act", + "arg": [ + 1, + 3 + ], + "edited": [ + "mu_W rows 1,3", + "mu_b[1],[3]" + ] + } + }, + "num_envs": 4096, + "seed": 100000, + "success_rate": 0.0, + "mean_return": -107.97726440429688, + "mean_steps_to_success": null, + "terminated_rate": 0.0, + "mean_max_progress": 0.00016678044630680233, + "eval_seconds": 17.60513925552368 + }, + { + "policy": "policies/franka/repair/M_s0/natural/collapse_ep500.npz", + "meta": { + "source": "logs/rl_games/franka_cabinet_direct/franka_M_s0/nn/last_franka_cabinet_direct_ep_500_rew_3122.486.pth", + "epoch": 500, + "frame": 32768000, + "task": "franka", + "size": "M", + "run": "franka_M_s0", + "kind": "ppo" + }, + "num_envs": 4096, + "seed": 100000, + "success_rate": 0.01123046875, + "mean_return": 3155.778564453125, + "mean_steps_to_success": 233.978271484375, + "terminated_rate": 0.0, + "mean_max_progress": 0.2718912959098816, + "eval_seconds": 19.483251094818115 + }, + { + "policy": "policies/franka/repair/M_s0/natural/early_ep15.npz", + "meta": { + "source": "logs/rl_games/franka_cabinet_direct/franka_M_s0_early/nn/last_franka_cabinet_direct_ep_15_rew_339.6565.pth", + "epoch": 15, + "frame": 983040, + "task": "franka", + "size": "M", + "run": "franka_M_s0_early", + "kind": "ppo" + }, + "num_envs": 4096, + "seed": 100000, + "success_rate": 0.0068359375, + "mean_return": 630.4943237304688, + "mean_steps_to_success": 128.82144165039062, + "terminated_rate": 0.00634765625, + "mean_max_progress": 0.024646878242492676, + "eval_seconds": 16.68050265312195 + }, + { + "policy": "policies/franka/repair/M_s0/natural/early_ep20.npz", + "meta": { + "source": "logs/rl_games/franka_cabinet_direct/franka_M_s0_early/nn/last_franka_cabinet_direct_ep_20_rew_422.66306.pth", + "epoch": 20, + "frame": 1310720, + "task": "franka", + "size": "M", + "run": "franka_M_s0_early", + "kind": "ppo" + }, + "num_envs": 4096, + "seed": 100000, + "success_rate": 0.15673828125, + "mean_return": 960.3773193359375, + "mean_steps_to_success": 131.97195434570312, + "terminated_rate": 0.099853515625, + "mean_max_progress": 0.11908940225839615, + "eval_seconds": 19.079452991485596 + }, + { + "policy": "policies/franka/repair/M_s0/natural/early_ep40.npz", + "meta": { + "source": "logs/rl_games/franka_cabinet_direct/franka_M_s0_early/nn/last_franka_cabinet_direct_ep_40_rew_394.78815.pth", + "epoch": 40, + "frame": 2621440, + "task": "franka", + "size": "M", + "run": "franka_M_s0_early", + "kind": "ppo" + }, + "num_envs": 4096, + "seed": 100000, + "success_rate": 0.226318359375, + "mean_return": 1287.75634765625, + "mean_steps_to_success": 172.64292907714844, + "terminated_rate": 0.2197265625, + "mean_max_progress": 0.2360568642616272, + "eval_seconds": 18.935123920440674 + } +] diff --git a/metadata/data/robot_policy/baselines/franka_repair_M_prompt.json b/metadata/data/robot_policy/baselines/franka_repair_M_prompt.json new file mode 100644 index 0000000000000000000000000000000000000000..e6453426d2a2025f38fb07f0faf6f25b5c42d167 --- /dev/null +++ b/metadata/data/robot_policy/baselines/franka_repair_M_prompt.json @@ -0,0 +1,390 @@ +[ + { + "policy": "policies/franka/repair/M_s0/damaged/dead_units_0_0p5.npz", + "meta": { + "source": "logs/rl_games/franka_cabinet_direct/franka_M_s0/nn/last_franka_cabinet_direct_ep_1500_rew__3492.267_.pth", + "epoch": 1500, + "frame": 98304000, + "task": "franka", + "size": "M", + "run": "franka_M_s0", + "kind": "broken", + "expert": "policies/franka/ppo/franka_M_s0/ep01500.npz", + "damage": { + "kind": "dead_units", + "arg": [ + 0, + 0.5 + ], + "edited": [ + "L0_W rows / L0_b [0, 1, 2, 3, 4, 5, 10, 12, 13, 14, 18, 21, 23, 24, 26, 28, 35, 36, 40, 41, 45, 49, 50, 53, 55, 56, 58, 60, 62, 66, 68, 69, 70, 71, 72, 73, 74, 79, 80, 82, 83, 85, 86, 88, 90, 91, 98, 101, 104, 105, 108, 109, 112, 113, 115, 116, 120, 122, 127, 130, 132, 134, 136, 137, 138, 139, 140, 141, 145, 146, 147, 148, 153, 155, 158, 159, 165, 166, 167, 168, 170, 172, 174, 177, 180, 181, 182, 184, 185, 186, 187, 189, 190, 192, 194, 195, 198, 200, 203, 204, 205, 206, 207, 209, 211, 213, 217, 218, 223, 224, 226, 228, 229, 234, 235, 236, 237, 239, 240, 241, 242, 243, 246, 248, 249, 250, 251, 254]" + ] + } + }, + "num_envs": 1024, + "seed": 500, + "success_rate": 0.5888671875, + "mean_return": 1749.35302734375, + "mean_steps_to_success": 154.53897094726562, + "terminated_rate": 0.0205078125, + "mean_max_progress": 0.22983703017234802, + "eval_seconds": 16.939881086349487, + "trajectory_file": "policies/franka/repair/M_s0/prompt_rollouts/dead_units_0_0p5_seed500.npz" + }, + { + "policy": "policies/franka/repair/M_s0/damaged/dead_units_1_0p5.npz", + "meta": { + "source": "logs/rl_games/franka_cabinet_direct/franka_M_s0/nn/last_franka_cabinet_direct_ep_1500_rew__3492.267_.pth", + "epoch": 1500, + "frame": 98304000, + "task": "franka", + "size": "M", + "run": "franka_M_s0", + "kind": "broken", + "expert": "policies/franka/ppo/franka_M_s0/ep01500.npz", + "damage": { + "kind": "dead_units", + "arg": [ + 1, + 0.5 + ], + "edited": [ + "L1_W rows / L1_b [1, 2, 4, 7, 8, 9, 11, 12, 14, 15, 19, 21, 22, 30, 32, 33, 35, 36, 40, 41, 43, 44, 50, 52, 53, 55, 56, 60, 61, 66, 67, 69, 73, 76, 77, 80, 81, 83, 85, 86, 87, 90, 91, 92, 95, 96, 97, 98, 99, 102, 103, 105, 108, 109, 112, 113, 114, 116, 117, 118, 120, 123, 125, 127]" + ] + } + }, + "num_envs": 1024, + "seed": 500, + "success_rate": 0.0966796875, + "mean_return": 522.7793579101562, + "mean_steps_to_success": 160.4545440673828, + "terminated_rate": 0.0654296875, + "mean_max_progress": 0.06325940787792206, + "eval_seconds": 17.458275079727173, + "trajectory_file": "policies/franka/repair/M_s0/prompt_rollouts/dead_units_1_0p5_seed500.npz" + }, + { + "policy": "policies/franka/repair/M_s0/damaged/flip_obs_18.npz", + "meta": { + "source": "logs/rl_games/franka_cabinet_direct/franka_M_s0/nn/last_franka_cabinet_direct_ep_1500_rew__3492.267_.pth", + "epoch": 1500, + "frame": 98304000, + "task": "franka", + "size": "M", + "run": "franka_M_s0", + "kind": "broken", + "expert": "policies/franka/ppo/franka_M_s0/ep01500.npz", + "damage": { + "kind": "flip_obs", + "arg": 18, + "edited": [ + "L0_W[:, 18]" + ] + } + }, + "num_envs": 1024, + "seed": 500, + "success_rate": 0.0, + "mean_return": -435.83404541015625, + "mean_steps_to_success": null, + "terminated_rate": 0.0, + "mean_max_progress": 0.0, + "eval_seconds": 13.742571115493774, + "trajectory_file": "policies/franka/repair/M_s0/prompt_rollouts/flip_obs_18_seed500.npz" + }, + { + "policy": "policies/franka/repair/M_s0/damaged/flip_obs_20.npz", + "meta": { + "source": "logs/rl_games/franka_cabinet_direct/franka_M_s0/nn/last_franka_cabinet_direct_ep_1500_rew__3492.267_.pth", + "epoch": 1500, + "frame": 98304000, + "task": "franka", + "size": "M", + "run": "franka_M_s0", + "kind": "broken", + "expert": "policies/franka/ppo/franka_M_s0/ep01500.npz", + "damage": { + "kind": "flip_obs", + "arg": 20, + "edited": [ + "L0_W[:, 20]" + ] + } + }, + "num_envs": 1024, + "seed": 500, + "success_rate": 0.0, + "mean_return": -228.39093017578125, + "mean_steps_to_success": null, + "terminated_rate": 0.0, + "mean_max_progress": 0.0005809733993373811, + "eval_seconds": 14.313901424407959, + "trajectory_file": "policies/franka/repair/M_s0/prompt_rollouts/flip_obs_20_seed500.npz" + }, + { + "policy": "policies/franka/repair/M_s0/damaged/noise_0p5.npz", + "meta": { + "source": "logs/rl_games/franka_cabinet_direct/franka_M_s0/nn/last_franka_cabinet_direct_ep_1500_rew__3492.267_.pth", + "epoch": 1500, + "frame": 98304000, + "task": "franka", + "size": "M", + "run": "franka_M_s0", + "kind": "broken", + "expert": "policies/franka/ppo/franka_M_s0/ep01500.npz", + "damage": { + "kind": "noise", + "arg": 0.5, + "edited": [ + "all W (Gaussian)" + ] + } + }, + "num_envs": 1024, + "seed": 500, + "success_rate": 0.001953125, + "mean_return": -36.63417434692383, + "mean_steps_to_success": 313.0, + "terminated_rate": 0.0009765625, + "mean_max_progress": 0.0031453617848455906, + "eval_seconds": 17.85616660118103, + "trajectory_file": "policies/franka/repair/M_s0/prompt_rollouts/noise_0p5_seed500.npz" + }, + { + "policy": "policies/franka/repair/M_s0/damaged/noise_1p0.npz", + "meta": { + "source": "logs/rl_games/franka_cabinet_direct/franka_M_s0/nn/last_franka_cabinet_direct_ep_1500_rew__3492.267_.pth", + "epoch": 1500, + "frame": 98304000, + "task": "franka", + "size": "M", + "run": "franka_M_s0", + "kind": "broken", + "expert": "policies/franka/ppo/franka_M_s0/ep01500.npz", + "damage": { + "kind": "noise", + "arg": 1.0, + "edited": [ + "all W (Gaussian)" + ] + } + }, + "num_envs": 1024, + "seed": 500, + "success_rate": 0.0107421875, + "mean_return": 108.55462646484375, + "mean_steps_to_success": 317.0, + "terminated_rate": 0.005859375, + "mean_max_progress": 0.015900548547506332, + "eval_seconds": 18.863226175308228, + "trajectory_file": "policies/franka/repair/M_s0/prompt_rollouts/noise_1p0_seed500.npz" + }, + { + "policy": "policies/franka/repair/M_s0/damaged/norm_shift_20.npz", + "meta": { + "source": "logs/rl_games/franka_cabinet_direct/franka_M_s0/nn/last_franka_cabinet_direct_ep_1500_rew__3492.267_.pth", + "epoch": 1500, + "frame": 98304000, + "task": "franka", + "size": "M", + "run": "franka_M_s0", + "kind": "broken", + "expert": "policies/franka/ppo/franka_M_s0/ep01500.npz", + "damage": { + "kind": "norm_shift", + "arg": 20, + "edited": [ + "norm_mean[20]" + ] + } + }, + "num_envs": 1024, + "seed": 500, + "success_rate": 0.052734375, + "mean_return": 567.72314453125, + "mean_steps_to_success": 310.3148193359375, + "terminated_rate": 0.0322265625, + "mean_max_progress": 0.061275433748960495, + "eval_seconds": 17.705883264541626, + "trajectory_file": "policies/franka/repair/M_s0/prompt_rollouts/norm_shift_20_seed500.npz" + }, + { + "policy": "policies/franka/repair/M_s0/damaged/norm_shift_21.npz", + "meta": { + "source": "logs/rl_games/franka_cabinet_direct/franka_M_s0/nn/last_franka_cabinet_direct_ep_1500_rew__3492.267_.pth", + "epoch": 1500, + "frame": 98304000, + "task": "franka", + "size": "M", + "run": "franka_M_s0", + "kind": "broken", + "expert": "policies/franka/ppo/franka_M_s0/ep01500.npz", + "damage": { + "kind": "norm_shift", + "arg": 21, + "edited": [ + "norm_mean[21]" + ] + } + }, + "num_envs": 1024, + "seed": 500, + "success_rate": 0.005859375, + "mean_return": 291.67578125, + "mean_steps_to_success": 259.8333435058594, + "terminated_rate": 0.0009765625, + "mean_max_progress": 0.020002463832497597, + "eval_seconds": 17.92686676979065, + "trajectory_file": "policies/franka/repair/M_s0/prompt_rollouts/norm_shift_21_seed500.npz" + }, + { + "policy": "policies/franka/repair/M_s0/damaged/swap_act_0_6.npz", + "meta": { + "source": "logs/rl_games/franka_cabinet_direct/franka_M_s0/nn/last_franka_cabinet_direct_ep_1500_rew__3492.267_.pth", + "epoch": 1500, + "frame": 98304000, + "task": "franka", + "size": "M", + "run": "franka_M_s0", + "kind": "broken", + "expert": "policies/franka/ppo/franka_M_s0/ep01500.npz", + "damage": { + "kind": "swap_act", + "arg": [ + 0, + 6 + ], + "edited": [ + "mu_W rows 0,6", + "mu_b[0],[6]" + ] + } + }, + "num_envs": 1024, + "seed": 500, + "success_rate": 0.0068359375, + "mean_return": 149.83828735351562, + "mean_steps_to_success": 176.42857360839844, + "terminated_rate": 0.005859375, + "mean_max_progress": 0.012269087135791779, + "eval_seconds": 17.99741291999817, + "trajectory_file": "policies/franka/repair/M_s0/prompt_rollouts/swap_act_0_6_seed500.npz" + }, + { + "policy": "policies/franka/repair/M_s0/damaged/swap_act_1_3.npz", + "meta": { + "source": "logs/rl_games/franka_cabinet_direct/franka_M_s0/nn/last_franka_cabinet_direct_ep_1500_rew__3492.267_.pth", + "epoch": 1500, + "frame": 98304000, + "task": "franka", + "size": "M", + "run": "franka_M_s0", + "kind": "broken", + "expert": "policies/franka/ppo/franka_M_s0/ep01500.npz", + "damage": { + "kind": "swap_act", + "arg": [ + 1, + 3 + ], + "edited": [ + "mu_W rows 1,3", + "mu_b[1],[3]" + ] + } + }, + "num_envs": 1024, + "seed": 500, + "success_rate": 0.0, + "mean_return": -103.7619400024414, + "mean_steps_to_success": null, + "terminated_rate": 0.0, + "mean_max_progress": 0.00016098172636702657, + "eval_seconds": 12.88846731185913, + "trajectory_file": "policies/franka/repair/M_s0/prompt_rollouts/swap_act_1_3_seed500.npz" + }, + { + "policy": "policies/franka/repair/M_s0/natural/collapse_ep500.npz", + "meta": { + "source": "logs/rl_games/franka_cabinet_direct/franka_M_s0/nn/last_franka_cabinet_direct_ep_500_rew_3122.486.pth", + "epoch": 500, + "frame": 32768000, + "task": "franka", + "size": "M", + "run": "franka_M_s0", + "kind": "ppo" + }, + "num_envs": 1024, + "seed": 500, + "success_rate": 0.00390625, + "mean_return": 3157.68896484375, + "mean_steps_to_success": 262.5, + "terminated_rate": 0.0, + "mean_max_progress": 0.2711697220802307, + "eval_seconds": 15.490287065505981, + "trajectory_file": "policies/franka/repair/M_s0/prompt_rollouts/collapse_ep500_seed500.npz" + }, + { + "policy": "policies/franka/repair/M_s0/natural/early_ep15.npz", + "meta": { + "source": "logs/rl_games/franka_cabinet_direct/franka_M_s0_early/nn/last_franka_cabinet_direct_ep_15_rew_339.6565.pth", + "epoch": 15, + "frame": 983040, + "task": "franka", + "size": "M", + "run": "franka_M_s0_early", + "kind": "ppo" + }, + "num_envs": 1024, + "seed": 500, + "success_rate": 0.0078125, + "mean_return": 644.9630126953125, + "mean_steps_to_success": 129.0, + "terminated_rate": 0.005859375, + "mean_max_progress": 0.03225094452500343, + "eval_seconds": 11.715023517608643, + "trajectory_file": "policies/franka/repair/M_s0/prompt_rollouts/early_ep15_seed500.npz" + }, + { + "policy": "policies/franka/repair/M_s0/natural/early_ep20.npz", + "meta": { + "source": "logs/rl_games/franka_cabinet_direct/franka_M_s0_early/nn/last_franka_cabinet_direct_ep_20_rew_422.66306.pth", + "epoch": 20, + "frame": 1310720, + "task": "franka", + "size": "M", + "run": "franka_M_s0_early", + "kind": "ppo" + }, + "num_envs": 1024, + "seed": 500, + "success_rate": 0.138671875, + "mean_return": 962.1654663085938, + "mean_steps_to_success": 131.2957763671875, + "terminated_rate": 0.091796875, + "mean_max_progress": 0.11596935987472534, + "eval_seconds": 16.345983505249023, + "trajectory_file": "policies/franka/repair/M_s0/prompt_rollouts/early_ep20_seed500.npz" + }, + { + "policy": "policies/franka/repair/M_s0/natural/early_ep40.npz", + "meta": { + "source": "logs/rl_games/franka_cabinet_direct/franka_M_s0_early/nn/last_franka_cabinet_direct_ep_40_rew_394.78815.pth", + "epoch": 40, + "frame": 2621440, + "task": "franka", + "size": "M", + "run": "franka_M_s0_early", + "kind": "ppo" + }, + "num_envs": 1024, + "seed": 500, + "success_rate": 0.2431640625, + "mean_return": 1276.7930908203125, + "mean_steps_to_success": 173.34939575195312, + "terminated_rate": 0.2333984375, + "mean_max_progress": 0.23903708159923553, + "eval_seconds": 15.748563528060913, + "trajectory_file": "policies/franka/repair/M_s0/prompt_rollouts/early_ep40_seed500.npz" + } +] diff --git a/metadata/data/robot_policy/baselines/shadow_early_heldout.json b/metadata/data/robot_policy/baselines/shadow_early_heldout.json new file mode 100644 index 0000000000000000000000000000000000000000..4d7699c89a9b084cf922d4719f06a5bfcd32084e --- /dev/null +++ b/metadata/data/robot_policy/baselines/shadow_early_heldout.json @@ -0,0 +1,502 @@ +[ + { + "policy": "policies/shadow/ppo/shadow_L_s0_early/ep00010.npz", + "meta": { + "source": "logs/rl_games/shadow_hand/shadow_L_s0_early/nn/last_shadow_hand_ep_10_rew_-22.706238.pth", + "epoch": 10, + "frame": 1310720, + "task": "shadow", + "size": "L", + "run": "shadow_L_s0_early", + "kind": "ppo" + }, + "num_envs": 2048, + "seed": 100000, + "success_rate": 0.0009765625, + "mean_return": 167.26687622070312, + "mean_steps_to_success": 1.0, + "terminated_rate": 0.0, + "mean_max_progress": null, + "eval_seconds": 18.613210916519165, + "mean_goals": 0.0009765625, + "median_goals": 0.0, + "drop_rate": 0.0, + "mean_survival_frac": 0.9983333349227905, + "mean_best_rot_err_first_goal": 2.082810878753662 + }, + { + "policy": "policies/shadow/ppo/shadow_L_s0_early/ep00020.npz", + "meta": { + "source": "logs/rl_games/shadow_hand/shadow_L_s0_early/nn/last_shadow_hand_ep_20_rew_-37.01027.pth", + "epoch": 20, + "frame": 2621440, + "task": "shadow", + "size": "L", + "run": "shadow_L_s0_early", + "kind": "ppo" + }, + "num_envs": 2048, + "seed": 100000, + "success_rate": 0.0029296875, + "mean_return": 193.64102172851562, + "mean_steps_to_success": 32.333335876464844, + "terminated_rate": 0.0068359375, + "mean_max_progress": null, + "eval_seconds": 18.46553373336792, + "mean_goals": 0.0029296875, + "median_goals": 0.0, + "drop_rate": 0.0068359375, + "mean_survival_frac": 0.9938932657241821, + "mean_best_rot_err_first_goal": 2.02432918548584 + }, + { + "policy": "policies/shadow/ppo/shadow_L_s0_early/ep00030.npz", + "meta": { + "source": "logs/rl_games/shadow_hand/shadow_L_s0_early/nn/last_shadow_hand_ep_30_rew_-45.811543.pth", + "epoch": 30, + "frame": 3932160, + "task": "shadow", + "size": "L", + "run": "shadow_L_s0_early", + "kind": "ppo" + }, + "num_envs": 2048, + "seed": 100000, + "success_rate": 0.0048828125, + "mean_return": 173.8318328857422, + "mean_steps_to_success": 51.20000076293945, + "terminated_rate": 0.205078125, + "mean_max_progress": null, + "eval_seconds": 16.506211042404175, + "mean_goals": 0.0048828125, + "median_goals": 0.0, + "drop_rate": 0.205078125, + "mean_survival_frac": 0.8141211271286011, + "mean_best_rot_err_first_goal": 1.8401607275009155 + }, + { + "policy": "policies/shadow/ppo/shadow_L_s0_early/ep00040.npz", + "meta": { + "source": "logs/rl_games/shadow_hand/shadow_L_s0_early/nn/last_shadow_hand_ep_40_rew_6.339822.pth", + "epoch": 40, + "frame": 5242880, + "task": "shadow", + "size": "L", + "run": "shadow_L_s0_early", + "kind": "ppo" + }, + "num_envs": 2048, + "seed": 100000, + "success_rate": 0.00244140625, + "mean_return": 223.78640747070312, + "mean_steps_to_success": 10.600000381469727, + "terminated_rate": 0.02001953125, + "mean_max_progress": null, + "eval_seconds": 15.556045532226562, + "mean_goals": 0.00244140625, + "median_goals": 0.0, + "drop_rate": 0.02001953125, + "mean_survival_frac": 0.9825894832611084, + "mean_best_rot_err_first_goal": 1.9328758716583252 + }, + { + "policy": "policies/shadow/ppo/shadow_L_s0_early/ep00050.npz", + "meta": { + "source": "logs/rl_games/shadow_hand/shadow_L_s0_early/nn/last_shadow_hand_ep_50_rew_80.160614.pth", + "epoch": 50, + "frame": 6553600, + "task": "shadow", + "size": "L", + "run": "shadow_L_s0_early", + "kind": "ppo" + }, + "num_envs": 2048, + "seed": 100000, + "success_rate": 0.00341796875, + "mean_return": 247.21401977539062, + "mean_steps_to_success": 8.0, + "terminated_rate": 0.00390625, + "mean_max_progress": null, + "eval_seconds": 16.821653127670288, + "mean_goals": 0.00341796875, + "median_goals": 0.0, + "drop_rate": 0.00390625, + "mean_survival_frac": 0.9954313039779663, + "mean_best_rot_err_first_goal": 1.970900058746338 + }, + { + "policy": "policies/shadow/ppo/shadow_L_s0_early/ep00060.npz", + "meta": { + "source": "logs/rl_games/shadow_hand/shadow_L_s0_early/nn/last_shadow_hand_ep_60_rew_119.654564.pth", + "epoch": 60, + "frame": 7864320, + "task": "shadow", + "size": "L", + "run": "shadow_L_s0_early", + "kind": "ppo" + }, + "num_envs": 2048, + "seed": 100000, + "success_rate": 0.0029296875, + "mean_return": 259.8006286621094, + "mean_steps_to_success": 59.0, + "terminated_rate": 0.00390625, + "mean_max_progress": null, + "eval_seconds": 15.721859216690063, + "mean_goals": 0.0029296875, + "median_goals": 0.0, + "drop_rate": 0.00390625, + "mean_survival_frac": 0.9955614805221558, + "mean_best_rot_err_first_goal": 1.960098385810852 + }, + { + "policy": "policies/shadow/ppo/shadow_L_s0_early/ep00070.npz", + "meta": { + "source": "logs/rl_games/shadow_hand/shadow_L_s0_early/nn/last_shadow_hand_ep_70_rew_194.63724.pth", + "epoch": 70, + "frame": 9175040, + "task": "shadow", + "size": "L", + "run": "shadow_L_s0_early", + "kind": "ppo" + }, + "num_envs": 2048, + "seed": 100000, + "success_rate": 0.0029296875, + "mean_return": 272.0272216796875, + "mean_steps_to_success": 41.16666793823242, + "terminated_rate": 0.0009765625, + "mean_max_progress": null, + "eval_seconds": 15.527508020401001, + "mean_goals": 0.0029296875, + "median_goals": 0.0, + "drop_rate": 0.0009765625, + "mean_survival_frac": 0.9975000023841858, + "mean_best_rot_err_first_goal": 1.9193514585494995 + }, + { + "policy": "policies/shadow/ppo/shadow_L_s0_early/ep00080.npz", + "meta": { + "source": "logs/rl_games/shadow_hand/shadow_L_s0_early/nn/last_shadow_hand_ep_80_rew_218.68536.pth", + "epoch": 80, + "frame": 10485760, + "task": "shadow", + "size": "L", + "run": "shadow_L_s0_early", + "kind": "ppo" + }, + "num_envs": 2048, + "seed": 100000, + "success_rate": 0.00732421875, + "mean_return": 286.2230529785156, + "mean_steps_to_success": 18.33333396911621, + "terminated_rate": 0.00146484375, + "mean_max_progress": null, + "eval_seconds": 16.965334177017212, + "mean_goals": 0.00732421875, + "median_goals": 0.0, + "drop_rate": 0.00146484375, + "mean_survival_frac": 0.9973583817481995, + "mean_best_rot_err_first_goal": 1.8658255338668823 + }, + { + "policy": "policies/shadow/ppo/shadow_L_s0_early/ep00090.npz", + "meta": { + "source": "logs/rl_games/shadow_hand/shadow_L_s0_early/nn/last_shadow_hand_ep_90_rew_201.5467.pth", + "epoch": 90, + "frame": 11796480, + "task": "shadow", + "size": "L", + "run": "shadow_L_s0_early", + "kind": "ppo" + }, + "num_envs": 2048, + "seed": 100000, + "success_rate": 0.01318359375, + "mean_return": 296.7847900390625, + "mean_steps_to_success": 39.14814758300781, + "terminated_rate": 0.001953125, + "mean_max_progress": null, + "eval_seconds": 16.968181371688843, + "mean_goals": 0.01318359375, + "median_goals": 0.0, + "drop_rate": 0.001953125, + "mean_survival_frac": 0.9966617822647095, + "mean_best_rot_err_first_goal": 1.8323951959609985 + }, + { + "policy": "policies/shadow/ppo/shadow_L_s0_early/ep00100.npz", + "meta": { + "source": "logs/rl_games/shadow_hand/shadow_L_s0_early/nn/last_shadow_hand_ep_100_rew_290.59326.pth", + "epoch": 100, + "frame": 13107200, + "task": "shadow", + "size": "L", + "run": "shadow_L_s0_early", + "kind": "ppo" + }, + "num_envs": 2048, + "seed": 100000, + "success_rate": 0.01708984375, + "mean_return": 321.2593994140625, + "mean_steps_to_success": 31.828571319580078, + "terminated_rate": 0.00537109375, + "mean_max_progress": null, + "eval_seconds": 15.120555400848389, + "mean_goals": 0.01708984375, + "median_goals": 0.0, + "drop_rate": 0.00537109375, + "mean_survival_frac": 0.9934879541397095, + "mean_best_rot_err_first_goal": 1.781111478805542 + }, + { + "policy": "policies/shadow/ppo/shadow_L_s0_early/ep00110.npz", + "meta": { + "source": "logs/rl_games/shadow_hand/shadow_L_s0_early/nn/last_shadow_hand_ep_110_rew_318.08954.pth", + "epoch": 110, + "frame": 14417920, + "task": "shadow", + "size": "L", + "run": "shadow_L_s0_early", + "kind": "ppo" + }, + "num_envs": 2048, + "seed": 100000, + "success_rate": 0.021484375, + "mean_return": 341.4014892578125, + "mean_steps_to_success": 45.47727584838867, + "terminated_rate": 0.00341796875, + "mean_max_progress": null, + "eval_seconds": 17.18002963066101, + "mean_goals": 0.02197265625, + "median_goals": 0.0, + "drop_rate": 0.00341796875, + "mean_survival_frac": 0.9955078363418579, + "mean_best_rot_err_first_goal": 1.7293307781219482 + }, + { + "policy": "policies/shadow/ppo/shadow_L_s0_early/ep00120.npz", + "meta": { + "source": "logs/rl_games/shadow_hand/shadow_L_s0_early/nn/last_shadow_hand_ep_120_rew_323.98572.pth", + "epoch": 120, + "frame": 15728640, + "task": "shadow", + "size": "L", + "run": "shadow_L_s0_early", + "kind": "ppo" + }, + "num_envs": 2048, + "seed": 100000, + "success_rate": 0.02685546875, + "mean_return": 370.82879638671875, + "mean_steps_to_success": 43.23636245727539, + "terminated_rate": 0.0068359375, + "mean_max_progress": null, + "eval_seconds": 15.29853630065918, + "mean_goals": 0.0283203125, + "median_goals": 0.0, + "drop_rate": 0.0068359375, + "mean_survival_frac": 0.9927636981010437, + "mean_best_rot_err_first_goal": 1.6373438835144043 + }, + { + "policy": "policies/shadow/ppo/shadow_L_s0_early/ep00130.npz", + "meta": { + "source": "logs/rl_games/shadow_hand/shadow_L_s0_early/nn/last_shadow_hand_ep_130_rew_363.20062.pth", + "epoch": 130, + "frame": 17039360, + "task": "shadow", + "size": "L", + "run": "shadow_L_s0_early", + "kind": "ppo" + }, + "num_envs": 2048, + "seed": 100000, + "success_rate": 0.0419921875, + "mean_return": 412.3703308105469, + "mean_steps_to_success": 62.848838806152344, + "terminated_rate": 0.00830078125, + "mean_max_progress": null, + "eval_seconds": 16.71039628982544, + "mean_goals": 0.04345703125, + "median_goals": 0.0, + "drop_rate": 0.00830078125, + "mean_survival_frac": 0.9920613765716553, + "mean_best_rot_err_first_goal": 1.5366895198822021 + }, + { + "policy": "policies/shadow/ppo/shadow_L_s0_early/ep00140.npz", + "meta": { + "source": "logs/rl_games/shadow_hand/shadow_L_s0_early/nn/last_shadow_hand_ep_140_rew_430.50668.pth", + "epoch": 140, + "frame": 18350080, + "task": "shadow", + "size": "L", + "run": "shadow_L_s0_early", + "kind": "ppo" + }, + "num_envs": 2048, + "seed": 100000, + "success_rate": 0.052734375, + "mean_return": 449.4071960449219, + "mean_steps_to_success": 67.1388931274414, + "terminated_rate": 0.0078125, + "mean_max_progress": null, + "eval_seconds": 16.128864526748657, + "mean_goals": 0.052734375, + "median_goals": 0.0, + "drop_rate": 0.0078125, + "mean_survival_frac": 0.9924576282501221, + "mean_best_rot_err_first_goal": 1.443040370941162 + }, + { + "policy": "policies/shadow/ppo/shadow_L_s0_early/ep00150.npz", + "meta": { + "source": "logs/rl_games/shadow_hand/shadow_L_s0_early/nn/last_shadow_hand_ep_150_rew_579.27576.pth", + "epoch": 150, + "frame": 19660800, + "task": "shadow", + "size": "L", + "run": "shadow_L_s0_early", + "kind": "ppo" + }, + "num_envs": 2048, + "seed": 100000, + "success_rate": 0.0830078125, + "mean_return": 501.4048767089844, + "mean_steps_to_success": 91.39411926269531, + "terminated_rate": 0.00927734375, + "mean_max_progress": null, + "eval_seconds": 16.589701890945435, + "mean_goals": 0.08984375, + "median_goals": 0.0, + "drop_rate": 0.00927734375, + "mean_survival_frac": 0.9909513592720032, + "mean_best_rot_err_first_goal": 1.3116565942764282 + }, + { + "policy": "policies/shadow/ppo/shadow_L_s0_early/ep00160.npz", + "meta": { + "source": "logs/rl_games/shadow_hand/shadow_L_s0_early/nn/last_shadow_hand_ep_160_rew_538.89435.pth", + "epoch": 160, + "frame": 20971520, + "task": "shadow", + "size": "L", + "run": "shadow_L_s0_early", + "kind": "ppo" + }, + "num_envs": 2048, + "seed": 100000, + "success_rate": 0.16162109375, + "mean_return": 611.2735595703125, + "mean_steps_to_success": 87.55891418457031, + "terminated_rate": 0.0166015625, + "mean_max_progress": null, + "eval_seconds": 17.363032817840576, + "mean_goals": 0.185546875, + "median_goals": 0.0, + "drop_rate": 0.0166015625, + "mean_survival_frac": 0.985511064529419, + "mean_best_rot_err_first_goal": 1.0482988357543945 + }, + { + "policy": "policies/shadow/ppo/shadow_L_s0_early/ep00170.npz", + "meta": { + "source": "logs/rl_games/shadow_hand/shadow_L_s0_early/nn/last_shadow_hand_ep_170_rew_609.14655.pth", + "epoch": 170, + "frame": 22282240, + "task": "shadow", + "size": "L", + "run": "shadow_L_s0_early", + "kind": "ppo" + }, + "num_envs": 2048, + "seed": 100000, + "success_rate": 0.20556640625, + "mean_return": 747.8974609375, + "mean_steps_to_success": 88.64608764648438, + "terminated_rate": 0.01611328125, + "mean_max_progress": null, + "eval_seconds": 17.667461395263672, + "mean_goals": 0.251953125, + "median_goals": 0.0, + "drop_rate": 0.01611328125, + "mean_survival_frac": 0.9859700798988342, + "mean_best_rot_err_first_goal": 0.7842912673950195 + }, + { + "policy": "policies/shadow/ppo/shadow_L_s0_early/ep00180.npz", + "meta": { + "source": "logs/rl_games/shadow_hand/shadow_L_s0_early/nn/last_shadow_hand_ep_180_rew_813.52704.pth", + "epoch": 180, + "frame": 23592960, + "task": "shadow", + "size": "L", + "run": "shadow_L_s0_early", + "kind": "ppo" + }, + "num_envs": 2048, + "seed": 100000, + "success_rate": 0.2841796875, + "mean_return": 896.87451171875, + "mean_steps_to_success": 88.54639434814453, + "terminated_rate": 0.021484375, + "mean_max_progress": null, + "eval_seconds": 18.710469007492065, + "mean_goals": 0.3671875, + "median_goals": 0.0, + "drop_rate": 0.021484375, + "mean_survival_frac": 0.9820767641067505, + "mean_best_rot_err_first_goal": 0.5636105537414551 + }, + { + "policy": "policies/shadow/ppo/shadow_L_s0_early/ep00190.npz", + "meta": { + "source": "logs/rl_games/shadow_hand/shadow_L_s0_early/nn/last_shadow_hand_ep_190_rew_886.62646.pth", + "epoch": 190, + "frame": 24903680, + "task": "shadow", + "size": "L", + "run": "shadow_L_s0_early", + "kind": "ppo" + }, + "num_envs": 2048, + "seed": 100000, + "success_rate": 0.41748046875, + "mean_return": 1004.9165649414062, + "mean_steps_to_success": 86.34152221679688, + "terminated_rate": 0.029296875, + "mean_max_progress": null, + "eval_seconds": 18.818400621414185, + "mean_goals": 0.6474609375, + "median_goals": 0.0, + "drop_rate": 0.029296875, + "mean_survival_frac": 0.9745743274688721, + "mean_best_rot_err_first_goal": 0.4075639247894287 + }, + { + "policy": "policies/shadow/ppo/shadow_L_s0_early/ep00200.npz", + "meta": { + "source": "logs/rl_games/shadow_hand/shadow_L_s0_early/nn/last_shadow_hand_ep_200_rew__1001.15106_.pth", + "epoch": 200, + "frame": 26214400, + "task": "shadow", + "size": "L", + "run": "shadow_L_s0_early", + "kind": "ppo" + }, + "num_envs": 2048, + "seed": 100000, + "success_rate": 0.49609375, + "mean_return": 1164.235107421875, + "mean_steps_to_success": 83.52362060546875, + "terminated_rate": 0.02783203125, + "mean_max_progress": null, + "eval_seconds": 18.854731798171997, + "mean_goals": 0.87109375, + "median_goals": 0.0, + "drop_rate": 0.02783203125, + "mean_survival_frac": 0.9784138798713684, + "mean_best_rot_err_first_goal": 0.2894991934299469 + } +] diff --git a/metadata/data/robot_policy/baselines/shadow_nvidia_heldout.json b/metadata/data/robot_policy/baselines/shadow_nvidia_heldout.json new file mode 100644 index 0000000000000000000000000000000000000000..35e60525bc192cb1b5889211227707968d63159e --- /dev/null +++ b/metadata/data/robot_policy/baselines/shadow_nvidia_heldout.json @@ -0,0 +1,26 @@ +[ + { + "policy": "policies/shadow/nvidia/pretrained.npz", + "meta": { + "source": "downloads/nvidia_pretrained/Isaac-Repose-Cube-Shadow-Direct-v0/checkpoint.pth", + "epoch": 4193, + "frame": 549584896, + "task": "shadow", + "kind": "nvidia_pretrained", + "env_id": "Isaac-Repose-Cube-Shadow-Direct-v0" + }, + "num_envs": 4096, + "seed": 100000, + "success_rate": 0.99609375, + "mean_return": 8436.625, + "mean_steps_to_success": 22.372304916381836, + "terminated_rate": 0.066650390625, + "mean_max_progress": null, + "eval_seconds": 34.998640298843384, + "mean_goals": 30.94482421875, + "median_goals": 32.0, + "drop_rate": 0.066650390625, + "mean_survival_frac": 0.9641158580780029, + "mean_best_rot_err_first_goal": 0.0032312783878296614 + } +] diff --git a/metadata/data/robot_policy/baselines/shadow_repair_L_heldout.json b/metadata/data/robot_policy/baselines/shadow_repair_L_heldout.json new file mode 100644 index 0000000000000000000000000000000000000000..dc2ace92d0dfa1011d0750aed83db213bb14f6be --- /dev/null +++ b/metadata/data/robot_policy/baselines/shadow_repair_L_heldout.json @@ -0,0 +1,120 @@ +[ + { + "policy": "policies/shadow/repair/L/damaged/norm_shift_69.npz", + "meta": { + "source": "downloads/nvidia_pretrained/Isaac-Repose-Cube-Shadow-Direct-v0/checkpoint.pth", + "epoch": 4193, + "frame": 549584896, + "task": "shadow", + "kind": "broken", + "env_id": "Isaac-Repose-Cube-Shadow-Direct-v0", + "expert": "policies/shadow/nvidia/pretrained.npz", + "damage": { + "kind": "norm_shift", + "arg": 69, + "edited": [ + "norm_mean[69]" + ] + } + }, + "num_envs": 4096, + "seed": 100000, + "success_rate": 0.00732421875, + "mean_return": 19.778594970703125, + "mean_steps_to_success": 79.0666732788086, + "terminated_rate": 0.501953125, + "mean_max_progress": null, + "eval_seconds": 37.518584966659546, + "mean_goals": 0.00732421875, + "median_goals": 0.0, + "drop_rate": 0.501953125, + "mean_survival_frac": 0.727307915687561, + "mean_best_rot_err_first_goal": 0.7544765472412109 + }, + { + "policy": "policies/shadow/repair/L/damaged/swap_act_7_11.npz", + "meta": { + "source": "downloads/nvidia_pretrained/Isaac-Repose-Cube-Shadow-Direct-v0/checkpoint.pth", + "epoch": 4193, + "frame": 549584896, + "task": "shadow", + "kind": "broken", + "env_id": "Isaac-Repose-Cube-Shadow-Direct-v0", + "expert": "policies/shadow/nvidia/pretrained.npz", + "damage": { + "kind": "swap_act", + "arg": [ + 7, + 11 + ], + "edited": [ + "mu_W rows 7,11", + "mu_b[7],[11]" + ] + } + }, + "num_envs": 4096, + "seed": 100000, + "success_rate": 0.005615234375, + "mean_return": -30.24211883544922, + "mean_steps_to_success": 26.869565963745117, + "terminated_rate": 0.58056640625, + "mean_max_progress": null, + "eval_seconds": 23.604309558868408, + "mean_goals": 0.005615234375, + "median_goals": 0.0, + "drop_rate": 0.58056640625, + "mean_survival_frac": 0.4608813524246216, + "mean_best_rot_err_first_goal": 1.5354273319244385 + }, + { + "policy": "policies/shadow/repair/L/natural/early_ep100.npz", + "meta": { + "source": "logs/rl_games/shadow_hand/shadow_L_s0_early/nn/last_shadow_hand_ep_100_rew_290.59326.pth", + "epoch": 100, + "frame": 13107200, + "task": "shadow", + "size": "L", + "run": "shadow_L_s0_early", + "kind": "ppo" + }, + "num_envs": 4096, + "seed": 100000, + "success_rate": 0.016357421875, + "mean_return": 319.2789306640625, + "mean_steps_to_success": 47.70149230957031, + "terminated_rate": 0.00537109375, + "mean_max_progress": null, + "eval_seconds": 23.998307704925537, + "mean_goals": 0.016357421875, + "median_goals": 0.0, + "drop_rate": 0.00537109375, + "mean_survival_frac": 0.994306206703186, + "mean_best_rot_err_first_goal": 1.7719988822937012 + }, + { + "policy": "policies/shadow/repair/L/natural/early_ep200.npz", + "meta": { + "source": "logs/rl_games/shadow_hand/shadow_L_s0_early/nn/last_shadow_hand_ep_200_rew__1001.15106_.pth", + "epoch": 200, + "frame": 26214400, + "task": "shadow", + "size": "L", + "run": "shadow_L_s0_early", + "kind": "ppo" + }, + "num_envs": 4096, + "seed": 100000, + "success_rate": 0.484619140625, + "mean_return": 1132.052001953125, + "mean_steps_to_success": 83.60150909423828, + "terminated_rate": 0.048828125, + "mean_max_progress": null, + "eval_seconds": 29.050029277801514, + "mean_goals": 0.858154296875, + "median_goals": 0.0, + "drop_rate": 0.048828125, + "mean_survival_frac": 0.9592057466506958, + "mean_best_rot_err_first_goal": 0.31097495555877686 + } +] diff --git a/metadata/data/robot_policy/baselines/shadow_repair_L_prompt.json b/metadata/data/robot_policy/baselines/shadow_repair_L_prompt.json new file mode 100644 index 0000000000000000000000000000000000000000..4dde1e74ab4afaaab5727aaec6c7a4f2f66ceb32 --- /dev/null +++ b/metadata/data/robot_policy/baselines/shadow_repair_L_prompt.json @@ -0,0 +1,124 @@ +[ + { + "policy": "policies/shadow/repair/L/damaged/norm_shift_69.npz", + "meta": { + "source": "downloads/nvidia_pretrained/Isaac-Repose-Cube-Shadow-Direct-v0/checkpoint.pth", + "epoch": 4193, + "frame": 549584896, + "task": "shadow", + "kind": "broken", + "env_id": "Isaac-Repose-Cube-Shadow-Direct-v0", + "expert": "policies/shadow/nvidia/pretrained.npz", + "damage": { + "kind": "norm_shift", + "arg": 69, + "edited": [ + "norm_mean[69]" + ] + } + }, + "num_envs": 1024, + "seed": 500, + "success_rate": 0.009765625, + "mean_return": 21.877500534057617, + "mean_steps_to_success": 75.80000305175781, + "terminated_rate": 0.4150390625, + "mean_max_progress": null, + "eval_seconds": 23.037562608718872, + "mean_goals": 0.009765625, + "median_goals": 0.0, + "drop_rate": 0.4150390625, + "mean_survival_frac": 0.792407214641571, + "mean_best_rot_err_first_goal": 0.7359496355056763, + "trajectory_file": "policies/shadow/repair/L/prompt_rollouts/norm_shift_69_seed500.npz" + }, + { + "policy": "policies/shadow/repair/L/damaged/swap_act_7_11.npz", + "meta": { + "source": "downloads/nvidia_pretrained/Isaac-Repose-Cube-Shadow-Direct-v0/checkpoint.pth", + "epoch": 4193, + "frame": 549584896, + "task": "shadow", + "kind": "broken", + "env_id": "Isaac-Repose-Cube-Shadow-Direct-v0", + "expert": "policies/shadow/nvidia/pretrained.npz", + "damage": { + "kind": "swap_act", + "arg": [ + 7, + 11 + ], + "edited": [ + "mu_W rows 7,11", + "mu_b[7],[11]" + ] + } + }, + "num_envs": 1024, + "seed": 500, + "success_rate": 0.0068359375, + "mean_return": -22.779251098632812, + "mean_steps_to_success": 45.000003814697266, + "terminated_rate": 0.5966796875, + "mean_max_progress": null, + "eval_seconds": 17.74618697166443, + "mean_goals": 0.0068359375, + "median_goals": 0.0, + "drop_rate": 0.5966796875, + "mean_survival_frac": 0.44440919160842896, + "mean_best_rot_err_first_goal": 1.5147645473480225, + "trajectory_file": "policies/shadow/repair/L/prompt_rollouts/swap_act_7_11_seed500.npz" + }, + { + "policy": "policies/shadow/repair/L/natural/early_ep100.npz", + "meta": { + "source": "logs/rl_games/shadow_hand/shadow_L_s0_early/nn/last_shadow_hand_ep_100_rew_290.59326.pth", + "epoch": 100, + "frame": 13107200, + "task": "shadow", + "size": "L", + "run": "shadow_L_s0_early", + "kind": "ppo" + }, + "num_envs": 1024, + "seed": 500, + "success_rate": 0.0087890625, + "mean_return": 312.6812744140625, + "mean_steps_to_success": 44.22222137451172, + "terminated_rate": 0.00390625, + "mean_max_progress": null, + "eval_seconds": 14.899807691574097, + "mean_goals": 0.0087890625, + "median_goals": 0.0, + "drop_rate": 0.00390625, + "mean_survival_frac": 0.9947754144668579, + "mean_best_rot_err_first_goal": 1.7847404479980469, + "trajectory_file": "policies/shadow/repair/L/prompt_rollouts/early_ep100_seed500.npz" + }, + { + "policy": "policies/shadow/repair/L/natural/early_ep200.npz", + "meta": { + "source": "logs/rl_games/shadow_hand/shadow_L_s0_early/nn/last_shadow_hand_ep_200_rew__1001.15106_.pth", + "epoch": 200, + "frame": 26214400, + "task": "shadow", + "size": "L", + "run": "shadow_L_s0_early", + "kind": "ppo" + }, + "num_envs": 1024, + "seed": 500, + "success_rate": 0.5029296875, + "mean_return": 1184.4261474609375, + "mean_steps_to_success": 88.2330093383789, + "terminated_rate": 0.015625, + "mean_max_progress": null, + "eval_seconds": 17.18170952796936, + "mean_goals": 0.880859375, + "median_goals": 1.0, + "drop_rate": 0.015625, + "mean_survival_frac": 0.9860286712646484, + "mean_best_rot_err_first_goal": 0.287282258272171, + "trajectory_file": "policies/shadow/repair/L/prompt_rollouts/early_ep200_seed500.npz" + } +] diff --git a/metadata/data/robot_policy/schedules/blind_jobs.jsonl b/metadata/data/robot_policy/schedules/blind_jobs.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..f48f0147c6b7f517bc62e50658b600249d0ed593 --- /dev/null +++ b/metadata/data/robot_policy/schedules/blind_jobs.jsonl @@ -0,0 +1,10 @@ +{"exp": "main", "task": "franka", "size": "M", "cond": "repair", "case": "swap_act_1_3", "model": "claude-opus-5-5", "rules": "A", "tools": "T0", "budget": 0, "init": "policies/franka/repair/M_s0/damaged/swap_act_1_3.npz", "attempt-id": "blind0", "init-eval": "results/baselines/franka_repair_M_prompt.json", "init-rollout": "policies/franka/repair/M_s0/prompt_rollouts/swap_act_1_3_seed500.npz", "max-change-frac": 0.1} +{"exp": "main", "task": "franka", "size": "M", "cond": "repair", "case": "swap_act_1_3", "model": "claude-opus-5-5", "rules": "A", "tools": "T2", "budget": 5, "init": "policies/franka/repair/M_s0/damaged/swap_act_1_3.npz", "attempt-id": "blind0", "init-eval": "results/baselines/franka_repair_M_prompt.json", "init-rollout": "policies/franka/repair/M_s0/prompt_rollouts/swap_act_1_3_seed500.npz", "max-change-frac": 0.1} +{"exp": "main", "task": "franka", "size": "M", "cond": "repair", "case": "swap_act_1_3", "model": "claude-opus-5-5", "rules": "B", "tools": "T0", "budget": 0, "init": "policies/franka/repair/M_s0/damaged/swap_act_1_3.npz", "attempt-id": "blind0", "init-eval": "results/baselines/franka_repair_M_prompt.json", "init-rollout": "policies/franka/repair/M_s0/prompt_rollouts/swap_act_1_3_seed500.npz", "max-change-frac": 0.1} +{"exp": "main", "task": "franka", "size": "M", "cond": "repair", "case": "norm_shift_21", "model": "claude-opus-5-5", "rules": "A", "tools": "T0", "budget": 0, "init": "policies/franka/repair/M_s0/damaged/norm_shift_21.npz", "attempt-id": "blind0", "init-eval": "results/baselines/franka_repair_M_prompt.json", "init-rollout": "policies/franka/repair/M_s0/prompt_rollouts/norm_shift_21_seed500.npz", "max-change-frac": 0.1} +{"exp": "main", "task": "franka", "size": "M", "cond": "repair", "case": "norm_shift_21", "model": "claude-opus-5-5", "rules": "A", "tools": "T2", "budget": 5, "init": "policies/franka/repair/M_s0/damaged/norm_shift_21.npz", "attempt-id": "blind0", "init-eval": "results/baselines/franka_repair_M_prompt.json", "init-rollout": "policies/franka/repair/M_s0/prompt_rollouts/norm_shift_21_seed500.npz", "max-change-frac": 0.1} +{"exp": "main", "task": "franka", "size": "M", "cond": "repair", "case": "norm_shift_21", "model": "claude-opus-5-5", "rules": "B", "tools": "T0", "budget": 0, "init": "policies/franka/repair/M_s0/damaged/norm_shift_21.npz", "attempt-id": "blind0", "init-eval": "results/baselines/franka_repair_M_prompt.json", "init-rollout": "policies/franka/repair/M_s0/prompt_rollouts/norm_shift_21_seed500.npz", "max-change-frac": 0.1} +{"exp": "hands", "task": "allegro", "size": "L", "cond": "repair", "case": "swap_act_4_7", "model": "claude-opus-5-5", "rules": "A", "tools": "T2", "budget": 5, "init": "policies/allegro/repair/L/damaged/swap_act_4_7.npz", "attempt-id": "blind0", "init-eval": "results/baselines/allegro_repair_L_prompt.json", "init-rollout": "policies/allegro/repair/L/prompt_rollouts/swap_act_4_7_seed500.npz", "max-change-frac": 0.1} +{"exp": "hands", "task": "allegro", "size": "L", "cond": "repair", "case": "norm_shift_53", "model": "claude-fable-5-1", "rules": "A", "tools": "T2", "budget": 5, "init": "policies/allegro/repair/L/damaged/norm_shift_53.npz", "attempt-id": "blind0", "init-eval": "results/baselines/allegro_repair_L_prompt.json", "init-rollout": "policies/allegro/repair/L/prompt_rollouts/norm_shift_53_seed500.npz", "max-change-frac": 0.1} +{"exp": "hands", "task": "shadow", "size": "L", "cond": "repair", "case": "swap_act_7_11", "model": "claude-opus-5-5", "rules": "A", "tools": "T2", "budget": 5, "init": "policies/shadow/repair/L/damaged/swap_act_7_11.npz", "attempt-id": "blind0", "init-eval": "results/baselines/shadow_repair_L_prompt.json", "init-rollout": "policies/shadow/repair/L/prompt_rollouts/swap_act_7_11_seed500.npz", "max-change-frac": 0.1} +{"exp": "hands", "task": "shadow", "size": "L", "cond": "repair", "case": "norm_shift_69", "model": "claude-fable-5-1", "rules": "A", "tools": "T2", "budget": 5, "init": "policies/shadow/repair/L/damaged/norm_shift_69.npz", "attempt-id": "blind0", "init-eval": "results/baselines/shadow_repair_L_prompt.json", "init-rollout": "policies/shadow/repair/L/prompt_rollouts/norm_shift_69_seed500.npz", "max-change-frac": 0.1} diff --git a/metadata/data/robot_policy/schedules/hands_jobs.jsonl b/metadata/data/robot_policy/schedules/hands_jobs.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..ed63005e82c4acab492cc759abdde54c5c479d2c --- /dev/null +++ b/metadata/data/robot_policy/schedules/hands_jobs.jsonl @@ -0,0 +1,16 @@ +{"exp": "hands", "task": "allegro", "rules": "A", "tools": "T2", "budget": 5, "attempt-id": "a0", "model": "claude-opus-5-5", "size": "L", "cond": "repair", "case": "swap_act_4_7", "init": "policies/allegro/repair/L/damaged/swap_act_4_7.npz", "init-eval": "results/baselines/allegro_repair_L_prompt.json", "init-rollout": "policies/allegro/repair/L/prompt_rollouts/swap_act_4_7_seed500.npz", "max-change-frac": 0.1} +{"exp": "hands", "task": "allegro", "rules": "A", "tools": "T2", "budget": 5, "attempt-id": "a0", "model": "claude-fable-5-1", "size": "L", "cond": "repair", "case": "norm_shift_53", "init": "policies/allegro/repair/L/damaged/norm_shift_53.npz", "init-eval": "results/baselines/allegro_repair_L_prompt.json", "init-rollout": "policies/allegro/repair/L/prompt_rollouts/norm_shift_53_seed500.npz", "max-change-frac": 0.1} +{"exp": "hands", "task": "allegro", "rules": "A", "tools": "T2", "budget": 5, "attempt-id": "a0", "model": "claude-opus-5-5", "size": "L", "cond": "repair", "case": "early_ep50", "init": "policies/allegro/repair/L/natural/early_ep50.npz", "init-eval": "results/baselines/allegro_repair_L_prompt.json", "init-rollout": "policies/allegro/repair/L/prompt_rollouts/early_ep50_seed500.npz", "max-change-frac": 0.1} +{"exp": "hands", "task": "allegro", "rules": "A", "tools": "T2", "budget": 5, "attempt-id": "a0", "model": "claude-fable-5-1", "size": "L", "cond": "repair", "case": "early_ep50", "init": "policies/allegro/repair/L/natural/early_ep50.npz", "init-eval": "results/baselines/allegro_repair_L_prompt.json", "init-rollout": "policies/allegro/repair/L/prompt_rollouts/early_ep50_seed500.npz", "max-change-frac": 0.1} +{"exp": "hands", "task": "allegro", "rules": "A", "tools": "T2", "budget": 5, "attempt-id": "a0", "model": "claude-opus-5-5", "size": "L", "cond": "repair", "case": "early_ep200", "init": "policies/allegro/repair/L/natural/early_ep200.npz", "init-eval": "results/baselines/allegro_repair_L_prompt.json", "init-rollout": "policies/allegro/repair/L/prompt_rollouts/early_ep200_seed500.npz", "max-change-frac": 0.1} +{"exp": "hands", "task": "allegro", "rules": "A", "tools": "T2", "budget": 5, "attempt-id": "a0", "model": "claude-fable-5-1", "size": "L", "cond": "repair", "case": "early_ep200", "init": "policies/allegro/repair/L/natural/early_ep200.npz", "init-eval": "results/baselines/allegro_repair_L_prompt.json", "init-rollout": "policies/allegro/repair/L/prompt_rollouts/early_ep200_seed500.npz", "max-change-frac": 0.1} +{"exp": "hands", "task": "allegro", "rules": "A", "tools": "T2", "budget": 5, "attempt-id": "a0", "model": "claude-opus-5-5", "size": "XS", "cond": "write"} +{"exp": "hands", "task": "allegro", "rules": "A", "tools": "T2", "budget": 5, "attempt-id": "a0", "model": "claude-fable-5-1", "size": "XS", "cond": "write"} +{"exp": "hands", "task": "shadow", "rules": "A", "tools": "T2", "budget": 5, "attempt-id": "a0", "model": "claude-opus-5-5", "size": "L", "cond": "repair", "case": "swap_act_7_11", "init": "policies/shadow/repair/L/damaged/swap_act_7_11.npz", "init-eval": "results/baselines/shadow_repair_L_prompt.json", "init-rollout": "policies/shadow/repair/L/prompt_rollouts/swap_act_7_11_seed500.npz", "max-change-frac": 0.1} +{"exp": "hands", "task": "shadow", "rules": "A", "tools": "T2", "budget": 5, "attempt-id": "a0", "model": "claude-fable-5-1", "size": "L", "cond": "repair", "case": "norm_shift_69", "init": "policies/shadow/repair/L/damaged/norm_shift_69.npz", "init-eval": "results/baselines/shadow_repair_L_prompt.json", "init-rollout": "policies/shadow/repair/L/prompt_rollouts/norm_shift_69_seed500.npz", "max-change-frac": 0.1} +{"exp": "hands", "task": "shadow", "rules": "A", "tools": "T2", "budget": 5, "attempt-id": "a0", "model": "claude-opus-5-5", "size": "L", "cond": "repair", "case": "early_ep100", "init": "policies/shadow/repair/L/natural/early_ep100.npz", "init-eval": "results/baselines/shadow_repair_L_prompt.json", "init-rollout": "policies/shadow/repair/L/prompt_rollouts/early_ep100_seed500.npz", "max-change-frac": 0.1} +{"exp": "hands", "task": "shadow", "rules": "A", "tools": "T2", "budget": 5, "attempt-id": "a0", "model": "claude-fable-5-1", "size": "L", "cond": "repair", "case": "early_ep100", "init": "policies/shadow/repair/L/natural/early_ep100.npz", "init-eval": "results/baselines/shadow_repair_L_prompt.json", "init-rollout": "policies/shadow/repair/L/prompt_rollouts/early_ep100_seed500.npz", "max-change-frac": 0.1} +{"exp": "hands", "task": "shadow", "rules": "A", "tools": "T2", "budget": 5, "attempt-id": "a0", "model": "claude-opus-5-5", "size": "L", "cond": "repair", "case": "early_ep200", "init": "policies/shadow/repair/L/natural/early_ep200.npz", "init-eval": "results/baselines/shadow_repair_L_prompt.json", "init-rollout": "policies/shadow/repair/L/prompt_rollouts/early_ep200_seed500.npz", "max-change-frac": 0.1} +{"exp": "hands", "task": "shadow", "rules": "A", "tools": "T2", "budget": 5, "attempt-id": "a0", "model": "claude-fable-5-1", "size": "L", "cond": "repair", "case": "early_ep200", "init": "policies/shadow/repair/L/natural/early_ep200.npz", "init-eval": "results/baselines/shadow_repair_L_prompt.json", "init-rollout": "policies/shadow/repair/L/prompt_rollouts/early_ep200_seed500.npz", "max-change-frac": 0.1} +{"exp": "hands", "task": "shadow", "rules": "A", "tools": "T2", "budget": 5, "attempt-id": "a0", "model": "claude-opus-5-5", "size": "XS", "cond": "write"} +{"exp": "hands", "task": "shadow", "rules": "A", "tools": "T2", "budget": 5, "attempt-id": "a0", "model": "claude-fable-5-1", "size": "XS", "cond": "write"} diff --git a/metadata/data/robot_policy/schedules/main_jobs.jsonl b/metadata/data/robot_policy/schedules/main_jobs.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..dfacc532057334671ca61d2521d2e53d562f654f --- /dev/null +++ b/metadata/data/robot_policy/schedules/main_jobs.jsonl @@ -0,0 +1,120 @@ +{"exp": "main", "task": "franka", "model": "claude-opus-5-5", "rules": "A", "tools": "T0", "budget": 0, "attempt-id": "a0", "size": "XS", "cond": "write"} +{"exp": "main", "task": "franka", "model": "claude-opus-5-5", "rules": "A", "tools": "T0", "budget": 0, "attempt-id": "a0", "size": "M", "cond": "repair", "case": "early_ep15", "init": "policies/franka/repair/M_s0/natural/early_ep15.npz", "init-eval": "results/baselines/franka_repair_M_prompt.json", "init-rollout": "policies/franka/repair/M_s0/prompt_rollouts/early_ep15_seed500.npz", "max-change-frac": 0.1} +{"exp": "main", "task": "franka", "model": "claude-opus-5-5", "rules": "A", "tools": "T0", "budget": 0, "attempt-id": "a0", "size": "M", "cond": "repair", "case": "early_ep20", "init": "policies/franka/repair/M_s0/natural/early_ep20.npz", "init-eval": "results/baselines/franka_repair_M_prompt.json", "init-rollout": "policies/franka/repair/M_s0/prompt_rollouts/early_ep20_seed500.npz", "max-change-frac": 0.1} +{"exp": "main", "task": "franka", "model": "claude-opus-5-5", "rules": "A", "tools": "T0", "budget": 0, "attempt-id": "a0", "size": "M", "cond": "repair", "case": "early_ep40", "init": "policies/franka/repair/M_s0/natural/early_ep40.npz", "init-eval": "results/baselines/franka_repair_M_prompt.json", "init-rollout": "policies/franka/repair/M_s0/prompt_rollouts/early_ep40_seed500.npz", "max-change-frac": 0.1} +{"exp": "main", "task": "franka", "model": "claude-opus-5-5", "rules": "A", "tools": "T0", "budget": 0, "attempt-id": "a0", "size": "M", "cond": "repair", "case": "collapse_ep500", "init": "policies/franka/repair/M_s0/natural/collapse_ep500.npz", "init-eval": "results/baselines/franka_repair_M_prompt.json", "init-rollout": "policies/franka/repair/M_s0/prompt_rollouts/collapse_ep500_seed500.npz", "max-change-frac": 0.1} +{"exp": "main", "task": "franka", "model": "claude-opus-5-5", "rules": "A", "tools": "T0", "budget": 0, "attempt-id": "a0", "size": "M", "cond": "repair", "case": "swap_act_1_3", "init": "policies/franka/repair/M_s0/damaged/swap_act_1_3.npz", "init-eval": "results/baselines/franka_repair_M_prompt.json", "init-rollout": "policies/franka/repair/M_s0/prompt_rollouts/swap_act_1_3_seed500.npz", "max-change-frac": 0.1} +{"exp": "main", "task": "franka", "model": "claude-opus-5-5", "rules": "A", "tools": "T0", "budget": 0, "attempt-id": "a0", "size": "M", "cond": "repair", "case": "norm_shift_21", "init": "policies/franka/repair/M_s0/damaged/norm_shift_21.npz", "init-eval": "results/baselines/franka_repair_M_prompt.json", "init-rollout": "policies/franka/repair/M_s0/prompt_rollouts/norm_shift_21_seed500.npz", "max-change-frac": 0.1} +{"exp": "main", "task": "franka", "model": "claude-opus-5-5", "rules": "A", "tools": "T2", "budget": 5, "attempt-id": "a0", "size": "XS", "cond": "write"} +{"exp": "main", "task": "franka", "model": "claude-opus-5-5", "rules": "A", "tools": "T2", "budget": 5, "attempt-id": "a0", "size": "M", "cond": "repair", "case": "early_ep15", "init": "policies/franka/repair/M_s0/natural/early_ep15.npz", "init-eval": "results/baselines/franka_repair_M_prompt.json", "init-rollout": "policies/franka/repair/M_s0/prompt_rollouts/early_ep15_seed500.npz", "max-change-frac": 0.1} +{"exp": "main", "task": "franka", "model": "claude-opus-5-5", "rules": "A", "tools": "T2", "budget": 5, "attempt-id": "a0", "size": "M", "cond": "repair", "case": "early_ep20", "init": "policies/franka/repair/M_s0/natural/early_ep20.npz", "init-eval": "results/baselines/franka_repair_M_prompt.json", "init-rollout": "policies/franka/repair/M_s0/prompt_rollouts/early_ep20_seed500.npz", "max-change-frac": 0.1} +{"exp": "main", "task": "franka", "model": "claude-opus-5-5", "rules": "A", "tools": "T2", "budget": 5, "attempt-id": "a0", "size": "M", "cond": "repair", "case": "early_ep40", "init": "policies/franka/repair/M_s0/natural/early_ep40.npz", "init-eval": "results/baselines/franka_repair_M_prompt.json", "init-rollout": "policies/franka/repair/M_s0/prompt_rollouts/early_ep40_seed500.npz", "max-change-frac": 0.1} +{"exp": "main", "task": "franka", "model": "claude-opus-5-5", "rules": "A", "tools": "T2", "budget": 5, "attempt-id": "a0", "size": "M", "cond": "repair", "case": "collapse_ep500", "init": "policies/franka/repair/M_s0/natural/collapse_ep500.npz", "init-eval": "results/baselines/franka_repair_M_prompt.json", "init-rollout": "policies/franka/repair/M_s0/prompt_rollouts/collapse_ep500_seed500.npz", "max-change-frac": 0.1} +{"exp": "main", "task": "franka", "model": "claude-opus-5-5", "rules": "A", "tools": "T2", "budget": 5, "attempt-id": "a0", "size": "M", "cond": "repair", "case": "swap_act_1_3", "init": "policies/franka/repair/M_s0/damaged/swap_act_1_3.npz", "init-eval": "results/baselines/franka_repair_M_prompt.json", "init-rollout": "policies/franka/repair/M_s0/prompt_rollouts/swap_act_1_3_seed500.npz", "max-change-frac": 0.1} +{"exp": "main", "task": "franka", "model": "claude-opus-5-5", "rules": "A", "tools": "T2", "budget": 5, "attempt-id": "a0", "size": "M", "cond": "repair", "case": "norm_shift_21", "init": "policies/franka/repair/M_s0/damaged/norm_shift_21.npz", "init-eval": "results/baselines/franka_repair_M_prompt.json", "init-rollout": "policies/franka/repair/M_s0/prompt_rollouts/norm_shift_21_seed500.npz", "max-change-frac": 0.1} +{"exp": "main", "task": "franka", "model": "claude-opus-5-5", "rules": "B", "tools": "T0", "budget": 0, "attempt-id": "a0", "size": "XS", "cond": "write"} +{"exp": "main", "task": "franka", "model": "claude-opus-5-5", "rules": "B", "tools": "T0", "budget": 0, "attempt-id": "a0", "size": "M", "cond": "repair", "case": "early_ep15", "init": "policies/franka/repair/M_s0/natural/early_ep15.npz", "init-eval": "results/baselines/franka_repair_M_prompt.json", "init-rollout": "policies/franka/repair/M_s0/prompt_rollouts/early_ep15_seed500.npz", "max-change-frac": 0.1} +{"exp": "main", "task": "franka", "model": "claude-opus-5-5", "rules": "B", "tools": "T0", "budget": 0, "attempt-id": "a0", "size": "M", "cond": "repair", "case": "early_ep20", "init": "policies/franka/repair/M_s0/natural/early_ep20.npz", "init-eval": "results/baselines/franka_repair_M_prompt.json", "init-rollout": "policies/franka/repair/M_s0/prompt_rollouts/early_ep20_seed500.npz", "max-change-frac": 0.1} +{"exp": "main", "task": "franka", "model": "claude-opus-5-5", "rules": "B", "tools": "T0", "budget": 0, "attempt-id": "a0", "size": "M", "cond": "repair", "case": "early_ep40", "init": "policies/franka/repair/M_s0/natural/early_ep40.npz", "init-eval": "results/baselines/franka_repair_M_prompt.json", "init-rollout": "policies/franka/repair/M_s0/prompt_rollouts/early_ep40_seed500.npz", "max-change-frac": 0.1} +{"exp": "main", "task": "franka", "model": "claude-opus-5-5", "rules": "B", "tools": "T0", "budget": 0, "attempt-id": "a0", "size": "M", "cond": "repair", "case": "collapse_ep500", "init": "policies/franka/repair/M_s0/natural/collapse_ep500.npz", "init-eval": "results/baselines/franka_repair_M_prompt.json", "init-rollout": "policies/franka/repair/M_s0/prompt_rollouts/collapse_ep500_seed500.npz", "max-change-frac": 0.1} +{"exp": "main", "task": "franka", "model": "claude-opus-5-5", "rules": "B", "tools": "T0", "budget": 0, "attempt-id": "a0", "size": "M", "cond": "repair", "case": "swap_act_1_3", "init": "policies/franka/repair/M_s0/damaged/swap_act_1_3.npz", "init-eval": "results/baselines/franka_repair_M_prompt.json", "init-rollout": "policies/franka/repair/M_s0/prompt_rollouts/swap_act_1_3_seed500.npz", "max-change-frac": 0.1} +{"exp": "main", "task": "franka", "model": "claude-opus-5-5", "rules": "B", "tools": "T0", "budget": 0, "attempt-id": "a0", "size": "M", "cond": "repair", "case": "norm_shift_21", "init": "policies/franka/repair/M_s0/damaged/norm_shift_21.npz", "init-eval": "results/baselines/franka_repair_M_prompt.json", "init-rollout": "policies/franka/repair/M_s0/prompt_rollouts/norm_shift_21_seed500.npz", "max-change-frac": 0.1} +{"exp": "main", "task": "franka", "model": "claude-opus-5-5", "rules": "B", "tools": "T2", "budget": 5, "attempt-id": "a0", "size": "XS", "cond": "write"} +{"exp": "main", "task": "franka", "model": "claude-opus-5-5", "rules": "B", "tools": "T2", "budget": 5, "attempt-id": "a0", "size": "M", "cond": "repair", "case": "early_ep15", "init": "policies/franka/repair/M_s0/natural/early_ep15.npz", "init-eval": "results/baselines/franka_repair_M_prompt.json", "init-rollout": "policies/franka/repair/M_s0/prompt_rollouts/early_ep15_seed500.npz", "max-change-frac": 0.1} +{"exp": "main", "task": "franka", "model": "claude-opus-5-5", "rules": "B", "tools": "T2", "budget": 5, "attempt-id": "a0", "size": "M", "cond": "repair", "case": "early_ep20", "init": "policies/franka/repair/M_s0/natural/early_ep20.npz", "init-eval": "results/baselines/franka_repair_M_prompt.json", "init-rollout": "policies/franka/repair/M_s0/prompt_rollouts/early_ep20_seed500.npz", "max-change-frac": 0.1} +{"exp": "main", "task": "franka", "model": "claude-opus-5-5", "rules": "B", "tools": "T2", "budget": 5, "attempt-id": "a0", "size": "M", "cond": "repair", "case": "early_ep40", "init": "policies/franka/repair/M_s0/natural/early_ep40.npz", "init-eval": "results/baselines/franka_repair_M_prompt.json", "init-rollout": "policies/franka/repair/M_s0/prompt_rollouts/early_ep40_seed500.npz", "max-change-frac": 0.1} +{"exp": "main", "task": "franka", "model": "claude-opus-5-5", "rules": "B", "tools": "T2", "budget": 5, "attempt-id": "a0", "size": "M", "cond": "repair", "case": "collapse_ep500", "init": "policies/franka/repair/M_s0/natural/collapse_ep500.npz", "init-eval": "results/baselines/franka_repair_M_prompt.json", "init-rollout": "policies/franka/repair/M_s0/prompt_rollouts/collapse_ep500_seed500.npz", "max-change-frac": 0.1} +{"exp": "main", "task": "franka", "model": "claude-opus-5-5", "rules": "B", "tools": "T2", "budget": 5, "attempt-id": "a0", "size": "M", "cond": "repair", "case": "swap_act_1_3", "init": "policies/franka/repair/M_s0/damaged/swap_act_1_3.npz", "init-eval": "results/baselines/franka_repair_M_prompt.json", "init-rollout": "policies/franka/repair/M_s0/prompt_rollouts/swap_act_1_3_seed500.npz", "max-change-frac": 0.1} +{"exp": "main", "task": "franka", "model": "claude-opus-5-5", "rules": "B", "tools": "T2", "budget": 5, "attempt-id": "a0", "size": "M", "cond": "repair", "case": "norm_shift_21", "init": "policies/franka/repair/M_s0/damaged/norm_shift_21.npz", "init-eval": "results/baselines/franka_repair_M_prompt.json", "init-rollout": "policies/franka/repair/M_s0/prompt_rollouts/norm_shift_21_seed500.npz", "max-change-frac": 0.1} +{"exp": "main", "task": "franka", "model": "claude-fable-5-1", "rules": "A", "tools": "T0", "budget": 0, "attempt-id": "a0", "size": "XS", "cond": "write"} +{"exp": "main", "task": "franka", "model": "claude-fable-5-1", "rules": "A", "tools": "T0", "budget": 0, "attempt-id": "a0", "size": "M", "cond": "repair", "case": "early_ep15", "init": "policies/franka/repair/M_s0/natural/early_ep15.npz", "init-eval": "results/baselines/franka_repair_M_prompt.json", "init-rollout": "policies/franka/repair/M_s0/prompt_rollouts/early_ep15_seed500.npz", "max-change-frac": 0.1} +{"exp": "main", "task": "franka", "model": "claude-fable-5-1", "rules": "A", "tools": "T0", "budget": 0, "attempt-id": "a0", "size": "M", "cond": "repair", "case": "early_ep20", "init": "policies/franka/repair/M_s0/natural/early_ep20.npz", "init-eval": "results/baselines/franka_repair_M_prompt.json", "init-rollout": "policies/franka/repair/M_s0/prompt_rollouts/early_ep20_seed500.npz", "max-change-frac": 0.1} +{"exp": "main", "task": "franka", "model": "claude-fable-5-1", "rules": "A", "tools": "T0", "budget": 0, "attempt-id": "a0", "size": "M", "cond": "repair", "case": "early_ep40", "init": "policies/franka/repair/M_s0/natural/early_ep40.npz", "init-eval": "results/baselines/franka_repair_M_prompt.json", "init-rollout": "policies/franka/repair/M_s0/prompt_rollouts/early_ep40_seed500.npz", "max-change-frac": 0.1} +{"exp": "main", "task": "franka", "model": "claude-fable-5-1", "rules": "A", "tools": "T0", "budget": 0, "attempt-id": "a0", "size": "M", "cond": "repair", "case": "collapse_ep500", "init": "policies/franka/repair/M_s0/natural/collapse_ep500.npz", "init-eval": "results/baselines/franka_repair_M_prompt.json", "init-rollout": "policies/franka/repair/M_s0/prompt_rollouts/collapse_ep500_seed500.npz", "max-change-frac": 0.1} +{"exp": "main", "task": "franka", "model": "claude-fable-5-1", "rules": "A", "tools": "T0", "budget": 0, "attempt-id": "a0", "size": "M", "cond": "repair", "case": "swap_act_1_3", "init": "policies/franka/repair/M_s0/damaged/swap_act_1_3.npz", "init-eval": "results/baselines/franka_repair_M_prompt.json", "init-rollout": "policies/franka/repair/M_s0/prompt_rollouts/swap_act_1_3_seed500.npz", "max-change-frac": 0.1} +{"exp": "main", "task": "franka", "model": "claude-fable-5-1", "rules": "A", "tools": "T0", "budget": 0, "attempt-id": "a0", "size": "M", "cond": "repair", "case": "norm_shift_21", "init": "policies/franka/repair/M_s0/damaged/norm_shift_21.npz", "init-eval": "results/baselines/franka_repair_M_prompt.json", "init-rollout": "policies/franka/repair/M_s0/prompt_rollouts/norm_shift_21_seed500.npz", "max-change-frac": 0.1} +{"exp": "main", "task": "franka", "model": "claude-fable-5-1", "rules": "A", "tools": "T2", "budget": 5, "attempt-id": "a0", "size": "XS", "cond": "write"} +{"exp": "main", "task": "franka", "model": "claude-fable-5-1", "rules": "A", "tools": "T2", "budget": 5, "attempt-id": "a0", "size": "M", "cond": "repair", "case": "early_ep15", "init": "policies/franka/repair/M_s0/natural/early_ep15.npz", "init-eval": "results/baselines/franka_repair_M_prompt.json", "init-rollout": "policies/franka/repair/M_s0/prompt_rollouts/early_ep15_seed500.npz", "max-change-frac": 0.1} +{"exp": "main", "task": "franka", "model": "claude-fable-5-1", "rules": "A", "tools": "T2", "budget": 5, "attempt-id": "a0", "size": "M", "cond": "repair", "case": "early_ep20", "init": "policies/franka/repair/M_s0/natural/early_ep20.npz", "init-eval": "results/baselines/franka_repair_M_prompt.json", "init-rollout": "policies/franka/repair/M_s0/prompt_rollouts/early_ep20_seed500.npz", "max-change-frac": 0.1} +{"exp": "main", "task": "franka", "model": "claude-fable-5-1", "rules": "A", "tools": "T2", "budget": 5, "attempt-id": "a0", "size": "M", "cond": "repair", "case": "early_ep40", "init": "policies/franka/repair/M_s0/natural/early_ep40.npz", "init-eval": "results/baselines/franka_repair_M_prompt.json", "init-rollout": "policies/franka/repair/M_s0/prompt_rollouts/early_ep40_seed500.npz", "max-change-frac": 0.1} +{"exp": "main", "task": "franka", "model": "claude-fable-5-1", "rules": "A", "tools": "T2", "budget": 5, "attempt-id": "a0", "size": "M", "cond": "repair", "case": "collapse_ep500", "init": "policies/franka/repair/M_s0/natural/collapse_ep500.npz", "init-eval": "results/baselines/franka_repair_M_prompt.json", "init-rollout": "policies/franka/repair/M_s0/prompt_rollouts/collapse_ep500_seed500.npz", "max-change-frac": 0.1} +{"exp": "main", "task": "franka", "model": "claude-fable-5-1", "rules": "A", "tools": "T2", "budget": 5, "attempt-id": "a0", "size": "M", "cond": "repair", "case": "swap_act_1_3", "init": "policies/franka/repair/M_s0/damaged/swap_act_1_3.npz", "init-eval": "results/baselines/franka_repair_M_prompt.json", "init-rollout": "policies/franka/repair/M_s0/prompt_rollouts/swap_act_1_3_seed500.npz", "max-change-frac": 0.1} +{"exp": "main", "task": "franka", "model": "claude-fable-5-1", "rules": "A", "tools": "T2", "budget": 5, "attempt-id": "a0", "size": "M", "cond": "repair", "case": "norm_shift_21", "init": "policies/franka/repair/M_s0/damaged/norm_shift_21.npz", "init-eval": "results/baselines/franka_repair_M_prompt.json", "init-rollout": "policies/franka/repair/M_s0/prompt_rollouts/norm_shift_21_seed500.npz", "max-change-frac": 0.1} +{"exp": "main", "task": "franka", "model": "claude-fable-5-1", "rules": "B", "tools": "T0", "budget": 0, "attempt-id": "a0", "size": "XS", "cond": "write"} +{"exp": "main", "task": "franka", "model": "claude-fable-5-1", "rules": "B", "tools": "T0", "budget": 0, "attempt-id": "a0", "size": "M", "cond": "repair", "case": "early_ep15", "init": "policies/franka/repair/M_s0/natural/early_ep15.npz", "init-eval": "results/baselines/franka_repair_M_prompt.json", "init-rollout": "policies/franka/repair/M_s0/prompt_rollouts/early_ep15_seed500.npz", "max-change-frac": 0.1} +{"exp": "main", "task": "franka", "model": "claude-fable-5-1", "rules": "B", "tools": "T0", "budget": 0, "attempt-id": "a0", "size": "M", "cond": "repair", "case": "early_ep20", "init": "policies/franka/repair/M_s0/natural/early_ep20.npz", "init-eval": "results/baselines/franka_repair_M_prompt.json", "init-rollout": "policies/franka/repair/M_s0/prompt_rollouts/early_ep20_seed500.npz", "max-change-frac": 0.1} +{"exp": "main", "task": "franka", "model": "claude-fable-5-1", "rules": "B", "tools": "T0", "budget": 0, "attempt-id": "a0", "size": "M", "cond": "repair", "case": "early_ep40", "init": "policies/franka/repair/M_s0/natural/early_ep40.npz", "init-eval": "results/baselines/franka_repair_M_prompt.json", "init-rollout": "policies/franka/repair/M_s0/prompt_rollouts/early_ep40_seed500.npz", "max-change-frac": 0.1} +{"exp": "main", "task": "franka", "model": "claude-fable-5-1", "rules": "B", "tools": "T0", "budget": 0, "attempt-id": "a0", "size": "M", "cond": "repair", "case": "collapse_ep500", "init": "policies/franka/repair/M_s0/natural/collapse_ep500.npz", "init-eval": "results/baselines/franka_repair_M_prompt.json", "init-rollout": "policies/franka/repair/M_s0/prompt_rollouts/collapse_ep500_seed500.npz", "max-change-frac": 0.1} +{"exp": "main", "task": "franka", "model": "claude-fable-5-1", "rules": "B", "tools": "T0", "budget": 0, "attempt-id": "a0", "size": "M", "cond": "repair", "case": "swap_act_1_3", "init": "policies/franka/repair/M_s0/damaged/swap_act_1_3.npz", "init-eval": "results/baselines/franka_repair_M_prompt.json", "init-rollout": "policies/franka/repair/M_s0/prompt_rollouts/swap_act_1_3_seed500.npz", "max-change-frac": 0.1} +{"exp": "main", "task": "franka", "model": "claude-fable-5-1", "rules": "B", "tools": "T0", "budget": 0, "attempt-id": "a0", "size": "M", "cond": "repair", "case": "norm_shift_21", "init": "policies/franka/repair/M_s0/damaged/norm_shift_21.npz", "init-eval": "results/baselines/franka_repair_M_prompt.json", "init-rollout": "policies/franka/repair/M_s0/prompt_rollouts/norm_shift_21_seed500.npz", "max-change-frac": 0.1} +{"exp": "main", "task": "franka", "model": "claude-fable-5-1", "rules": "B", "tools": "T2", "budget": 5, "attempt-id": "a0", "size": "XS", "cond": "write"} +{"exp": "main", "task": "franka", "model": "claude-fable-5-1", "rules": "B", "tools": "T2", "budget": 5, "attempt-id": "a0", "size": "M", "cond": "repair", "case": "early_ep15", "init": "policies/franka/repair/M_s0/natural/early_ep15.npz", "init-eval": "results/baselines/franka_repair_M_prompt.json", "init-rollout": "policies/franka/repair/M_s0/prompt_rollouts/early_ep15_seed500.npz", "max-change-frac": 0.1} +{"exp": "main", "task": "franka", "model": "claude-fable-5-1", "rules": "B", "tools": "T2", "budget": 5, "attempt-id": "a0", "size": "M", "cond": "repair", "case": "early_ep20", "init": "policies/franka/repair/M_s0/natural/early_ep20.npz", "init-eval": "results/baselines/franka_repair_M_prompt.json", "init-rollout": "policies/franka/repair/M_s0/prompt_rollouts/early_ep20_seed500.npz", "max-change-frac": 0.1} +{"exp": "main", "task": "franka", "model": "claude-fable-5-1", "rules": "B", "tools": "T2", "budget": 5, "attempt-id": "a0", "size": "M", "cond": "repair", "case": "early_ep40", "init": "policies/franka/repair/M_s0/natural/early_ep40.npz", "init-eval": "results/baselines/franka_repair_M_prompt.json", "init-rollout": "policies/franka/repair/M_s0/prompt_rollouts/early_ep40_seed500.npz", "max-change-frac": 0.1} +{"exp": "main", "task": "franka", "model": "claude-fable-5-1", "rules": "B", "tools": "T2", "budget": 5, "attempt-id": "a0", "size": "M", "cond": "repair", "case": "collapse_ep500", "init": "policies/franka/repair/M_s0/natural/collapse_ep500.npz", "init-eval": "results/baselines/franka_repair_M_prompt.json", "init-rollout": "policies/franka/repair/M_s0/prompt_rollouts/collapse_ep500_seed500.npz", "max-change-frac": 0.1} +{"exp": "main", "task": "franka", "model": "claude-fable-5-1", "rules": "B", "tools": "T2", "budget": 5, "attempt-id": "a0", "size": "M", "cond": "repair", "case": "swap_act_1_3", "init": "policies/franka/repair/M_s0/damaged/swap_act_1_3.npz", "init-eval": "results/baselines/franka_repair_M_prompt.json", "init-rollout": "policies/franka/repair/M_s0/prompt_rollouts/swap_act_1_3_seed500.npz", "max-change-frac": 0.1} +{"exp": "main", "task": "franka", "model": "claude-fable-5-1", "rules": "B", "tools": "T2", "budget": 5, "attempt-id": "a0", "size": "M", "cond": "repair", "case": "norm_shift_21", "init": "policies/franka/repair/M_s0/damaged/norm_shift_21.npz", "init-eval": "results/baselines/franka_repair_M_prompt.json", "init-rollout": "policies/franka/repair/M_s0/prompt_rollouts/norm_shift_21_seed500.npz", "max-change-frac": 0.1} +{"exp": "main", "task": "franka", "model": "claude-opus-5-5", "rules": "A", "tools": "T0", "budget": 0, "attempt-id": "a1", "size": "XS", "cond": "write"} +{"exp": "main", "task": "franka", "model": "claude-opus-5-5", "rules": "A", "tools": "T0", "budget": 0, "attempt-id": "a1", "size": "M", "cond": "repair", "case": "early_ep15", "init": "policies/franka/repair/M_s0/natural/early_ep15.npz", "init-eval": "results/baselines/franka_repair_M_prompt.json", "init-rollout": "policies/franka/repair/M_s0/prompt_rollouts/early_ep15_seed500.npz", "max-change-frac": 0.1} +{"exp": "main", "task": "franka", "model": "claude-opus-5-5", "rules": "A", "tools": "T0", "budget": 0, "attempt-id": "a1", "size": "M", "cond": "repair", "case": "early_ep20", "init": "policies/franka/repair/M_s0/natural/early_ep20.npz", "init-eval": "results/baselines/franka_repair_M_prompt.json", "init-rollout": "policies/franka/repair/M_s0/prompt_rollouts/early_ep20_seed500.npz", "max-change-frac": 0.1} +{"exp": "main", "task": "franka", "model": "claude-opus-5-5", "rules": "A", "tools": "T0", "budget": 0, "attempt-id": "a1", "size": "M", "cond": "repair", "case": "early_ep40", "init": "policies/franka/repair/M_s0/natural/early_ep40.npz", "init-eval": "results/baselines/franka_repair_M_prompt.json", "init-rollout": "policies/franka/repair/M_s0/prompt_rollouts/early_ep40_seed500.npz", "max-change-frac": 0.1} +{"exp": "main", "task": "franka", "model": "claude-opus-5-5", "rules": "A", "tools": "T0", "budget": 0, "attempt-id": "a1", "size": "M", "cond": "repair", "case": "collapse_ep500", "init": "policies/franka/repair/M_s0/natural/collapse_ep500.npz", "init-eval": "results/baselines/franka_repair_M_prompt.json", "init-rollout": "policies/franka/repair/M_s0/prompt_rollouts/collapse_ep500_seed500.npz", "max-change-frac": 0.1} +{"exp": "main", "task": "franka", "model": "claude-opus-5-5", "rules": "A", "tools": "T0", "budget": 0, "attempt-id": "a1", "size": "M", "cond": "repair", "case": "swap_act_1_3", "init": "policies/franka/repair/M_s0/damaged/swap_act_1_3.npz", "init-eval": "results/baselines/franka_repair_M_prompt.json", "init-rollout": "policies/franka/repair/M_s0/prompt_rollouts/swap_act_1_3_seed500.npz", "max-change-frac": 0.1} +{"exp": "main", "task": "franka", "model": "claude-opus-5-5", "rules": "A", "tools": "T0", "budget": 0, "attempt-id": "a1", "size": "M", "cond": "repair", "case": "norm_shift_21", "init": "policies/franka/repair/M_s0/damaged/norm_shift_21.npz", "init-eval": "results/baselines/franka_repair_M_prompt.json", "init-rollout": "policies/franka/repair/M_s0/prompt_rollouts/norm_shift_21_seed500.npz", "max-change-frac": 0.1} +{"exp": "main", "task": "franka", "model": "claude-opus-5-5", "rules": "A", "tools": "T2", "budget": 5, "attempt-id": "a1", "size": "XS", "cond": "write"} +{"exp": "main", "task": "franka", "model": "claude-opus-5-5", "rules": "A", "tools": "T2", "budget": 5, "attempt-id": "a1", "size": "M", "cond": "repair", "case": "early_ep15", "init": "policies/franka/repair/M_s0/natural/early_ep15.npz", "init-eval": "results/baselines/franka_repair_M_prompt.json", "init-rollout": "policies/franka/repair/M_s0/prompt_rollouts/early_ep15_seed500.npz", "max-change-frac": 0.1} +{"exp": "main", "task": "franka", "model": "claude-opus-5-5", "rules": "A", "tools": "T2", "budget": 5, "attempt-id": "a1", "size": "M", "cond": "repair", "case": "early_ep20", "init": "policies/franka/repair/M_s0/natural/early_ep20.npz", "init-eval": "results/baselines/franka_repair_M_prompt.json", "init-rollout": "policies/franka/repair/M_s0/prompt_rollouts/early_ep20_seed500.npz", "max-change-frac": 0.1} +{"exp": "main", "task": "franka", "model": "claude-opus-5-5", "rules": "A", "tools": "T2", "budget": 5, "attempt-id": "a1", "size": "M", "cond": "repair", "case": "early_ep40", "init": "policies/franka/repair/M_s0/natural/early_ep40.npz", "init-eval": "results/baselines/franka_repair_M_prompt.json", "init-rollout": "policies/franka/repair/M_s0/prompt_rollouts/early_ep40_seed500.npz", "max-change-frac": 0.1} +{"exp": "main", "task": "franka", "model": "claude-opus-5-5", "rules": "A", "tools": "T2", "budget": 5, "attempt-id": "a1", "size": "M", "cond": "repair", "case": "collapse_ep500", "init": "policies/franka/repair/M_s0/natural/collapse_ep500.npz", "init-eval": "results/baselines/franka_repair_M_prompt.json", "init-rollout": "policies/franka/repair/M_s0/prompt_rollouts/collapse_ep500_seed500.npz", "max-change-frac": 0.1} +{"exp": "main", "task": "franka", "model": "claude-opus-5-5", "rules": "A", "tools": "T2", "budget": 5, "attempt-id": "a1", "size": "M", "cond": "repair", "case": "swap_act_1_3", "init": "policies/franka/repair/M_s0/damaged/swap_act_1_3.npz", "init-eval": "results/baselines/franka_repair_M_prompt.json", "init-rollout": "policies/franka/repair/M_s0/prompt_rollouts/swap_act_1_3_seed500.npz", "max-change-frac": 0.1} +{"exp": "main", "task": "franka", "model": "claude-opus-5-5", "rules": "A", "tools": "T2", "budget": 5, "attempt-id": "a1", "size": "M", "cond": "repair", "case": "norm_shift_21", "init": "policies/franka/repair/M_s0/damaged/norm_shift_21.npz", "init-eval": "results/baselines/franka_repair_M_prompt.json", "init-rollout": "policies/franka/repair/M_s0/prompt_rollouts/norm_shift_21_seed500.npz", "max-change-frac": 0.1} +{"exp": "main", "task": "franka", "model": "claude-opus-5-5", "rules": "B", "tools": "T0", "budget": 0, "attempt-id": "a1", "size": "XS", "cond": "write"} +{"exp": "main", "task": "franka", "model": "claude-opus-5-5", "rules": "B", "tools": "T0", "budget": 0, "attempt-id": "a1", "size": "M", "cond": "repair", "case": "early_ep15", "init": "policies/franka/repair/M_s0/natural/early_ep15.npz", "init-eval": "results/baselines/franka_repair_M_prompt.json", "init-rollout": "policies/franka/repair/M_s0/prompt_rollouts/early_ep15_seed500.npz", "max-change-frac": 0.1} +{"exp": "main", "task": "franka", "model": "claude-opus-5-5", "rules": "B", "tools": "T0", "budget": 0, "attempt-id": "a1", "size": "M", "cond": "repair", "case": "early_ep20", "init": "policies/franka/repair/M_s0/natural/early_ep20.npz", "init-eval": "results/baselines/franka_repair_M_prompt.json", "init-rollout": "policies/franka/repair/M_s0/prompt_rollouts/early_ep20_seed500.npz", "max-change-frac": 0.1} +{"exp": "main", "task": "franka", "model": "claude-opus-5-5", "rules": "B", "tools": "T0", "budget": 0, "attempt-id": "a1", "size": "M", "cond": "repair", "case": "early_ep40", "init": "policies/franka/repair/M_s0/natural/early_ep40.npz", "init-eval": "results/baselines/franka_repair_M_prompt.json", "init-rollout": "policies/franka/repair/M_s0/prompt_rollouts/early_ep40_seed500.npz", "max-change-frac": 0.1} +{"exp": "main", "task": "franka", "model": "claude-opus-5-5", "rules": "B", "tools": "T0", "budget": 0, "attempt-id": "a1", "size": "M", "cond": "repair", "case": "collapse_ep500", "init": "policies/franka/repair/M_s0/natural/collapse_ep500.npz", "init-eval": "results/baselines/franka_repair_M_prompt.json", "init-rollout": "policies/franka/repair/M_s0/prompt_rollouts/collapse_ep500_seed500.npz", "max-change-frac": 0.1} +{"exp": "main", "task": "franka", "model": "claude-opus-5-5", "rules": "B", "tools": "T0", "budget": 0, "attempt-id": "a1", "size": "M", "cond": "repair", "case": "swap_act_1_3", "init": "policies/franka/repair/M_s0/damaged/swap_act_1_3.npz", "init-eval": "results/baselines/franka_repair_M_prompt.json", "init-rollout": "policies/franka/repair/M_s0/prompt_rollouts/swap_act_1_3_seed500.npz", "max-change-frac": 0.1} +{"exp": "main", "task": "franka", "model": "claude-opus-5-5", "rules": "B", "tools": "T0", "budget": 0, "attempt-id": "a1", "size": "M", "cond": "repair", "case": "norm_shift_21", "init": "policies/franka/repair/M_s0/damaged/norm_shift_21.npz", "init-eval": "results/baselines/franka_repair_M_prompt.json", "init-rollout": "policies/franka/repair/M_s0/prompt_rollouts/norm_shift_21_seed500.npz", "max-change-frac": 0.1} +{"exp": "main", "task": "franka", "model": "claude-opus-5-5", "rules": "B", "tools": "T2", "budget": 5, "attempt-id": "a1", "size": "XS", "cond": "write"} +{"exp": "main", "task": "franka", "model": "claude-opus-5-5", "rules": "B", "tools": "T2", "budget": 5, "attempt-id": "a1", "size": "M", "cond": "repair", "case": "early_ep15", "init": "policies/franka/repair/M_s0/natural/early_ep15.npz", "init-eval": "results/baselines/franka_repair_M_prompt.json", "init-rollout": "policies/franka/repair/M_s0/prompt_rollouts/early_ep15_seed500.npz", "max-change-frac": 0.1} +{"exp": "main", "task": "franka", "model": "claude-opus-5-5", "rules": "B", "tools": "T2", "budget": 5, "attempt-id": "a1", "size": "M", "cond": "repair", "case": "early_ep20", "init": "policies/franka/repair/M_s0/natural/early_ep20.npz", "init-eval": "results/baselines/franka_repair_M_prompt.json", "init-rollout": "policies/franka/repair/M_s0/prompt_rollouts/early_ep20_seed500.npz", "max-change-frac": 0.1} +{"exp": "main", "task": "franka", "model": "claude-opus-5-5", "rules": "B", "tools": "T2", "budget": 5, "attempt-id": "a1", "size": "M", "cond": "repair", "case": "early_ep40", "init": "policies/franka/repair/M_s0/natural/early_ep40.npz", "init-eval": "results/baselines/franka_repair_M_prompt.json", "init-rollout": "policies/franka/repair/M_s0/prompt_rollouts/early_ep40_seed500.npz", "max-change-frac": 0.1} +{"exp": "main", "task": "franka", "model": "claude-opus-5-5", "rules": "B", "tools": "T2", "budget": 5, "attempt-id": "a1", "size": "M", "cond": "repair", "case": "collapse_ep500", "init": "policies/franka/repair/M_s0/natural/collapse_ep500.npz", "init-eval": "results/baselines/franka_repair_M_prompt.json", "init-rollout": "policies/franka/repair/M_s0/prompt_rollouts/collapse_ep500_seed500.npz", "max-change-frac": 0.1} +{"exp": "main", "task": "franka", "model": "claude-opus-5-5", "rules": "B", "tools": "T2", "budget": 5, "attempt-id": "a1", "size": "M", "cond": "repair", "case": "swap_act_1_3", "init": "policies/franka/repair/M_s0/damaged/swap_act_1_3.npz", "init-eval": "results/baselines/franka_repair_M_prompt.json", "init-rollout": "policies/franka/repair/M_s0/prompt_rollouts/swap_act_1_3_seed500.npz", "max-change-frac": 0.1} +{"exp": "main", "task": "franka", "model": "claude-opus-5-5", "rules": "B", "tools": "T2", "budget": 5, "attempt-id": "a1", "size": "M", "cond": "repair", "case": "norm_shift_21", "init": "policies/franka/repair/M_s0/damaged/norm_shift_21.npz", "init-eval": "results/baselines/franka_repair_M_prompt.json", "init-rollout": "policies/franka/repair/M_s0/prompt_rollouts/norm_shift_21_seed500.npz", "max-change-frac": 0.1} +{"exp": "main", "task": "franka", "model": "claude-fable-5-1", "rules": "A", "tools": "T0", "budget": 0, "attempt-id": "a1", "size": "XS", "cond": "write"} +{"exp": "main", "task": "franka", "model": "claude-fable-5-1", "rules": "A", "tools": "T0", "budget": 0, "attempt-id": "a1", "size": "M", "cond": "repair", "case": "early_ep15", "init": "policies/franka/repair/M_s0/natural/early_ep15.npz", "init-eval": "results/baselines/franka_repair_M_prompt.json", "init-rollout": "policies/franka/repair/M_s0/prompt_rollouts/early_ep15_seed500.npz", "max-change-frac": 0.1} +{"exp": "main", "task": "franka", "model": "claude-fable-5-1", "rules": "A", "tools": "T0", "budget": 0, "attempt-id": "a1", "size": "M", "cond": "repair", "case": "early_ep20", "init": "policies/franka/repair/M_s0/natural/early_ep20.npz", "init-eval": "results/baselines/franka_repair_M_prompt.json", "init-rollout": "policies/franka/repair/M_s0/prompt_rollouts/early_ep20_seed500.npz", "max-change-frac": 0.1} +{"exp": "main", "task": "franka", "model": "claude-fable-5-1", "rules": "A", "tools": "T0", "budget": 0, "attempt-id": "a1", "size": "M", "cond": "repair", "case": "early_ep40", "init": "policies/franka/repair/M_s0/natural/early_ep40.npz", "init-eval": "results/baselines/franka_repair_M_prompt.json", "init-rollout": "policies/franka/repair/M_s0/prompt_rollouts/early_ep40_seed500.npz", "max-change-frac": 0.1} +{"exp": "main", "task": "franka", "model": "claude-fable-5-1", "rules": "A", "tools": "T0", "budget": 0, "attempt-id": "a1", "size": "M", "cond": "repair", "case": "collapse_ep500", "init": "policies/franka/repair/M_s0/natural/collapse_ep500.npz", "init-eval": "results/baselines/franka_repair_M_prompt.json", "init-rollout": "policies/franka/repair/M_s0/prompt_rollouts/collapse_ep500_seed500.npz", "max-change-frac": 0.1} +{"exp": "main", "task": "franka", "model": "claude-fable-5-1", "rules": "A", "tools": "T0", "budget": 0, "attempt-id": "a1", "size": "M", "cond": "repair", "case": "swap_act_1_3", "init": "policies/franka/repair/M_s0/damaged/swap_act_1_3.npz", "init-eval": "results/baselines/franka_repair_M_prompt.json", "init-rollout": "policies/franka/repair/M_s0/prompt_rollouts/swap_act_1_3_seed500.npz", "max-change-frac": 0.1} +{"exp": "main", "task": "franka", "model": "claude-fable-5-1", "rules": "A", "tools": "T0", "budget": 0, "attempt-id": "a1", "size": "M", "cond": "repair", "case": "norm_shift_21", "init": "policies/franka/repair/M_s0/damaged/norm_shift_21.npz", "init-eval": "results/baselines/franka_repair_M_prompt.json", "init-rollout": "policies/franka/repair/M_s0/prompt_rollouts/norm_shift_21_seed500.npz", "max-change-frac": 0.1} +{"exp": "main", "task": "franka", "model": "claude-fable-5-1", "rules": "A", "tools": "T2", "budget": 5, "attempt-id": "a1", "size": "XS", "cond": "write"} +{"exp": "main", "task": "franka", "model": "claude-fable-5-1", "rules": "A", "tools": "T2", "budget": 5, "attempt-id": "a1", "size": "M", "cond": "repair", "case": "early_ep15", "init": "policies/franka/repair/M_s0/natural/early_ep15.npz", "init-eval": "results/baselines/franka_repair_M_prompt.json", "init-rollout": "policies/franka/repair/M_s0/prompt_rollouts/early_ep15_seed500.npz", "max-change-frac": 0.1} +{"exp": "main", "task": "franka", "model": "claude-fable-5-1", "rules": "A", "tools": "T2", "budget": 5, "attempt-id": "a1", "size": "M", "cond": "repair", "case": "early_ep20", "init": "policies/franka/repair/M_s0/natural/early_ep20.npz", "init-eval": "results/baselines/franka_repair_M_prompt.json", "init-rollout": "policies/franka/repair/M_s0/prompt_rollouts/early_ep20_seed500.npz", "max-change-frac": 0.1} +{"exp": "main", "task": "franka", "model": "claude-fable-5-1", "rules": "A", "tools": "T2", "budget": 5, "attempt-id": "a1", "size": "M", "cond": "repair", "case": "early_ep40", "init": "policies/franka/repair/M_s0/natural/early_ep40.npz", "init-eval": "results/baselines/franka_repair_M_prompt.json", "init-rollout": "policies/franka/repair/M_s0/prompt_rollouts/early_ep40_seed500.npz", "max-change-frac": 0.1} +{"exp": "main", "task": "franka", "model": "claude-fable-5-1", "rules": "A", "tools": "T2", "budget": 5, "attempt-id": "a1", "size": "M", "cond": "repair", "case": "collapse_ep500", "init": "policies/franka/repair/M_s0/natural/collapse_ep500.npz", "init-eval": "results/baselines/franka_repair_M_prompt.json", "init-rollout": "policies/franka/repair/M_s0/prompt_rollouts/collapse_ep500_seed500.npz", "max-change-frac": 0.1} +{"exp": "main", "task": "franka", "model": "claude-fable-5-1", "rules": "A", "tools": "T2", "budget": 5, "attempt-id": "a1", "size": "M", "cond": "repair", "case": "swap_act_1_3", "init": "policies/franka/repair/M_s0/damaged/swap_act_1_3.npz", "init-eval": "results/baselines/franka_repair_M_prompt.json", "init-rollout": "policies/franka/repair/M_s0/prompt_rollouts/swap_act_1_3_seed500.npz", "max-change-frac": 0.1} +{"exp": "main", "task": "franka", "model": "claude-fable-5-1", "rules": "A", "tools": "T2", "budget": 5, "attempt-id": "a1", "size": "M", "cond": "repair", "case": "norm_shift_21", "init": "policies/franka/repair/M_s0/damaged/norm_shift_21.npz", "init-eval": "results/baselines/franka_repair_M_prompt.json", "init-rollout": "policies/franka/repair/M_s0/prompt_rollouts/norm_shift_21_seed500.npz", "max-change-frac": 0.1} +{"exp": "main", "task": "franka", "model": "claude-fable-5-1", "rules": "B", "tools": "T0", "budget": 0, "attempt-id": "a1", "size": "XS", "cond": "write"} +{"exp": "main", "task": "franka", "model": "claude-fable-5-1", "rules": "B", "tools": "T0", "budget": 0, "attempt-id": "a1", "size": "M", "cond": "repair", "case": "early_ep15", "init": "policies/franka/repair/M_s0/natural/early_ep15.npz", "init-eval": "results/baselines/franka_repair_M_prompt.json", "init-rollout": "policies/franka/repair/M_s0/prompt_rollouts/early_ep15_seed500.npz", "max-change-frac": 0.1} +{"exp": "main", "task": "franka", "model": "claude-fable-5-1", "rules": "B", "tools": "T0", "budget": 0, "attempt-id": "a1", "size": "M", "cond": "repair", "case": "early_ep20", "init": "policies/franka/repair/M_s0/natural/early_ep20.npz", "init-eval": "results/baselines/franka_repair_M_prompt.json", "init-rollout": "policies/franka/repair/M_s0/prompt_rollouts/early_ep20_seed500.npz", "max-change-frac": 0.1} +{"exp": "main", "task": "franka", "model": "claude-fable-5-1", "rules": "B", "tools": "T0", "budget": 0, "attempt-id": "a1", "size": "M", "cond": "repair", "case": "early_ep40", "init": "policies/franka/repair/M_s0/natural/early_ep40.npz", "init-eval": "results/baselines/franka_repair_M_prompt.json", "init-rollout": "policies/franka/repair/M_s0/prompt_rollouts/early_ep40_seed500.npz", "max-change-frac": 0.1} +{"exp": "main", "task": "franka", "model": "claude-fable-5-1", "rules": "B", "tools": "T0", "budget": 0, "attempt-id": "a1", "size": "M", "cond": "repair", "case": "collapse_ep500", "init": "policies/franka/repair/M_s0/natural/collapse_ep500.npz", "init-eval": "results/baselines/franka_repair_M_prompt.json", "init-rollout": "policies/franka/repair/M_s0/prompt_rollouts/collapse_ep500_seed500.npz", "max-change-frac": 0.1} +{"exp": "main", "task": "franka", "model": "claude-fable-5-1", "rules": "B", "tools": "T0", "budget": 0, "attempt-id": "a1", "size": "M", "cond": "repair", "case": "swap_act_1_3", "init": "policies/franka/repair/M_s0/damaged/swap_act_1_3.npz", "init-eval": "results/baselines/franka_repair_M_prompt.json", "init-rollout": "policies/franka/repair/M_s0/prompt_rollouts/swap_act_1_3_seed500.npz", "max-change-frac": 0.1} +{"exp": "main", "task": "franka", "model": "claude-fable-5-1", "rules": "B", "tools": "T0", "budget": 0, "attempt-id": "a1", "size": "M", "cond": "repair", "case": "norm_shift_21", "init": "policies/franka/repair/M_s0/damaged/norm_shift_21.npz", "init-eval": "results/baselines/franka_repair_M_prompt.json", "init-rollout": "policies/franka/repair/M_s0/prompt_rollouts/norm_shift_21_seed500.npz", "max-change-frac": 0.1} +{"exp": "main", "task": "franka", "model": "claude-fable-5-1", "rules": "B", "tools": "T2", "budget": 5, "attempt-id": "a1", "size": "XS", "cond": "write"} +{"exp": "main", "task": "franka", "model": "claude-fable-5-1", "rules": "B", "tools": "T2", "budget": 5, "attempt-id": "a1", "size": "M", "cond": "repair", "case": "early_ep15", "init": "policies/franka/repair/M_s0/natural/early_ep15.npz", "init-eval": "results/baselines/franka_repair_M_prompt.json", "init-rollout": "policies/franka/repair/M_s0/prompt_rollouts/early_ep15_seed500.npz", "max-change-frac": 0.1} +{"exp": "main", "task": "franka", "model": "claude-fable-5-1", "rules": "B", "tools": "T2", "budget": 5, "attempt-id": "a1", "size": "M", "cond": "repair", "case": "early_ep20", "init": "policies/franka/repair/M_s0/natural/early_ep20.npz", "init-eval": "results/baselines/franka_repair_M_prompt.json", "init-rollout": "policies/franka/repair/M_s0/prompt_rollouts/early_ep20_seed500.npz", "max-change-frac": 0.1} +{"exp": "main", "task": "franka", "model": "claude-fable-5-1", "rules": "B", "tools": "T2", "budget": 5, "attempt-id": "a1", "size": "M", "cond": "repair", "case": "early_ep40", "init": "policies/franka/repair/M_s0/natural/early_ep40.npz", "init-eval": "results/baselines/franka_repair_M_prompt.json", "init-rollout": "policies/franka/repair/M_s0/prompt_rollouts/early_ep40_seed500.npz", "max-change-frac": 0.1} +{"exp": "main", "task": "franka", "model": "claude-fable-5-1", "rules": "B", "tools": "T2", "budget": 5, "attempt-id": "a1", "size": "M", "cond": "repair", "case": "collapse_ep500", "init": "policies/franka/repair/M_s0/natural/collapse_ep500.npz", "init-eval": "results/baselines/franka_repair_M_prompt.json", "init-rollout": "policies/franka/repair/M_s0/prompt_rollouts/collapse_ep500_seed500.npz", "max-change-frac": 0.1} +{"exp": "main", "task": "franka", "model": "claude-fable-5-1", "rules": "B", "tools": "T2", "budget": 5, "attempt-id": "a1", "size": "M", "cond": "repair", "case": "swap_act_1_3", "init": "policies/franka/repair/M_s0/damaged/swap_act_1_3.npz", "init-eval": "results/baselines/franka_repair_M_prompt.json", "init-rollout": "policies/franka/repair/M_s0/prompt_rollouts/swap_act_1_3_seed500.npz", "max-change-frac": 0.1} +{"exp": "main", "task": "franka", "model": "claude-fable-5-1", "rules": "B", "tools": "T2", "budget": 5, "attempt-id": "a1", "size": "M", "cond": "repair", "case": "norm_shift_21", "init": "policies/franka/repair/M_s0/damaged/norm_shift_21.npz", "init-eval": "results/baselines/franka_repair_M_prompt.json", "init-rollout": "policies/franka/repair/M_s0/prompt_rollouts/norm_shift_21_seed500.npz", "max-change-frac": 0.1} +{"exp": "main", "task": "franka", "model": "claude-opus-5-5", "rules": "A", "tools": "T0", "budget": 0, "attempt-id": "a2", "size": "XS", "cond": "write"} +{"exp": "main", "task": "franka", "model": "claude-opus-5-5", "rules": "A", "tools": "T2", "budget": 5, "attempt-id": "a2", "size": "XS", "cond": "write"} +{"exp": "main", "task": "franka", "model": "claude-opus-5-5", "rules": "B", "tools": "T0", "budget": 0, "attempt-id": "a2", "size": "XS", "cond": "write"} +{"exp": "main", "task": "franka", "model": "claude-opus-5-5", "rules": "B", "tools": "T2", "budget": 5, "attempt-id": "a2", "size": "XS", "cond": "write"} +{"exp": "main", "task": "franka", "model": "claude-fable-5-1", "rules": "A", "tools": "T0", "budget": 0, "attempt-id": "a2", "size": "XS", "cond": "write"} +{"exp": "main", "task": "franka", "model": "claude-fable-5-1", "rules": "A", "tools": "T2", "budget": 5, "attempt-id": "a2", "size": "XS", "cond": "write"} +{"exp": "main", "task": "franka", "model": "claude-fable-5-1", "rules": "B", "tools": "T0", "budget": 0, "attempt-id": "a2", "size": "XS", "cond": "write"} +{"exp": "main", "task": "franka", "model": "claude-fable-5-1", "rules": "B", "tools": "T2", "budget": 5, "attempt-id": "a2", "size": "XS", "cond": "write"} diff --git a/metadata/data/robot_policy/summary.json b/metadata/data/robot_policy/summary.json new file mode 100644 index 0000000000000000000000000000000000000000..554523366b4e70fc0527f5826cc06bcb80eae54f --- /dev/null +++ b/metadata/data/robot_policy/summary.json @@ -0,0 +1,1506 @@ +{ + "generated_at": "2026-11-05T00:19:33.674929+00:00", + "runs": [ + { + "dir": "data/robot_policy/attempts/main/franka_M_repair_collapse_ep500/opus-5-5/A_T0_K0_cap10/a0", + "status": "scored", + "counted": true, + "task": "franka", + "cond": "repair", + "case": "collapse_ep500", + "kind": "natural", + "model": "opus-5-5", + "rules": "A", + "tools": "T0", + "blind": false, + "metric": "success_rate", + "after": 0.947509765625, + "after_drop": null, + "before": 0.01123046875, + "before_drop": null, + "best_rot_err_first_goal": null, + "frac_changed": 0.016942433283952677, + "n_changed": 812, + "rollouts_used": null, + "minutes": 56.6, + "output_tokens": 228621, + "audit": "compliant", + "audit_compliant": true, + "doc": "docs/markdowns/franka/collapse_ep500__opus-5-5__A_T0.md" + }, + { + "dir": "data/robot_policy/attempts/main/franka_M_repair_collapse_ep500/opus-5-5/A_T2_K5_cap10/a0", + "status": "scored", + "counted": true, + "task": "franka", + "cond": "repair", + "case": "collapse_ep500", + "kind": "natural", + "model": "opus-5-5", + "rules": "A", + "tools": "T2", + "blind": false, + "metric": "success_rate", + "after": 1.0, + "after_drop": null, + "before": 0.01123046875, + "before_drop": null, + "best_rot_err_first_goal": null, + "frac_changed": 0.01685897302147015, + "n_changed": 808, + "rollouts_used": 1, + "minutes": 10.4, + "output_tokens": 49402, + "audit": "compliant", + "audit_compliant": true, + "doc": "docs/markdowns/franka/collapse_ep500__opus-5-5__A_T2.md" + }, + { + "dir": "data/robot_policy/attempts/main/franka_M_repair_early_ep15/opus-5-5/A_T0_K0_cap10/a0", + "status": "scored", + "counted": true, + "task": "franka", + "cond": "repair", + "case": "early_ep15", + "kind": "natural", + "model": "opus-5-5", + "rules": "A", + "tools": "T0", + "blind": false, + "metric": "success_rate", + "after": 0.00439453125, + "after_drop": null, + "before": 0.0068359375, + "before_drop": null, + "best_rot_err_first_goal": null, + "frac_changed": 0.09055438479354018, + "n_changed": 4340, + "rollouts_used": null, + "minutes": 48.9, + "output_tokens": 245900, + "audit": "compliant", + "audit_compliant": true, + "doc": "docs/markdowns/franka/early_ep15__opus-5-5__A_T0.md" + }, + { + "dir": "data/robot_policy/attempts/main/franka_M_repair_early_ep15/opus-5-5/A_T2_K5_cap10/a0", + "status": "scored", + "counted": true, + "task": "franka", + "cond": "repair", + "case": "early_ep15", + "kind": "natural", + "model": "opus-5-5", + "rules": "A", + "tools": "T2", + "blind": false, + "metric": "success_rate", + "after": 0.99853515625, + "after_drop": null, + "before": 0.0068359375, + "before_drop": null, + "best_rot_err_first_goal": null, + "frac_changed": 0.06215703048386087, + "n_changed": 2979, + "rollouts_used": 5, + "minutes": 45.6, + "output_tokens": 250428, + "audit": "compliant", + "audit_compliant": true, + "doc": "docs/markdowns/franka/early_ep15__opus-5-5__A_T2.md" + }, + { + "dir": "data/robot_policy/attempts/main/franka_M_repair_early_ep20/opus-5-5/A_T0_K0_cap10/a0", + "status": "scored", + "counted": true, + "task": "franka", + "cond": "repair", + "case": "early_ep20", + "kind": "natural", + "model": "opus-5-5", + "rules": "A", + "tools": "T0", + "blind": false, + "metric": "success_rate", + "after": 0.92529296875, + "after_drop": null, + "before": 0.15673828125, + "before_drop": null, + "best_rot_err_first_goal": null, + "frac_changed": 0.0861101258163457, + "n_changed": 4127, + "rollouts_used": null, + "minutes": 62.4, + "output_tokens": 378662, + "audit": "compliant", + "audit_compliant": true, + "doc": "docs/markdowns/franka/early_ep20__opus-5-5__A_T0.md" + }, + { + "dir": "data/robot_policy/attempts/main/franka_M_repair_early_ep20/opus-5-5/A_T2_K5_cap10/a0", + "status": "scored", + "counted": true, + "task": "franka", + "cond": "repair", + "case": "early_ep20", + "kind": "natural", + "model": "opus-5-5", + "rules": "A", + "tools": "T2", + "blind": false, + "metric": "success_rate", + "after": 0.73583984375, + "after_drop": null, + "before": 0.15673828125, + "before_drop": null, + "best_rot_err_first_goal": null, + "frac_changed": 0.06998143009159764, + "n_changed": 3354, + "rollouts_used": 5, + "minutes": 38.6, + "output_tokens": 201793, + "audit": "compliant", + "audit_compliant": true, + "doc": "docs/markdowns/franka/early_ep20__opus-5-5__A_T2.md" + }, + { + "dir": "data/robot_policy/attempts/main/franka_M_repair_early_ep40/opus-5-5/A_T0_K0_cap10/a0", + "status": "scored", + "counted": true, + "task": "franka", + "cond": "repair", + "case": "early_ep40", + "kind": "natural", + "model": "opus-5-5", + "rules": "A", + "tools": "T0", + "blind": false, + "metric": "success_rate", + "after": 0.00048828125, + "after_drop": null, + "before": 0.226318359375, + "before_drop": null, + "best_rot_err_first_goal": null, + "frac_changed": 0.05097335531120245, + "n_changed": 2443, + "rollouts_used": null, + "minutes": 54.7, + "output_tokens": 322525, + "audit": "compliant", + "audit_compliant": true, + "doc": "docs/markdowns/franka/early_ep40__opus-5-5__A_T0.md" + }, + { + "dir": "data/robot_policy/attempts/main/franka_M_repair_norm_shift_21/opus-5-5/A_T0_K0_cap10/a0", + "status": "scored", + "counted": true, + "task": "franka", + "cond": "repair", + "case": "norm_shift_21", + "kind": "damaged", + "model": "opus-5-5", + "rules": "A", + "tools": "T0", + "blind": false, + "metric": "success_rate", + "after": 0.94189453125, + "after_drop": null, + "before": 0.00439453125, + "before_drop": null, + "best_rot_err_first_goal": null, + "frac_changed": 2.0865065620631376e-05, + "n_changed": 1, + "rollouts_used": null, + "minutes": 13.6, + "output_tokens": 78667, + "audit": "compliant", + "audit_compliant": true, + "doc": "docs/markdowns/franka/norm_shift_21__opus-5-5__A_T0.md" + }, + { + "dir": "data/robot_policy/attempts/main/franka_M_repair_norm_shift_21/opus-5-5/A_T0_K0_cap10/blind0", + "status": "rate_limited", + "counted": false + }, + { + "dir": "data/robot_policy/attempts/main/franka_M_repair_norm_shift_21/opus-5-5/A_T2_K5_cap10/a0", + "status": "scored", + "counted": true, + "task": "franka", + "cond": "repair", + "case": "norm_shift_21", + "kind": "damaged", + "model": "opus-5-5", + "rules": "A", + "tools": "T2", + "blind": false, + "metric": "success_rate", + "after": 0.99951171875, + "after_drop": null, + "before": 0.00439453125, + "before_drop": null, + "best_rot_err_first_goal": null, + "frac_changed": 2.0865065620631376e-05, + "n_changed": 1, + "rollouts_used": 2, + "minutes": 15.6, + "output_tokens": 78349, + "audit": "compliant", + "audit_compliant": true, + "doc": "docs/markdowns/franka/norm_shift_21__opus-5-5__A_T2.md" + }, + { + "dir": "data/robot_policy/attempts/main/franka_M_repair_norm_shift_21/opus-5-5/A_T2_K5_cap10/blind0", + "status": "rate_limited", + "counted": false + }, + { + "dir": "data/robot_policy/attempts/main/franka_M_repair_norm_shift_21/opus-5-5/B_T0_K0_cap10/a0", + "status": "scored", + "counted": true, + "task": "franka", + "cond": "repair", + "case": "norm_shift_21", + "kind": "damaged", + "model": "opus-5-5", + "rules": "B", + "tools": "T0", + "blind": false, + "metric": "success_rate", + "after": 0.49609375, + "after_drop": null, + "before": 0.00439453125, + "before_drop": null, + "best_rot_err_first_goal": null, + "frac_changed": 2.0865065620631376e-05, + "n_changed": 1, + "rollouts_used": null, + "minutes": 12.2, + "output_tokens": 69006, + "audit": "compliant", + "audit_compliant": true, + "doc": "docs/markdowns/franka/norm_shift_21__opus-5-5__B_T0.md" + }, + { + "dir": "data/robot_policy/attempts/main/franka_M_repair_norm_shift_21/opus-5-5/B_T0_K0_cap10/blind0", + "status": "scored", + "counted": true, + "task": "franka", + "cond": "repair", + "case": "norm_shift_21", + "kind": "damaged", + "model": "opus-5-5", + "rules": "B", + "tools": "T0", + "blind": true, + "metric": "success_rate", + "after": 0.6142578125, + "after_drop": null, + "before": 0.00439453125, + "before_drop": null, + "best_rot_err_first_goal": null, + "frac_changed": 2.0865065620631376e-05, + "n_changed": 1, + "rollouts_used": null, + "minutes": 14.6, + "output_tokens": 86719, + "audit": "compliant", + "audit_compliant": true, + "doc": null + }, + { + "dir": "data/robot_policy/attempts/main/franka_M_repair_swap_act_1_3/opus-5-5/A_T0_K0_cap10/a0", + "status": "scored", + "counted": true, + "task": "franka", + "cond": "repair", + "case": "swap_act_1_3", + "kind": "damaged", + "model": "opus-5-5", + "rules": "A", + "tools": "T0", + "blind": false, + "metric": "success_rate", + "after": 0.999755859375, + "after_drop": null, + "before": 0.0, + "before_drop": null, + "best_rot_err_first_goal": null, + "frac_changed": 0.002712458530682079, + "n_changed": 130, + "rollouts_used": null, + "minutes": 4.0, + "output_tokens": 19045, + "audit": "compliant", + "audit_compliant": true, + "doc": "docs/markdowns/franka/swap_act_1_3__opus-5-5__A_T0.md" + }, + { + "dir": "data/robot_policy/attempts/main/franka_M_repair_swap_act_1_3/opus-5-5/A_T0_K0_cap10/blind0", + "status": "scored", + "counted": true, + "task": "franka", + "cond": "repair", + "case": "swap_act_1_3", + "kind": "damaged", + "model": "opus-5-5", + "rules": "A", + "tools": "T0", + "blind": true, + "metric": "success_rate", + "after": 0.0, + "after_drop": null, + "before": 0.0, + "before_drop": null, + "best_rot_err_first_goal": null, + "frac_changed": 0.000166920524965051, + "n_changed": 8, + "rollouts_used": null, + "minutes": 29.4, + "output_tokens": 180572, + "audit": "audit_failed", + "audit_compliant": false, + "doc": null + }, + { + "dir": "data/robot_policy/attempts/main/franka_M_repair_swap_act_1_3/opus-5-5/A_T2_K5_cap10/a0", + "status": "scored", + "counted": true, + "task": "franka", + "cond": "repair", + "case": "swap_act_1_3", + "kind": "damaged", + "model": "opus-5-5", + "rules": "A", + "tools": "T2", + "blind": false, + "metric": "success_rate", + "after": 0.999755859375, + "after_drop": null, + "before": 0.0, + "before_drop": null, + "best_rot_err_first_goal": null, + "frac_changed": 0.002712458530682079, + "n_changed": 130, + "rollouts_used": 1, + "minutes": 3.2, + "output_tokens": 7588, + "audit": "compliant", + "audit_compliant": true, + "doc": "docs/markdowns/franka/swap_act_1_3__opus-5-5__A_T2.md" + }, + { + "dir": "data/robot_policy/attempts/main/franka_M_repair_swap_act_1_3/opus-5-5/A_T2_K5_cap10/blind0", + "status": "scored", + "counted": true, + "task": "franka", + "cond": "repair", + "case": "swap_act_1_3", + "kind": "damaged", + "model": "opus-5-5", + "rules": "A", + "tools": "T2", + "blind": true, + "metric": "success_rate", + "after": 0.999755859375, + "after_drop": null, + "before": 0.0, + "before_drop": null, + "best_rot_err_first_goal": null, + "frac_changed": 0.002712458530682079, + "n_changed": 130, + "rollouts_used": 1, + "minutes": 8.2, + "output_tokens": 40486, + "audit": "compliant", + "audit_compliant": true, + "doc": null + }, + { + "dir": "data/robot_policy/attempts/main/franka_M_repair_swap_act_1_3/opus-5-5/B_T0_K0_cap10/a0", + "status": "scored", + "counted": true, + "task": "franka", + "cond": "repair", + "case": "swap_act_1_3", + "kind": "damaged", + "model": "opus-5-5", + "rules": "B", + "tools": "T0", + "blind": false, + "metric": "success_rate", + "after": 0.999755859375, + "after_drop": null, + "before": 0.0, + "before_drop": null, + "best_rot_err_first_goal": null, + "frac_changed": 0.002712458530682079, + "n_changed": 130, + "rollouts_used": null, + "minutes": 2.0, + "output_tokens": 5502, + "audit": "compliant", + "audit_compliant": true, + "doc": "docs/markdowns/franka/swap_act_1_3__opus-5-5__B_T0.md" + }, + { + "dir": "data/robot_policy/attempts/main/franka_M_repair_swap_act_1_3/opus-5-5/B_T0_K0_cap10/blind0", + "status": "scored", + "counted": true, + "task": "franka", + "cond": "repair", + "case": "swap_act_1_3", + "kind": "damaged", + "model": "opus-5-5", + "rules": "B", + "tools": "T0", + "blind": true, + "metric": "success_rate", + "after": 0.013916015625, + "after_drop": null, + "before": 0.0, + "before_drop": null, + "best_rot_err_first_goal": null, + "frac_changed": 0.037369332526550794, + "n_changed": 1791, + "rollouts_used": null, + "minutes": 41.4, + "output_tokens": 248819, + "audit": "audit_failed", + "audit_compliant": false, + "doc": null + }, + { + "dir": "data/robot_policy/attempts/main/franka_XS_write/opus-5-5/A_T0_K0/a0", + "status": "scored", + "counted": true, + "task": "franka", + "cond": "write", + "case": null, + "kind": null, + "model": "opus-5-5", + "rules": "A", + "tools": "T0", + "blind": false, + "metric": "success_rate", + "after": 0.0, + "after_drop": null, + "before": 0.0, + "before_drop": null, + "best_rot_err_first_goal": null, + "frac_changed": 0.06164058795637743, + "n_changed": 390, + "rollouts_used": null, + "minutes": 59.2, + "output_tokens": 233109, + "audit": "compliant", + "audit_compliant": true, + "doc": null + }, + { + "dir": "data/robot_policy/attempts/main/franka_XS_write/opus-5-5/A_T2_K5/a0", + "status": "scored", + "counted": true, + "task": "franka", + "cond": "write", + "case": null, + "kind": null, + "model": "opus-5-5", + "rules": "A", + "tools": "T2", + "blind": false, + "metric": "success_rate", + "after": 0.99853515625, + "after_drop": null, + "before": 0.0, + "before_drop": null, + "best_rot_err_first_goal": null, + "frac_changed": 0.0644855381697487, + "n_changed": 408, + "rollouts_used": 5, + "minutes": 49.8, + "output_tokens": 278731, + "audit": "compliant", + "audit_compliant": true, + "doc": null + }, + { + "dir": "data/robot_policy/attempts/hands/allegro_L_repair_early_ep200/fable-5-1/A_T2_K5_cap10/a0", + "status": "scored", + "counted": true, + "task": "allegro", + "cond": "repair", + "case": "early_ep200", + "kind": "natural", + "model": "fable-5-1", + "rules": "A", + "tools": "T2", + "blind": false, + "metric": "mean_goals", + "after": 0.13818359375, + "after_drop": 0.01806640625, + "before": 0.233642578125, + "before_drop": 0.019775390625, + "best_rot_err_first_goal": 1.3360732793807983, + "frac_changed": 0.025840436646259898, + "n_changed": 21172, + "rollouts_used": 5, + "minutes": 53.3, + "output_tokens": 180800, + "audit": "compliant", + "audit_compliant": true, + "doc": "docs/markdowns/allegro/early_ep200__fable-5-1__A_T2.md" + }, + { + "dir": "data/robot_policy/attempts/hands/allegro_L_repair_early_ep200/opus-5-5/A_T2_K5_cap10/a0", + "status": "scored", + "counted": true, + "task": "allegro", + "cond": "repair", + "case": "early_ep200", + "kind": "natural", + "model": "opus-5-5", + "rules": "A", + "tools": "T2", + "blind": false, + "metric": "mean_goals", + "after": 0.27490234375, + "after_drop": 0.041259765625, + "before": 0.233642578125, + "before_drop": 0.019775390625, + "best_rot_err_first_goal": 1.1451492309570312, + "frac_changed": 0.017954782897370555, + "n_changed": 14711, + "rollouts_used": 5, + "minutes": 29.1, + "output_tokens": 97578, + "audit": "compliant", + "audit_compliant": true, + "doc": "docs/markdowns/allegro/early_ep200__opus-5-5__A_T2.md" + }, + { + "dir": "data/robot_policy/attempts/hands/allegro_L_repair_early_ep50/fable-5-1/A_T2_K5_cap10/a0", + "status": "scored", + "counted": true, + "task": "allegro", + "cond": "repair", + "case": "early_ep50", + "kind": "natural", + "model": "fable-5-1", + "rules": "A", + "tools": "T2", + "blind": false, + "metric": "mean_goals", + "after": 0.01708984375, + "after_drop": 0.10595703125, + "before": 0.027587890625, + "before_drop": 0.01123046875, + "best_rot_err_first_goal": 1.784785270690918, + "frac_changed": 0.03218948026206587, + "n_changed": 26374, + "rollouts_used": 5, + "minutes": 59.1, + "output_tokens": 206462, + "audit": "compliant", + "audit_compliant": true, + "doc": "docs/markdowns/allegro/early_ep50__fable-5-1__A_T2.md" + }, + { + "dir": "data/robot_policy/attempts/hands/allegro_L_repair_early_ep50/opus-5-5/A_T2_K5_cap10/a0", + "status": "scored", + "counted": true, + "task": "allegro", + "cond": "repair", + "case": "early_ep50", + "kind": "natural", + "model": "opus-5-5", + "rules": "A", + "tools": "T2", + "blind": false, + "metric": "mean_goals", + "after": 0.04345703125, + "after_drop": 0.3623046875, + "before": 0.027587890625, + "before_drop": 0.01123046875, + "best_rot_err_first_goal": 1.5322929620742798, + "frac_changed": 0.010486540320454612, + "n_changed": 8592, + "rollouts_used": 5, + "minutes": 28.7, + "output_tokens": 138587, + "audit": "compliant", + "audit_compliant": true, + "doc": "docs/markdowns/allegro/early_ep50__opus-5-5__A_T2.md" + }, + { + "dir": "data/robot_policy/attempts/hands/allegro_L_repair_norm_shift_53/fable-5-1/A_T2_K5_cap10/a0", + "status": "scored", + "counted": true, + "task": "allegro", + "cond": "repair", + "case": "norm_shift_53", + "kind": "damaged", + "model": "fable-5-1", + "rules": "A", + "tools": "T2", + "blind": false, + "metric": "mean_goals", + "after": 12.4794921875, + "after_drop": 0.120849609375, + "before": 0.05322265625, + "before_drop": 0.29345703125, + "best_rot_err_first_goal": 0.006840783171355724, + "frac_changed": 1.2205005028462072e-06, + "n_changed": 1, + "rollouts_used": 1, + "minutes": 4.0, + "output_tokens": 8310, + "audit": "compliant", + "audit_compliant": true, + "doc": "docs/markdowns/allegro/norm_shift_53__fable-5-1__A_T2.md" + }, + { + "dir": "data/robot_policy/attempts/hands/allegro_L_repair_norm_shift_53/fable-5-1/A_T2_K5_cap10/blind0", + "status": "scored", + "counted": true, + "task": "allegro", + "cond": "repair", + "case": "norm_shift_53", + "kind": "damaged", + "model": "fable-5-1", + "rules": "A", + "tools": "T2", + "blind": true, + "metric": "mean_goals", + "after": 12.4794921875, + "after_drop": 0.120849609375, + "before": 0.05322265625, + "before_drop": 0.29345703125, + "best_rot_err_first_goal": 0.006840783171355724, + "frac_changed": 1.2205005028462072e-06, + "n_changed": 1, + "rollouts_used": 1, + "minutes": 5.4, + "output_tokens": 17403, + "audit": "compliant", + "audit_compliant": true, + "doc": null + }, + { + "dir": "data/robot_policy/attempts/hands/allegro_L_repair_swap_act_4_7/opus-5-5/A_T2_K5_cap10/a0", + "status": "scored", + "counted": true, + "task": "allegro", + "cond": "repair", + "case": "swap_act_4_7", + "kind": "damaged", + "model": "opus-5-5", + "rules": "A", + "tools": "T2", + "blind": false, + "metric": "mean_goals", + "after": 12.552001953125, + "after_drop": 0.117431640625, + "before": 0.18310546875, + "before_drop": 0.21142578125, + "best_rot_err_first_goal": 0.010117166675627232, + "frac_changed": 0.00031488912973432143, + "n_changed": 258, + "rollouts_used": 1, + "minutes": 5.6, + "output_tokens": 8534, + "audit": "compliant", + "audit_compliant": true, + "doc": "docs/markdowns/allegro/swap_act_4_7__opus-5-5__A_T2.md" + }, + { + "dir": "data/robot_policy/attempts/hands/allegro_L_repair_swap_act_4_7/opus-5-5/A_T2_K5_cap10/blind0", + "status": "rate_limited", + "counted": false + }, + { + "dir": "data/robot_policy/attempts/hands/shadow_L_repair_early_ep100/fable-5-1/A_T2_K5_cap10/a0", + "status": "scored", + "counted": true, + "task": "shadow", + "cond": "repair", + "case": "early_ep100", + "kind": "natural", + "model": "fable-5-1", + "rules": "A", + "tools": "T2", + "blind": false, + "metric": "mean_goals", + "after": 0.0224609375, + "after_drop": 0.087646484375, + "before": 0.016357421875, + "before_drop": 0.00537109375, + "best_rot_err_first_goal": 1.6580688953399658, + "frac_changed": 0.06303875301075058, + "n_changed": 32192, + "rollouts_used": 5, + "minutes": 90.1, + "output_tokens": 366553, + "audit": "compliant", + "audit_compliant": true, + "doc": "docs/markdowns/shadow/early_ep100__fable-5-1__A_T2.md" + }, + { + "dir": "data/robot_policy/attempts/hands/shadow_L_repair_early_ep100/opus-5-5/A_T2_K5_cap10/a0", + "status": "scored", + "counted": true, + "task": "shadow", + "cond": "repair", + "case": "early_ep100", + "kind": "natural", + "model": "opus-5-5", + "rules": "A", + "tools": "T2", + "blind": false, + "metric": "mean_goals", + "after": 0.04638671875, + "after_drop": 0.082275390625, + "before": 0.016357421875, + "before_drop": 0.00537109375, + "best_rot_err_first_goal": 1.4632103443145752, + "frac_changed": 0.02687645641999726, + "n_changed": 13725, + "rollouts_used": 5, + "minutes": 32.4, + "output_tokens": 160522, + "audit": "compliant", + "audit_compliant": true, + "doc": "docs/markdowns/shadow/early_ep100__opus-5-5__A_T2.md" + }, + { + "dir": "data/robot_policy/attempts/hands/shadow_L_repair_early_ep200/fable-5-1/A_T2_K5_cap10/a0", + "status": "scored", + "counted": true, + "task": "shadow", + "cond": "repair", + "case": "early_ep200", + "kind": "natural", + "model": "fable-5-1", + "rules": "A", + "tools": "T2", + "blind": false, + "metric": "mean_goals", + "after": 1.597900390625, + "after_drop": 0.06884765625, + "before": 0.858154296875, + "before_drop": 0.048828125, + "best_rot_err_first_goal": 0.22931504249572754, + "frac_changed": 0.024520727671490394, + "n_changed": 12522, + "rollouts_used": 5, + "minutes": 36.1, + "output_tokens": 117088, + "audit": "compliant", + "audit_compliant": true, + "doc": "docs/markdowns/shadow/early_ep200__fable-5-1__A_T2.md" + }, + { + "dir": "data/robot_policy/attempts/hands/shadow_L_repair_early_ep200/opus-5-5/A_T2_K5_cap10/a0", + "status": "scored", + "counted": true, + "task": "shadow", + "cond": "repair", + "case": "early_ep200", + "kind": "natural", + "model": "opus-5-5", + "rules": "A", + "tools": "T2", + "blind": false, + "metric": "mean_goals", + "after": 2.6728515625, + "after_drop": 0.25341796875, + "before": 0.858154296875, + "before_drop": 0.048828125, + "best_rot_err_first_goal": 0.11053228378295898, + "frac_changed": 0.049368476707071104, + "n_changed": 25211, + "rollouts_used": 5, + "minutes": 22.5, + "output_tokens": 79142, + "audit": "compliant", + "audit_compliant": true, + "doc": "docs/markdowns/shadow/early_ep200__opus-5-5__A_T2.md" + }, + { + "dir": "data/robot_policy/attempts/hands/shadow_L_repair_norm_shift_69/fable-5-1/A_T2_K5_cap10/a0", + "status": "scored", + "counted": true, + "task": "shadow", + "cond": "repair", + "case": "norm_shift_69", + "kind": "damaged", + "model": "fable-5-1", + "rules": "A", + "tools": "T2", + "blind": false, + "metric": "mean_goals", + "after": 30.692138671875, + "after_drop": 0.068603515625, + "before": 0.00732421875, + "before_drop": 0.501953125, + "best_rot_err_first_goal": 0.004023965913802385, + "frac_changed": 1.9582117610198366e-06, + "n_changed": 1, + "rollouts_used": 1, + "minutes": 5.0, + "output_tokens": 9987, + "audit": "compliant", + "audit_compliant": true, + "doc": "docs/markdowns/shadow/norm_shift_69__fable-5-1__A_T2.md" + }, + { + "dir": "data/robot_policy/attempts/hands/shadow_L_repair_norm_shift_69/fable-5-1/A_T2_K5_cap10/blind0", + "status": "scored", + "counted": true, + "task": "shadow", + "cond": "repair", + "case": "norm_shift_69", + "kind": "damaged", + "model": "fable-5-1", + "rules": "A", + "tools": "T2", + "blind": true, + "metric": "mean_goals", + "after": 30.692138671875, + "after_drop": 0.068603515625, + "before": 0.00732421875, + "before_drop": 0.501953125, + "best_rot_err_first_goal": 0.004023965913802385, + "frac_changed": 1.9582117610198366e-06, + "n_changed": 1, + "rollouts_used": 1, + "minutes": 6.9, + "output_tokens": 21321, + "audit": "compliant", + "audit_compliant": true, + "doc": null + }, + { + "dir": "data/robot_policy/attempts/hands/shadow_L_repair_swap_act_7_11/opus-5-5/A_T2_K5_cap10/a0", + "status": "scored", + "counted": true, + "task": "shadow", + "cond": "repair", + "case": "swap_act_7_11", + "kind": "damaged", + "model": "opus-5-5", + "rules": "A", + "tools": "T2", + "blind": false, + "metric": "mean_goals", + "after": 30.94482421875, + "after_drop": 0.066650390625, + "before": 0.005615234375, + "before_drop": 0.58056640625, + "best_rot_err_first_goal": 0.0032312783878296614, + "frac_changed": 0.0005052186343431178, + "n_changed": 258, + "rollouts_used": 1, + "minutes": 4.2, + "output_tokens": 8178, + "audit": "compliant", + "audit_compliant": true, + "doc": "docs/markdowns/shadow/swap_act_7_11__opus-5-5__A_T2.md" + }, + { + "dir": "data/robot_policy/attempts/hands/shadow_L_repair_swap_act_7_11/opus-5-5/A_T2_K5_cap10/blind0", + "status": "rate_limited", + "counted": false + }, + { + "dir": "data/robot_policy/attempts/hands/allegro_XS_write/fable-5-1/A_T2_K5/a0", + "status": "scored", + "counted": true, + "task": "allegro", + "cond": "write", + "case": null, + "kind": null, + "model": "fable-5-1", + "rules": "A", + "tools": "T2", + "blind": false, + "metric": "mean_goals", + "after": 0.02001953125, + "after_drop": 0.381103515625, + "before": 0.0, + "before_drop": null, + "best_rot_err_first_goal": 1.6483254432678223, + "frac_changed": 0.015764425936942297, + "n_changed": 212, + "rollouts_used": 5, + "minutes": 52.8, + "output_tokens": 197469, + "audit": "compliant", + "audit_compliant": true, + "doc": null + }, + { + "dir": "data/robot_policy/attempts/hands/allegro_XS_write/fable-5-1/B_T2_K5/a0", + "status": "scored", + "counted": true, + "task": "allegro", + "cond": "write", + "case": null, + "kind": null, + "model": "fable-5-1", + "rules": "B", + "tools": "T2", + "blind": false, + "metric": "mean_goals", + "after": 0.02685546875, + "after_drop": 0.15576171875, + "before": 0.0, + "before_drop": null, + "best_rot_err_first_goal": 1.6658480167388916, + "frac_changed": 0.011228435455086259, + "n_changed": 151, + "rollouts_used": 5, + "minutes": 53.0, + "output_tokens": 197447, + "audit": "compliant", + "audit_compliant": true, + "doc": null + }, + { + "dir": "data/robot_policy/attempts/hands/allegro_XS_write/opus-5-5/A_T2_K5/a0", + "status": "scored", + "counted": true, + "task": "allegro", + "cond": "write", + "case": null, + "kind": null, + "model": "opus-5-5", + "rules": "A", + "tools": "T2", + "blind": false, + "metric": "mean_goals", + "after": 0.05224609375, + "after_drop": 0.580322265625, + "before": 0.0, + "before_drop": null, + "best_rot_err_first_goal": 1.3383140563964844, + "frac_changed": 0.04327781082688876, + "n_changed": 582, + "rollouts_used": 5, + "minutes": 38.7, + "output_tokens": 197618, + "audit": "compliant", + "audit_compliant": true, + "doc": null + }, + { + "dir": "data/robot_policy/attempts/hands/allegro_XS_write/opus-5-5/B_T2_K5/a0", + "status": "scored", + "counted": true, + "task": "allegro", + "cond": "write", + "case": null, + "kind": null, + "model": "opus-5-5", + "rules": "B", + "tools": "T2", + "blind": false, + "metric": "mean_goals", + "after": 0.03466796875, + "after_drop": 0.229736328125, + "before": 0.0, + "before_drop": null, + "best_rot_err_first_goal": 1.7037235498428345, + "frac_changed": 0.02647233789411065, + "n_changed": 356, + "rollouts_used": 5, + "minutes": 48.8, + "output_tokens": 252424, + "audit": "compliant", + "audit_compliant": true, + "doc": null + }, + { + "dir": "data/robot_policy/attempts/hands/shadow_XS_write/fable-5-1/A_T2_K5/a0", + "status": "scored", + "counted": true, + "task": "shadow", + "cond": "write", + "case": null, + "kind": null, + "model": "fable-5-1", + "rules": "A", + "tools": "T2", + "blind": false, + "metric": "mean_goals", + "after": 0.01318359375, + "after_drop": 0.06591796875, + "before": 0.0, + "before_drop": null, + "best_rot_err_first_goal": 1.6238707304000854, + "frac_changed": 0.0063577993201561125, + "n_changed": 101, + "rollouts_used": 5, + "minutes": 49.2, + "output_tokens": 179497, + "audit": "compliant", + "audit_compliant": true, + "doc": null + }, + { + "dir": "data/robot_policy/attempts/hands/shadow_XS_write/fable-5-1/B_T2_K5/a0", + "status": "scored", + "counted": true, + "task": "shadow", + "cond": "write", + "case": null, + "kind": null, + "model": "fable-5-1", + "rules": "B", + "tools": "T2", + "blind": false, + "metric": "mean_goals", + "after": 0.01318359375, + "after_drop": 0.17431640625, + "before": 0.0, + "before_drop": null, + "best_rot_err_first_goal": 1.919116735458374, + "frac_changed": 0.008938688153090771, + "n_changed": 142, + "rollouts_used": 5, + "minutes": 44.3, + "output_tokens": 159526, + "audit": "compliant", + "audit_compliant": true, + "doc": null + }, + { + "dir": "data/robot_policy/attempts/hands/shadow_XS_write/opus-5-5/A_T2_K5/a0", + "status": "scored", + "counted": true, + "task": "shadow", + "cond": "write", + "case": null, + "kind": null, + "model": "opus-5-5", + "rules": "A", + "tools": "T2", + "blind": false, + "metric": "mean_goals", + "after": 0.02294921875, + "after_drop": 0.11083984375, + "before": 0.0, + "before_drop": null, + "best_rot_err_first_goal": 1.8377445936203003, + "frac_changed": 0.02549414578874481, + "n_changed": 405, + "rollouts_used": 5, + "minutes": 40.8, + "output_tokens": 207551, + "audit": "compliant", + "audit_compliant": true, + "doc": null + }, + { + "dir": "data/robot_policy/attempts/hands/shadow_XS_write/opus-5-5/B_T2_K5/a0", + "status": "scored", + "counted": true, + "task": "shadow", + "cond": "write", + "case": null, + "kind": null, + "model": "opus-5-5", + "rules": "B", + "tools": "T2", + "blind": false, + "metric": "mean_goals", + "after": 0.022216796875, + "after_drop": 0.006591796875, + "before": 0.0, + "before_drop": null, + "best_rot_err_first_goal": 1.8993313312530518, + "frac_changed": 0.024990557723781948, + "n_changed": 397, + "rollouts_used": 1, + "minutes": 117.4, + "output_tokens": 196926, + "audit": "compliant", + "audit_compliant": true, + "doc": null + } + ], + "experts": { + "franka": { + "success_rate": 0.999755859375, + "policy": "policies/franka/ppo/franka_M_s0/ep01500.npz" + }, + "allegro": { + "mean_goals": 12.552001953125, + "drop_rate": 0.117431640625 + }, + "shadow": { + "mean_goals": 30.94482421875, + "drop_rate": 0.066650390625 + } + }, + "ppo_early": { + "franka": [ + { + "frames": 983040, + "value": 0.0068359375, + "ckpt": "ep00015.npz" + }, + { + "frames": 1310720, + "value": 0.1572265625, + "ckpt": "ep00020.npz" + }, + { + "frames": 1638400, + "value": 0.650634765625, + "ckpt": "ep00025.npz" + }, + { + "frames": 1966080, + "value": 0.50830078125, + "ckpt": "ep00030.npz" + }, + { + "frames": 2293760, + "value": 0.679443359375, + "ckpt": "ep00035.npz" + }, + { + "frames": 2621440, + "value": 0.22265625, + "ckpt": "ep00040.npz" + }, + { + "frames": 2949120, + "value": 0.626220703125, + "ckpt": "ep00045.npz" + }, + { + "frames": 3276800, + "value": 0.88330078125, + "ckpt": "ep00050.npz" + }, + { + "frames": 3604480, + "value": 0.81201171875, + "ckpt": "ep00055.npz" + }, + { + "frames": 3932160, + "value": 0.947021484375, + "ckpt": "ep00060.npz" + } + ], + "allegro": [ + { + "frames": 1310720, + "value": 0.009765625, + "ckpt": "ep00010.npz" + }, + { + "frames": 2621440, + "value": 0.01416015625, + "ckpt": "ep00020.npz" + }, + { + "frames": 3932160, + "value": 0.017578125, + "ckpt": "ep00030.npz" + }, + { + "frames": 5242880, + "value": 0.0234375, + "ckpt": "ep00040.npz" + }, + { + "frames": 6553600, + "value": 0.0263671875, + "ckpt": "ep00050.npz" + }, + { + "frames": 7864320, + "value": 0.03173828125, + "ckpt": "ep00060.npz" + }, + { + "frames": 9175040, + "value": 0.033203125, + "ckpt": "ep00070.npz" + }, + { + "frames": 10485760, + "value": 0.04052734375, + "ckpt": "ep00080.npz" + }, + { + "frames": 11796480, + "value": 0.05029296875, + "ckpt": "ep00090.npz" + }, + { + "frames": 13107200, + "value": 0.05078125, + "ckpt": "ep00100.npz" + }, + { + "frames": 14417920, + "value": 0.048828125, + "ckpt": "ep00110.npz" + }, + { + "frames": 15728640, + "value": 0.060546875, + "ckpt": "ep00120.npz" + }, + { + "frames": 17039360, + "value": 0.07275390625, + "ckpt": "ep00130.npz" + }, + { + "frames": 18350080, + "value": 0.080078125, + "ckpt": "ep00140.npz" + }, + { + "frames": 19660800, + "value": 0.10302734375, + "ckpt": "ep00150.npz" + }, + { + "frames": 20971520, + "value": 0.09765625, + "ckpt": "ep00160.npz" + }, + { + "frames": 22282240, + "value": 0.1298828125, + "ckpt": "ep00170.npz" + }, + { + "frames": 23592960, + "value": 0.15283203125, + "ckpt": "ep00180.npz" + }, + { + "frames": 24903680, + "value": 0.18896484375, + "ckpt": "ep00190.npz" + }, + { + "frames": 26214400, + "value": 0.2529296875, + "ckpt": "ep00200.npz" + } + ], + "shadow": [ + { + "frames": 1310720, + "value": 0.0009765625, + "ckpt": "ep00010.npz" + }, + { + "frames": 2621440, + "value": 0.0029296875, + "ckpt": "ep00020.npz" + }, + { + "frames": 3932160, + "value": 0.0048828125, + "ckpt": "ep00030.npz" + }, + { + "frames": 5242880, + "value": 0.00244140625, + "ckpt": "ep00040.npz" + }, + { + "frames": 6553600, + "value": 0.00341796875, + "ckpt": "ep00050.npz" + }, + { + "frames": 7864320, + "value": 0.0029296875, + "ckpt": "ep00060.npz" + }, + { + "frames": 9175040, + "value": 0.0029296875, + "ckpt": "ep00070.npz" + }, + { + "frames": 10485760, + "value": 0.00732421875, + "ckpt": "ep00080.npz" + }, + { + "frames": 11796480, + "value": 0.01318359375, + "ckpt": "ep00090.npz" + }, + { + "frames": 13107200, + "value": 0.01708984375, + "ckpt": "ep00100.npz" + }, + { + "frames": 14417920, + "value": 0.02197265625, + "ckpt": "ep00110.npz" + }, + { + "frames": 15728640, + "value": 0.0283203125, + "ckpt": "ep00120.npz" + }, + { + "frames": 17039360, + "value": 0.04345703125, + "ckpt": "ep00130.npz" + }, + { + "frames": 18350080, + "value": 0.052734375, + "ckpt": "ep00140.npz" + }, + { + "frames": 19660800, + "value": 0.08984375, + "ckpt": "ep00150.npz" + }, + { + "frames": 20971520, + "value": 0.185546875, + "ckpt": "ep00160.npz" + }, + { + "frames": 22282240, + "value": 0.251953125, + "ckpt": "ep00170.npz" + }, + { + "frames": 23592960, + "value": 0.3671875, + "ckpt": "ep00180.npz" + }, + { + "frames": 24903680, + "value": 0.6474609375, + "ckpt": "ep00190.npz" + }, + { + "frames": 26214400, + "value": 0.87109375, + "ckpt": "ep00200.npz" + } + ] + }, + "blind_jobs": [ + { + "dir": "data/robot_policy/attempts/main/franka_M_repair_swap_act_1_3/opus-5-5/A_T0_K0_cap10/blind0", + "task": "franka", + "case": "swap_act_1_3", + "model": "opus-5-5", + "rules": "A", + "tools": "T0", + "status": "scored", + "quota_failures": 0 + }, + { + "dir": "data/robot_policy/attempts/main/franka_M_repair_swap_act_1_3/opus-5-5/A_T2_K5_cap10/blind0", + "task": "franka", + "case": "swap_act_1_3", + "model": "opus-5-5", + "rules": "A", + "tools": "T2", + "status": "scored", + "quota_failures": 0 + }, + { + "dir": "data/robot_policy/attempts/main/franka_M_repair_swap_act_1_3/opus-5-5/B_T0_K0_cap10/blind0", + "task": "franka", + "case": "swap_act_1_3", + "model": "opus-5-5", + "rules": "B", + "tools": "T0", + "status": "scored", + "quota_failures": 0 + }, + { + "dir": "data/robot_policy/attempts/main/franka_M_repair_norm_shift_21/opus-5-5/A_T0_K0_cap10/blind0", + "task": "franka", + "case": "norm_shift_21", + "model": "opus-5-5", + "rules": "A", + "tools": "T0", + "status": "rate_limited", + "quota_failures": 2 + }, + { + "dir": "data/robot_policy/attempts/main/franka_M_repair_norm_shift_21/opus-5-5/A_T2_K5_cap10/blind0", + "task": "franka", + "case": "norm_shift_21", + "model": "opus-5-5", + "rules": "A", + "tools": "T2", + "status": "rate_limited", + "quota_failures": 2 + }, + { + "dir": "data/robot_policy/attempts/main/franka_M_repair_norm_shift_21/opus-5-5/B_T0_K0_cap10/blind0", + "task": "franka", + "case": "norm_shift_21", + "model": "opus-5-5", + "rules": "B", + "tools": "T0", + "status": "scored", + "quota_failures": 0 + }, + { + "dir": "data/robot_policy/attempts/hands/allegro_L_repair_swap_act_4_7/opus-5-5/A_T2_K5_cap10/blind0", + "task": "allegro", + "case": "swap_act_4_7", + "model": "opus-5-5", + "rules": "A", + "tools": "T2", + "status": "rate_limited", + "quota_failures": 2 + }, + { + "dir": "data/robot_policy/attempts/hands/allegro_L_repair_norm_shift_53/fable-5-1/A_T2_K5_cap10/blind0", + "task": "allegro", + "case": "norm_shift_53", + "model": "fable-5-1", + "rules": "A", + "tools": "T2", + "status": "scored", + "quota_failures": 0 + }, + { + "dir": "data/robot_policy/attempts/hands/shadow_L_repair_swap_act_7_11/opus-5-5/A_T2_K5_cap10/blind0", + "task": "shadow", + "case": "swap_act_7_11", + "model": "opus-5-5", + "rules": "A", + "tools": "T2", + "status": "rate_limited", + "quota_failures": 2 + }, + { + "dir": "data/robot_policy/attempts/hands/shadow_L_repair_norm_shift_69/fable-5-1/A_T2_K5_cap10/blind0", + "task": "shadow", + "case": "norm_shift_69", + "model": "fable-5-1", + "rules": "A", + "tools": "T2", + "status": "scored", + "quota_failures": 0 + } + ] +}