Add trajectory-owned Stage3/Stage4 artifacts
Browse filesCreate-only foundation for the ten-phone Stage1-4 trajectory layout.
This view is limited to 50 files because it contains too many changes. See raw diff
- .gitattributes +5 -0
- objects/phone/trajectories/phone_02/stage4/data039-dr-rollouts-v1/checksums.json +8 -0
- objects/phone/trajectories/phone_02/stage4/data039-dr-rollouts-v1/data_card.md +96 -0
- objects/phone/trajectories/phone_02/stage4/data039-dr-rollouts-v1/dataset_manifest.json +0 -0
- objects/phone/trajectories/phone_02/stage4/data039-dr-rollouts-v1/distribution_calibration.json +120 -0
- objects/phone/trajectories/phone_02/stage4/data039-dr-rollouts-v1/dr_bank.json +3 -0
- objects/phone/trajectories/phone_03/stage4/data040-dr-rollouts-v1/checksums.json +8 -0
- objects/phone/trajectories/phone_03/stage4/data040-dr-rollouts-v1/data_card.md +96 -0
- objects/phone/trajectories/phone_03/stage4/data040-dr-rollouts-v1/dataset_manifest.json +0 -0
- objects/phone/trajectories/phone_03/stage4/data040-dr-rollouts-v1/distribution_calibration.json +120 -0
- objects/phone/trajectories/phone_03/stage4/data040-dr-rollouts-v1/dr_bank.json +3 -0
- objects/phone/trajectories/phone_04/stage3/selected-teacher-v1/checkpoint_receipt.json +17 -0
- objects/phone/trajectories/phone_04/stage3/selected-teacher-v1/model_7000.pt +3 -0
- objects/phone/trajectories/phone_04/stage3/selected-teacher-v1/training_card.md +108 -0
- objects/phone/trajectories/phone_04/stage4/data041-dr-rollouts-v1/checksums.json +8 -0
- objects/phone/trajectories/phone_04/stage4/data041-dr-rollouts-v1/data_card.md +96 -0
- objects/phone/trajectories/phone_04/stage4/data041-dr-rollouts-v1/dataset_manifest.json +0 -0
- objects/phone/trajectories/phone_04/stage4/data041-dr-rollouts-v1/distribution_calibration.json +120 -0
- objects/phone/trajectories/phone_04/stage4/data041-dr-rollouts-v1/dr_bank.json +3 -0
- objects/phone/trajectories/phone_05/stage4/data031-dr-rollouts-v1/checksums.json +8 -0
- objects/phone/trajectories/phone_05/stage4/data031-dr-rollouts-v1/data_card.md +139 -0
- objects/phone/trajectories/phone_05/stage4/data031-dr-rollouts-v1/dataset_manifest.json +0 -0
- objects/phone/trajectories/phone_05/stage4/data031-dr-rollouts-v1/rollout_dataset.tar +3 -0
- objects/phone/trajectories/phone_07/stage4/data019-dr-rollouts-v1/checksums.json +8 -0
- objects/phone/trajectories/phone_07/stage4/data019-dr-rollouts-v1/data_card.md +96 -0
- objects/phone/trajectories/phone_07/stage4/data019-dr-rollouts-v1/dataset_manifest.json +0 -0
- objects/phone/trajectories/phone_07/stage4/data019-dr-rollouts-v1/rollout_dataset.tar +3 -0
- objects/phone/trajectories/phone_09/stage3/selected-teacher-v1/checkpoint_receipt.json +17 -0
- objects/phone/trajectories/phone_09/stage3/selected-teacher-v1/model_8500.pt +3 -0
- objects/phone/trajectories/phone_09/stage3/selected-teacher-v1/training_card.md +118 -0
- objects/phone/trajectories/phone_09/stage4/data043-dr-rollouts-v1/checksums.json +8 -0
- objects/phone/trajectories/phone_09/stage4/data043-dr-rollouts-v1/data_card.md +96 -0
- objects/phone/trajectories/phone_09/stage4/data043-dr-rollouts-v1/dataset_manifest.json +0 -0
- objects/phone/trajectories/phone_09/stage4/data043-dr-rollouts-v1/distribution_calibration.json +120 -0
- objects/phone/trajectories/phone_09/stage4/data043-dr-rollouts-v1/dr_bank.json +3 -0
- objects/phone/trajectories/phone_10/stage4/data044-dr-rollouts-v1/checksums.json +8 -0
- objects/phone/trajectories/phone_10/stage4/data044-dr-rollouts-v1/data_card.md +96 -0
- objects/phone/trajectories/phone_10/stage4/data044-dr-rollouts-v1/dataset_manifest.json +0 -0
- objects/phone/trajectories/phone_10/stage4/data044-dr-rollouts-v1/distribution_calibration.json +120 -0
- objects/phone/trajectories/phone_10/stage4/data044-dr-rollouts-v1/dr_bank.json +3 -0
- objects/phone/trajectories/phone_14/stage4/data033-dr-rollouts-v1/checksums.json +8 -0
- objects/phone/trajectories/phone_14/stage4/data033-dr-rollouts-v1/data_card.md +139 -0
- objects/phone/trajectories/phone_14/stage4/data033-dr-rollouts-v1/dataset_manifest.json +0 -0
- objects/phone/trajectories/phone_14/stage4/data033-dr-rollouts-v1/rollout_dataset.tar +3 -0
- objects/phone/trajectories/phone_17/stage4/data034-dr-rollouts-v1/checksums.json +8 -0
- objects/phone/trajectories/phone_17/stage4/data034-dr-rollouts-v1/data_card.md +139 -0
- objects/phone/trajectories/phone_17/stage4/data034-dr-rollouts-v1/dataset_manifest.json +0 -0
- objects/phone/trajectories/phone_17/stage4/data034-dr-rollouts-v1/rollout_dataset.tar +3 -0
- objects/phone/trajectories/phone_18/stage4/data018-dr-rollouts-v1/checksums.json +8 -0
- objects/phone/trajectories/phone_18/stage4/data018-dr-rollouts-v1/data_card.md +154 -0
.gitattributes
CHANGED
|
@@ -60,3 +60,8 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
|
|
| 60 |
*.webm filter=lfs diff=lfs merge=lfs -text
|
| 61 |
objects/phone/releases/phone20-stage12-v1/common/cell_phone.stl filter=lfs diff=lfs merge=lfs -text
|
| 62 |
objects/phone/releases/phone20-stage12-v2/common/cell_phone.stl filter=lfs diff=lfs merge=lfs -text
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 60 |
*.webm filter=lfs diff=lfs merge=lfs -text
|
| 61 |
objects/phone/releases/phone20-stage12-v1/common/cell_phone.stl filter=lfs diff=lfs merge=lfs -text
|
| 62 |
objects/phone/releases/phone20-stage12-v2/common/cell_phone.stl filter=lfs diff=lfs merge=lfs -text
|
| 63 |
+
objects/phone/trajectories/phone_02/stage4/data039-dr-rollouts-v1/dr_bank.json filter=lfs diff=lfs merge=lfs -text
|
| 64 |
+
objects/phone/trajectories/phone_03/stage4/data040-dr-rollouts-v1/dr_bank.json filter=lfs diff=lfs merge=lfs -text
|
| 65 |
+
objects/phone/trajectories/phone_04/stage4/data041-dr-rollouts-v1/dr_bank.json filter=lfs diff=lfs merge=lfs -text
|
| 66 |
+
objects/phone/trajectories/phone_09/stage4/data043-dr-rollouts-v1/dr_bank.json filter=lfs diff=lfs merge=lfs -text
|
| 67 |
+
objects/phone/trajectories/phone_10/stage4/data044-dr-rollouts-v1/dr_bank.json filter=lfs diff=lfs merge=lfs -text
|
objects/phone/trajectories/phone_02/stage4/data039-dr-rollouts-v1/checksums.json
ADDED
|
@@ -0,0 +1,8 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"archive_sha256": "cf4dd6e00c0bf713b7717c2a186b73cfc1ed0a5b25823b8ec13850f48b4163d2",
|
| 3 |
+
"archive_size_bytes": 860088320,
|
| 4 |
+
"data_task_id": "SHARPA-BC-DATA-039",
|
| 5 |
+
"manifest_sha256": "ee4bcc6b2ce2daccc7506781095d68330dfa1286330b1efa28d34ed98b9e198f",
|
| 6 |
+
"phone_id": "phone_02",
|
| 7 |
+
"schema": "sharpa_stage4_release_checksums_v1"
|
| 8 |
+
}
|
objects/phone/trajectories/phone_02/stage4/data039-dr-rollouts-v1/data_card.md
ADDED
|
@@ -0,0 +1,96 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
# Data Card: SHARPA-BC-DATA-039 / 20260927_phone_02_s1_c4096_dr_rsi_v1
|
| 2 |
+
|
| 3 |
+
| Field | Value |
|
| 4 |
+
|---|---|
|
| 5 |
+
| Data Task ID | SHARPA-BC-DATA-039 |
|
| 6 |
+
| Dataset ID | 20260927_phone_02_s1_c4096_dr_rsi_v1 |
|
| 7 |
+
| Status | READY / C4096 / INFRA0 |
|
| 8 |
+
| Purpose | 保存 phone_02 residual Teacher 在冻结 DR+RSI 分布下的 closed-loop rollout，供 Stage4 BC/Student 独立复用。 |
|
| 9 |
+
| Source mode | teacher_rollout；每个attempt独立采样DR与RSI，保存Teacher真实执行后的deployable observation和absolute applied target。 |
|
| 10 |
+
| Source lineage | SHARPA-AC-P02-001 selected model_7499.pt |
|
| 11 |
+
| Episodes / splits | terminal 4096；train 2902；validation 486；heldout 489；reflection 219；infra 0。 |
|
| 12 |
+
| Sampling | 预先冻结的DR14 + uniform RSI bank；失败保留原split身份，禁止按结果补抽。 |
|
| 13 |
+
| Observation | GRU current-frame 71D与MLP history-3 213D deployable views；不含reference、phase、Teacher ID或DR latent。 |
|
| 14 |
+
| Action / label | 31D absolute final-applied target；不是Teacher raw residual。 |
|
| 15 |
+
| Normalization | admitted train only；attempt等权、attempt内transition等权；validation/heldout transform-only；std floor `1e-8`。 |
|
| 16 |
+
| Producer | formal039_recovery_v3 finalizer |
|
| 17 |
+
| Artifact / manifest | archive 860088320 bytes / SHA `cf4dd6e00c0bf713b7717c2a186b73cfc1ed0a5b25823b8ec13850f48b4163d2`；dataset manifest SHA `ee4bcc6b2ce2daccc7506781095d68330dfa1286330b1efa28d34ed98b9e198f`。 |
|
| 18 |
+
| Claim boundary | 仅支持 phone_02、同一Teacher与同一DR+RSI collection distribution的动作蒸馏；不证明跨trajectory、跨物体、P2 robustness或sim2real。 |
|
| 19 |
+
|
| 20 |
+
## Source and Provenance
|
| 21 |
+
|
| 22 |
+
| Artifact | Identity / value | Evidence |
|
| 23 |
+
|---|---|---|
|
| 24 |
+
| Teacher lineage | SHARPA-AC-P02-001 selected model_7499.pt | 冻结selected Teacher与Stage2 identity。 |
|
| 25 |
+
| Dataset manifest | `ee4bcc6b2ce2daccc7506781095d68330dfa1286330b1efa28d34ed98b9e198f` | READY finalizer输出；所有成员逐SHA绑定。 |
|
| 26 |
+
| Release archive | `cf4dd6e00c0bf713b7717c2a186b73cfc1ed0a5b25823b8ec13850f48b4163d2` / 860088320 bytes | create-only `dataset_v1/` tar。 |
|
| 27 |
+
|
| 28 |
+
## Schema
|
| 29 |
+
|
| 30 |
+
| Field | Shape / type | Semantics |
|
| 31 |
+
|---|---|---|
|
| 32 |
+
| observation | `(A,71)`或`(A,213)` float32 | 当前帧GRU view或attempt-local history-3 MLP view。 |
|
| 33 |
+
| target | `(A,31)` float32 | wrist xyz3 + rotation6D6 + 22 absolute finger targets。 |
|
| 34 |
+
| transition_index | `(A,)` int64 | attempt-local连续控制步。 |
|
| 35 |
+
| state_t1_student_frame | `(A,71)` float32 | physics step后的next deployable state，仅用于审计。 |
|
| 36 |
+
|
| 37 |
+
## Temporal Alignment
|
| 38 |
+
|
| 39 |
+
| Item | Contract |
|
| 40 |
+
|---|---|
|
| 41 |
+
| Physics / control | PhysX 120Hz；policy/control 30Hz；decimation4。 |
|
| 42 |
+
| Transition | `obs_t → Teacher → final_applied_target_t → 4 physics substeps → state_t+1`。 |
|
| 43 |
+
| Boundaries | history和future-action window不得跨attempt、split或index gap。 |
|
| 44 |
+
|
| 45 |
+
## Simulator Environment Randomization
|
| 46 |
+
|
| 47 |
+
| Field | Value |
|
| 48 |
+
|---|---|
|
| 49 |
+
| Enabled | true |
|
| 50 |
+
| Profile | reset-only DR14 + reference augmentation + uniform RSI。 |
|
| 51 |
+
| Exact sampling authority | 发布的DR bank及dataset manifest。 |
|
| 52 |
+
|
| 53 |
+
| Category | Parameter | Distribution / range | Sampling / effect |
|
| 54 |
+
|---|---|---|---|
|
| 55 |
+
| Reference augmentation | common X/Y | each `U[-0.05,0.05] m` | 同时作用hand和phone。 |
|
| 56 |
+
| Reference augmentation | common yaw | `U[-30°,30°]` | 同时作用hand和phone。 |
|
| 57 |
+
| Reset | phone XYZ | each `U[-0.006,0.006] m` | reference/RSI选择后应用。 |
|
| 58 |
+
| Reset | phone rotation | signed angle `U[-3°,3°]` | 每次reset独立采样。 |
|
| 59 |
+
| Reset | finger wrap/rub | wrap `U[0,0.06] rad`；rub `U[-0.025,0.025] rad` | grasp-shaped reset。 |
|
| 60 |
+
| RSI | normalized phase | `U[0,1)` | 映射到合法reference start index。 |
|
| 61 |
+
|
| 62 |
+
| Option | State | Meaning |
|
| 63 |
+
|---|---|---|
|
| 64 |
+
| Physics randomization | disabled | mass、COM、inertia、friction和drive固定。 |
|
| 65 |
+
| Observation/noise/lag | disabled | clean observation；zero lag。 |
|
| 66 |
+
| Disturbance/push | disabled | none。 |
|
| 67 |
+
|
| 68 |
+
## Splits, Sampling, and Leakage
|
| 69 |
+
|
| 70 |
+
| Split | Episodes / rows | Purpose |
|
| 71 |
+
|---|---:|---|
|
| 72 |
+
| train | 2902 | 唯一optimization和normalization来源。 |
|
| 73 |
+
| validation | 486 | offline diagnostic；不参与normalization。 |
|
| 74 |
+
| heldout | 489 | same-source unseen-attempt audit。 |
|
| 75 |
+
| reflection | 219 | 失败attempt；不进入监督数据。 |
|
| 76 |
+
| infra | 0 | 任一infra都会阻止READY。 |
|
| 77 |
+
|
| 78 |
+
Approved exception：本批次使用冻结的C4096或DATA-019 PARTIAL2763拓扑，而不是默认96 train + 32 heldout。
|
| 79 |
+
|
| 80 |
+
## Normalization
|
| 81 |
+
|
| 82 |
+
| Field | Contract |
|
| 83 |
+
|---|---|
|
| 84 |
+
| Source | admitted train attempts only。 |
|
| 85 |
+
| Weighting | attempt等权；attempt内transition等权。 |
|
| 86 |
+
| Validation / heldout | train statistics transform-only。 |
|
| 87 |
+
|
| 88 |
+
## Validation and Readiness
|
| 89 |
+
|
| 90 |
+
| Gate | Result | Evidence |
|
| 91 |
+
|---|---|---|
|
| 92 |
+
| Completeness | PASS | terminal 4096；infra 0。 |
|
| 93 |
+
| Split leakage | PASS | 原split身份保留；reflection不进入训练。 |
|
| 94 |
+
| Member identity | PASS | manifest `ee4bcc6b2ce2daccc7506781095d68330dfa1286330b1efa28d34ed98b9e198f`。 |
|
| 95 |
+
| Release identity | PASS | archive `cf4dd6e00c0bf713b7717c2a186b73cfc1ed0a5b25823b8ec13850f48b4163d2`。 |
|
| 96 |
+
| Dataset status | READY | READY / C4096 / INFRA0。 |
|
objects/phone/trajectories/phone_02/stage4/data039-dr-rollouts-v1/dataset_manifest.json
ADDED
|
The diff for this file is too large to render.
See raw diff
|
|
|
objects/phone/trajectories/phone_02/stage4/data039-dr-rollouts-v1/distribution_calibration.json
ADDED
|
@@ -0,0 +1,120 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"counts": {
|
| 3 |
+
"all": 4096,
|
| 4 |
+
"heldout": 512,
|
| 5 |
+
"train": 3072,
|
| 6 |
+
"validation": 512
|
| 7 |
+
},
|
| 8 |
+
"dimensions": [
|
| 9 |
+
{
|
| 10 |
+
"index": 0,
|
| 11 |
+
"max": 0.9999203216284513,
|
| 12 |
+
"mean": 0.4999999995343387,
|
| 13 |
+
"min": 9.857304394245148e-05
|
| 14 |
+
},
|
| 15 |
+
{
|
| 16 |
+
"index": 1,
|
| 17 |
+
"max": 0.999914282001555,
|
| 18 |
+
"mean": 0.4999999995343387,
|
| 19 |
+
"min": 0.00011155102401971817
|
| 20 |
+
},
|
| 21 |
+
{
|
| 22 |
+
"index": 2,
|
| 23 |
+
"max": 0.9999724524095654,
|
| 24 |
+
"mean": 0.4999999995343387,
|
| 25 |
+
"min": 0.00016929488629102707
|
| 26 |
+
},
|
| 27 |
+
{
|
| 28 |
+
"index": 3,
|
| 29 |
+
"max": 0.9998204549774528,
|
| 30 |
+
"mean": 0.4999999995343387,
|
| 31 |
+
"min": 0.00017310306429862976
|
| 32 |
+
},
|
| 33 |
+
{
|
| 34 |
+
"index": 4,
|
| 35 |
+
"max": 0.9997727330774069,
|
| 36 |
+
"mean": 0.4999999995343387,
|
| 37 |
+
"min": 7.149484008550644e-05
|
| 38 |
+
},
|
| 39 |
+
{
|
| 40 |
+
"index": 5,
|
| 41 |
+
"max": 0.999931494705379,
|
| 42 |
+
"mean": 0.4999999995343387,
|
| 43 |
+
"min": 0.00010371673852205276
|
| 44 |
+
},
|
| 45 |
+
{
|
| 46 |
+
"index": 6,
|
| 47 |
+
"max": 0.999803340062499,
|
| 48 |
+
"mean": 0.4999999995343387,
|
| 49 |
+
"min": 0.00022033601999282837
|
| 50 |
+
},
|
| 51 |
+
{
|
| 52 |
+
"index": 7,
|
| 53 |
+
"max": 0.9998374646529555,
|
| 54 |
+
"mean": 0.4999999995343387,
|
| 55 |
+
"min": 0.0002317270264029503
|
| 56 |
+
},
|
| 57 |
+
{
|
| 58 |
+
"index": 8,
|
| 59 |
+
"max": 0.9999275067821145,
|
| 60 |
+
"mean": 0.4999999995343387,
|
| 61 |
+
"min": 1.8942169845104218e-05
|
| 62 |
+
},
|
| 63 |
+
{
|
| 64 |
+
"index": 9,
|
| 65 |
+
"max": 0.9998596645891666,
|
| 66 |
+
"mean": 0.4999999995343387,
|
| 67 |
+
"min": 0.00019983481615781784
|
| 68 |
+
},
|
| 69 |
+
{
|
| 70 |
+
"index": 10,
|
| 71 |
+
"max": 0.9997780164703727,
|
| 72 |
+
"mean": 0.4999999995343387,
|
| 73 |
+
"min": 8.625909686088562e-05
|
| 74 |
+
},
|
| 75 |
+
{
|
| 76 |
+
"index": 11,
|
| 77 |
+
"max": 0.9998925756663084,
|
| 78 |
+
"mean": 0.4999999995343387,
|
| 79 |
+
"min": 0.00015900935977697372
|
| 80 |
+
},
|
| 81 |
+
{
|
| 82 |
+
"index": 12,
|
| 83 |
+
"max": 0.9997733179479837,
|
| 84 |
+
"mean": 0.4999999995343387,
|
| 85 |
+
"min": 9.434204548597336e-05
|
| 86 |
+
},
|
| 87 |
+
{
|
| 88 |
+
"index": 13,
|
| 89 |
+
"max": 0.999972146935761,
|
| 90 |
+
"mean": 0.4999999995343387,
|
| 91 |
+
"min": 0.00019931979477405548
|
| 92 |
+
},
|
| 93 |
+
{
|
| 94 |
+
"index": 14,
|
| 95 |
+
"max": 0.9999555619433522,
|
| 96 |
+
"mean": 0.4999999995343387,
|
| 97 |
+
"min": 0.0001388806849718094
|
| 98 |
+
}
|
| 99 |
+
],
|
| 100 |
+
"gates": {
|
| 101 |
+
"G10_dimension_means": "PASS",
|
| 102 |
+
"G11_rsi_full_discrete_support": "PASS",
|
| 103 |
+
"G12_action_state_contract": "PASS",
|
| 104 |
+
"G1_exact_cardinality": "PASS",
|
| 105 |
+
"G2_dimension_15d": "PASS",
|
| 106 |
+
"G3_finite": "PASS",
|
| 107 |
+
"G4_unit_cube": "PASS",
|
| 108 |
+
"G5_unique_case_ids": "PASS",
|
| 109 |
+
"G6_unique_case_seeds": "PASS",
|
| 110 |
+
"G7_unique_dr_seeds": "PASS",
|
| 111 |
+
"G8_split_membership_frozen": "PASS",
|
| 112 |
+
"G9_dimension_extrema": "PASS"
|
| 113 |
+
},
|
| 114 |
+
"rsi_start_phase": {
|
| 115 |
+
"max": 316,
|
| 116 |
+
"min": 0
|
| 117 |
+
},
|
| 118 |
+
"schema": "sharpa_coldclone_formal_distribution_calibration_v1",
|
| 119 |
+
"status": "PASS"
|
| 120 |
+
}
|
objects/phone/trajectories/phone_02/stage4/data039-dr-rollouts-v1/dr_bank.json
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:d48af3f71d901a9751140f68de847707e91b1a510936aebd8168b874b4841325
|
| 3 |
+
size 13927184
|
objects/phone/trajectories/phone_03/stage4/data040-dr-rollouts-v1/checksums.json
ADDED
|
@@ -0,0 +1,8 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"archive_sha256": "8b4eb65c1e77648557e0a5edd9f5e1c49870a3d0b843b343075fa7d7275026b8",
|
| 3 |
+
"archive_size_bytes": 923729920,
|
| 4 |
+
"data_task_id": "SHARPA-BC-DATA-040",
|
| 5 |
+
"manifest_sha256": "d0410b978257983f04599c7d0803fe91f8bd93080dfbb2210607106f932d0d21",
|
| 6 |
+
"phone_id": "phone_03",
|
| 7 |
+
"schema": "sharpa_stage4_release_checksums_v1"
|
| 8 |
+
}
|
objects/phone/trajectories/phone_03/stage4/data040-dr-rollouts-v1/data_card.md
ADDED
|
@@ -0,0 +1,96 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
# Data Card: SHARPA-BC-DATA-040 / 20260927_phone_03_s1_c4096_dr_rsi_v1
|
| 2 |
+
|
| 3 |
+
| Field | Value |
|
| 4 |
+
|---|---|
|
| 5 |
+
| Data Task ID | SHARPA-BC-DATA-040 |
|
| 6 |
+
| Dataset ID | 20260927_phone_03_s1_c4096_dr_rsi_v1 |
|
| 7 |
+
| Status | READY / C4096 / INFRA0 |
|
| 8 |
+
| Purpose | 保存 phone_03 residual Teacher 在冻结 DR+RSI 分布下的 closed-loop rollout，供 Stage4 BC/Student 独立复用。 |
|
| 9 |
+
| Source mode | teacher_rollout；每个attempt独立采样DR与RSI，保存Teacher真实执行后的deployable observation和absolute applied target。 |
|
| 10 |
+
| Source lineage | SHARPA-AC-P03-001 selected model_9500.pt |
|
| 11 |
+
| Episodes / splits | terminal 4096；train 2843；validation 465；heldout 466；reflection 322；infra 0。 |
|
| 12 |
+
| Sampling | 预先冻结的DR14 + uniform RSI bank；失败保留原split身份，禁止按结果补抽。 |
|
| 13 |
+
| Observation | GRU current-frame 71D与MLP history-3 213D deployable views；不含reference、phase、Teacher ID或DR latent。 |
|
| 14 |
+
| Action / label | 31D absolute final-applied target；不是Teacher raw residual。 |
|
| 15 |
+
| Normalization | admitted train only；attempt等权、attempt内transition等权；validation/heldout transform-only；std floor `1e-8`。 |
|
| 16 |
+
| Producer | formal040_recovery_v3 finalizer |
|
| 17 |
+
| Artifact / manifest | archive 923729920 bytes / SHA `8b4eb65c1e77648557e0a5edd9f5e1c49870a3d0b843b343075fa7d7275026b8`；dataset manifest SHA `d0410b978257983f04599c7d0803fe91f8bd93080dfbb2210607106f932d0d21`。 |
|
| 18 |
+
| Claim boundary | 仅支持 phone_03、同一Teacher与同一DR+RSI collection distribution的动作蒸馏；不证明跨trajectory、跨物体、P2 robustness或sim2real。 |
|
| 19 |
+
|
| 20 |
+
## Source and Provenance
|
| 21 |
+
|
| 22 |
+
| Artifact | Identity / value | Evidence |
|
| 23 |
+
|---|---|---|
|
| 24 |
+
| Teacher lineage | SHARPA-AC-P03-001 selected model_9500.pt | 冻结selected Teacher与Stage2 identity。 |
|
| 25 |
+
| Dataset manifest | `d0410b978257983f04599c7d0803fe91f8bd93080dfbb2210607106f932d0d21` | READY finalizer输出；所有成员逐SHA绑定。 |
|
| 26 |
+
| Release archive | `8b4eb65c1e77648557e0a5edd9f5e1c49870a3d0b843b343075fa7d7275026b8` / 923729920 bytes | create-only `dataset_v1/` tar。 |
|
| 27 |
+
|
| 28 |
+
## Schema
|
| 29 |
+
|
| 30 |
+
| Field | Shape / type | Semantics |
|
| 31 |
+
|---|---|---|
|
| 32 |
+
| observation | `(A,71)`或`(A,213)` float32 | 当前帧GRU view或attempt-local history-3 MLP view。 |
|
| 33 |
+
| target | `(A,31)` float32 | wrist xyz3 + rotation6D6 + 22 absolute finger targets。 |
|
| 34 |
+
| transition_index | `(A,)` int64 | attempt-local连续控制步。 |
|
| 35 |
+
| state_t1_student_frame | `(A,71)` float32 | physics step后的next deployable state，仅用于审计。 |
|
| 36 |
+
|
| 37 |
+
## Temporal Alignment
|
| 38 |
+
|
| 39 |
+
| Item | Contract |
|
| 40 |
+
|---|---|
|
| 41 |
+
| Physics / control | PhysX 120Hz；policy/control 30Hz；decimation4。 |
|
| 42 |
+
| Transition | `obs_t → Teacher → final_applied_target_t → 4 physics substeps → state_t+1`。 |
|
| 43 |
+
| Boundaries | history和future-action window不得跨attempt、split或index gap。 |
|
| 44 |
+
|
| 45 |
+
## Simulator Environment Randomization
|
| 46 |
+
|
| 47 |
+
| Field | Value |
|
| 48 |
+
|---|---|
|
| 49 |
+
| Enabled | true |
|
| 50 |
+
| Profile | reset-only DR14 + reference augmentation + uniform RSI。 |
|
| 51 |
+
| Exact sampling authority | 发布的DR bank及dataset manifest。 |
|
| 52 |
+
|
| 53 |
+
| Category | Parameter | Distribution / range | Sampling / effect |
|
| 54 |
+
|---|---|---|---|
|
| 55 |
+
| Reference augmentation | common X/Y | each `U[-0.05,0.05] m` | 同时作用hand和phone。 |
|
| 56 |
+
| Reference augmentation | common yaw | `U[-30°,30°]` | 同时作用hand和phone。 |
|
| 57 |
+
| Reset | phone XYZ | each `U[-0.006,0.006] m` | reference/RSI选择后应用。 |
|
| 58 |
+
| Reset | phone rotation | signed angle `U[-3°,3°]` | 每次reset独立采样。 |
|
| 59 |
+
| Reset | finger wrap/rub | wrap `U[0,0.06] rad`；rub `U[-0.025,0.025] rad` | grasp-shaped reset。 |
|
| 60 |
+
| RSI | normalized phase | `U[0,1)` | 映射到合法reference start index。 |
|
| 61 |
+
|
| 62 |
+
| Option | State | Meaning |
|
| 63 |
+
|---|---|---|
|
| 64 |
+
| Physics randomization | disabled | mass、COM、inertia、friction和drive固定。 |
|
| 65 |
+
| Observation/noise/lag | disabled | clean observation；zero lag。 |
|
| 66 |
+
| Disturbance/push | disabled | none。 |
|
| 67 |
+
|
| 68 |
+
## Splits, Sampling, and Leakage
|
| 69 |
+
|
| 70 |
+
| Split | Episodes / rows | Purpose |
|
| 71 |
+
|---|---:|---|
|
| 72 |
+
| train | 2843 | 唯一optimization和normalization来源。 |
|
| 73 |
+
| validation | 465 | offline diagnostic；不参与normalization。 |
|
| 74 |
+
| heldout | 466 | same-source unseen-attempt audit。 |
|
| 75 |
+
| reflection | 322 | 失败attempt；不进入监督数据。 |
|
| 76 |
+
| infra | 0 | 任一infra都会阻止READY。 |
|
| 77 |
+
|
| 78 |
+
Approved exception：本批次使用冻结的C4096或DATA-019 PARTIAL2763拓扑，而不是默认96 train + 32 heldout。
|
| 79 |
+
|
| 80 |
+
## Normalization
|
| 81 |
+
|
| 82 |
+
| Field | Contract |
|
| 83 |
+
|---|---|
|
| 84 |
+
| Source | admitted train attempts only。 |
|
| 85 |
+
| Weighting | attempt等权；attempt内transition等权。 |
|
| 86 |
+
| Validation / heldout | train statistics transform-only。 |
|
| 87 |
+
|
| 88 |
+
## Validation and Readiness
|
| 89 |
+
|
| 90 |
+
| Gate | Result | Evidence |
|
| 91 |
+
|---|---|---|
|
| 92 |
+
| Completeness | PASS | terminal 4096；infra 0。 |
|
| 93 |
+
| Split leakage | PASS | 原split身份保留；reflection不进入训练。 |
|
| 94 |
+
| Member identity | PASS | manifest `d0410b978257983f04599c7d0803fe91f8bd93080dfbb2210607106f932d0d21`。 |
|
| 95 |
+
| Release identity | PASS | archive `8b4eb65c1e77648557e0a5edd9f5e1c49870a3d0b843b343075fa7d7275026b8`。 |
|
| 96 |
+
| Dataset status | READY | READY / C4096 / INFRA0。 |
|
objects/phone/trajectories/phone_03/stage4/data040-dr-rollouts-v1/dataset_manifest.json
ADDED
|
The diff for this file is too large to render.
See raw diff
|
|
|
objects/phone/trajectories/phone_03/stage4/data040-dr-rollouts-v1/distribution_calibration.json
ADDED
|
@@ -0,0 +1,120 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"counts": {
|
| 3 |
+
"all": 4096,
|
| 4 |
+
"heldout": 512,
|
| 5 |
+
"train": 3072,
|
| 6 |
+
"validation": 512
|
| 7 |
+
},
|
| 8 |
+
"dimensions": [
|
| 9 |
+
{
|
| 10 |
+
"index": 0,
|
| 11 |
+
"max": 0.9998146072030067,
|
| 12 |
+
"mean": 0.4999999995343387,
|
| 13 |
+
"min": 3.0891969799995422e-06
|
| 14 |
+
},
|
| 15 |
+
{
|
| 16 |
+
"index": 1,
|
| 17 |
+
"max": 0.9999624211341143,
|
| 18 |
+
"mean": 0.4999999995343387,
|
| 19 |
+
"min": 0.00018163397908210754
|
| 20 |
+
},
|
| 21 |
+
{
|
| 22 |
+
"index": 2,
|
| 23 |
+
"max": 0.999921235255897,
|
| 24 |
+
"mean": 0.4999999995343387,
|
| 25 |
+
"min": 6.643403321504593e-05
|
| 26 |
+
},
|
| 27 |
+
{
|
| 28 |
+
"index": 3,
|
| 29 |
+
"max": 0.9999680146574974,
|
| 30 |
+
"mean": 0.4999999995343387,
|
| 31 |
+
"min": 0.00014361366629600525
|
| 32 |
+
},
|
| 33 |
+
{
|
| 34 |
+
"index": 4,
|
| 35 |
+
"max": 0.9997642720118165,
|
| 36 |
+
"mean": 0.4999999995343387,
|
| 37 |
+
"min": 0.0002243276685476303
|
| 38 |
+
},
|
| 39 |
+
{
|
| 40 |
+
"index": 5,
|
| 41 |
+
"max": 0.9999650986865163,
|
| 42 |
+
"mean": 0.4999999995343387,
|
| 43 |
+
"min": 0.00010893400758504868
|
| 44 |
+
},
|
| 45 |
+
{
|
| 46 |
+
"index": 6,
|
| 47 |
+
"max": 0.999795438721776,
|
| 48 |
+
"mean": 0.4999999995343387,
|
| 49 |
+
"min": 8.198525756597519e-05
|
| 50 |
+
},
|
| 51 |
+
{
|
| 52 |
+
"index": 7,
|
| 53 |
+
"max": 0.999818286858499,
|
| 54 |
+
"mean": 0.4999999995343387,
|
| 55 |
+
"min": 0.00015585403889417648
|
| 56 |
+
},
|
| 57 |
+
{
|
| 58 |
+
"index": 8,
|
| 59 |
+
"max": 0.9998348895460367,
|
| 60 |
+
"mean": 0.4999999995343387,
|
| 61 |
+
"min": 6.640609353780746e-05
|
| 62 |
+
},
|
| 63 |
+
{
|
| 64 |
+
"index": 9,
|
| 65 |
+
"max": 0.999998296611011,
|
| 66 |
+
"mean": 0.4999999995343387,
|
| 67 |
+
"min": 0.00013250019401311874
|
| 68 |
+
},
|
| 69 |
+
{
|
| 70 |
+
"index": 10,
|
| 71 |
+
"max": 0.9998271018266678,
|
| 72 |
+
"mean": 0.4999999995343387,
|
| 73 |
+
"min": 0.00017285626381635666
|
| 74 |
+
},
|
| 75 |
+
{
|
| 76 |
+
"index": 11,
|
| 77 |
+
"max": 0.9998014755547047,
|
| 78 |
+
"mean": 0.4999999995343387,
|
| 79 |
+
"min": 8.559133857488632e-05
|
| 80 |
+
},
|
| 81 |
+
{
|
| 82 |
+
"index": 12,
|
| 83 |
+
"max": 0.9997946731746197,
|
| 84 |
+
"mean": 0.4999999995343387,
|
| 85 |
+
"min": 2.6851892471313477e-05
|
| 86 |
+
},
|
| 87 |
+
{
|
| 88 |
+
"index": 13,
|
| 89 |
+
"max": 0.9997707791626453,
|
| 90 |
+
"mean": 0.4999999995343387,
|
| 91 |
+
"min": 3.535207360982895e-05
|
| 92 |
+
},
|
| 93 |
+
{
|
| 94 |
+
"index": 14,
|
| 95 |
+
"max": 0.9998546307906508,
|
| 96 |
+
"mean": 0.4999999995343387,
|
| 97 |
+
"min": 0.000133574940264225
|
| 98 |
+
}
|
| 99 |
+
],
|
| 100 |
+
"gates": {
|
| 101 |
+
"G10_dimension_means": "PASS",
|
| 102 |
+
"G11_rsi_full_discrete_support": "PASS",
|
| 103 |
+
"G12_action_state_contract": "PASS",
|
| 104 |
+
"G1_exact_cardinality": "PASS",
|
| 105 |
+
"G2_dimension_15d": "PASS",
|
| 106 |
+
"G3_finite": "PASS",
|
| 107 |
+
"G4_unit_cube": "PASS",
|
| 108 |
+
"G5_unique_case_ids": "PASS",
|
| 109 |
+
"G6_unique_case_seeds": "PASS",
|
| 110 |
+
"G7_unique_dr_seeds": "PASS",
|
| 111 |
+
"G8_split_membership_frozen": "PASS",
|
| 112 |
+
"G9_dimension_extrema": "PASS"
|
| 113 |
+
},
|
| 114 |
+
"rsi_start_phase": {
|
| 115 |
+
"max": 356,
|
| 116 |
+
"min": 0
|
| 117 |
+
},
|
| 118 |
+
"schema": "sharpa_coldclone_formal_distribution_calibration_v1",
|
| 119 |
+
"status": "PASS"
|
| 120 |
+
}
|
objects/phone/trajectories/phone_03/stage4/data040-dr-rollouts-v1/dr_bank.json
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:038211533432767596781f510942feb3e000eee614e78c47b6233e85260ea24f
|
| 3 |
+
size 13927518
|
objects/phone/trajectories/phone_04/stage3/selected-teacher-v1/checkpoint_receipt.json
ADDED
|
@@ -0,0 +1,17 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"checkpoint": {
|
| 3 |
+
"path": "model_7000.pt",
|
| 4 |
+
"sha256": "1872d4527d9fffe29c5ab9c7532ecd0a07f9b205d2720d0939e23e2fc672b74c",
|
| 5 |
+
"size_bytes": 21074741
|
| 6 |
+
},
|
| 7 |
+
"claim_boundary": "source identity only; no new evaluation or promotion claim",
|
| 8 |
+
"phone_id": "phone_04",
|
| 9 |
+
"schema": "sharpa_stage3_selected_teacher_receipt_v1",
|
| 10 |
+
"selection_basis": "frozen source Teacher for SHARPA-BC-DATA-041",
|
| 11 |
+
"task_id": "SHARPA-AC-P04-001",
|
| 12 |
+
"training_card": {
|
| 13 |
+
"path": "training_card.md",
|
| 14 |
+
"sha256": "1cc34146c145e5f60340fb9694f0b0f7acfaa5c5dbb6bfc522b98d80dcc50f66",
|
| 15 |
+
"size_bytes": 7190
|
| 16 |
+
}
|
| 17 |
+
}
|
objects/phone/trajectories/phone_04/stage3/selected-teacher-v1/model_7000.pt
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:1872d4527d9fffe29c5ab9c7532ecd0a07f9b205d2720d0939e23e2fc672b74c
|
| 3 |
+
size 21074741
|
objects/phone/trajectories/phone_04/stage3/selected-teacher-v1/training_card.md
ADDED
|
@@ -0,0 +1,108 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
# Training Card: phone_04_autocurriculum_v1
|
| 2 |
+
|
| 3 |
+
| Field | Value |
|
| 4 |
+
|---|---|
|
| 5 |
+
| Description | [SHARPA-AC-P04-001] accepted phone_04 Stage2 reference的`No-DR plateau → RefAug/reset-only DR plateau → 8-seed 2×3 → W&B artifact/readback`正式teacher候选。 |
|
| 6 |
+
| Role | T0 residual-teacher Auto-Curriculum mentor batch candidate |
|
| 7 |
+
| Policy | Asymmetric MLP residual PPO；actor/critic `[1024,512,256,128]`；ELU；S0 fresh；S1 model-only warm start、fresh optimizer |
|
| 8 |
+
| Policy input | 162D actor；199D asymmetric critic；history total 3 frames |
|
| 9 |
+
| Hidden state | false |
|
| 10 |
+
| Action | 28D normalized residual：6D floating root + 22D Sharpa fingers；no chunk `(1/1)` |
|
| 11 |
+
| Data | accepted phone_04 handoff SHA256 `6bc4e6e6996a784cf1b633bd4079f2c11adda086122b56dc9c6bfe999267b23e`；moving28 H5 SHA256 `400daea4bb1baad031a302371da89f0a87afc35ba7bd1bf7090ba113b87068fd` |
|
| 12 |
+
| Training | S0 `Regrind-Sharpa-Wuji2ActuatorParity-RouteA-v0`后S1 `Regrind-Sharpa-Wuji2ActuatorParity-RefAugDR-RouteA-v1`；各4096 env、24 steps/env、seed1、最多5000 iterations；checkpoint interval 500 |
|
| 13 |
+
| Reward profile | exact T0-153/WUJI2 RouteA reward、termination与PPO；S1只开启冻结RefAug/reset-only DR |
|
| 14 |
+
| Claim | 训练和发布完成只证明单episode remote reproduction；8-seed为diagnostic；P2仍`NOT_RUN`，不得声称robust teacher或真机成功。 |
|
| 15 |
+
| Metric | matching 16-env gate；S0/S1 plateau selected checkpoint；8/8 rollout与2×3；W&B checkpoint/video artifact下载SHA回读 |
|
| 16 |
+
| W&B | entity `dfnet`；project `dex-retarget`；group `sharpa-phone_04-autocurriculum-v1`；run IDs `sharpa-phone_04-autocurriculum-s0-v1`、`sharpa-phone_04-autocurriculum-s1-v1`、`sharpa-phone_04-autocurriculum-evaluation-seed0-7-v1` |
|
| 17 |
+
|
| 18 |
+
## Policy Inputs
|
| 19 |
+
|
| 20 |
+
| Field | Dim | Description |
|
| 21 |
+
|---|---:|---|
|
| 22 |
+
| object_position_world | 3 | 当前手机世界位置 |
|
| 23 |
+
| object_rotation_6d_world | 6 | 当前手机世界朝向连续 6D 表示 |
|
| 24 |
+
| wrist_position_history_3 | 9 | 当前及前两控制步 root 位置 |
|
| 25 |
+
| wrist_rotation_6d_history_3 | 18 | 当前及前两控制步 root 朝向 |
|
| 26 |
+
| joint_position_history_3 | 66 | 当前及前两控制步 22D Sharpa 关节角 |
|
| 27 |
+
| previous_raw_action | 28 | 上一控制步 raw residual action |
|
| 28 |
+
| demonstration_phase | 1 | 当前 reference index 的归一化 phase |
|
| 29 |
+
| action_base_wrist_position_rotation_6d | 9 | 当前 reference root target |
|
| 30 |
+
| action_base_joint_position | 22 | 当前 22D reference joint target |
|
| 31 |
+
|
| 32 |
+
Critic 在 actor 162D 上增加五个语义 fingertip world positions（15D）和 22D joint velocity，共 199D；这些 privileged 量不进入 actor。
|
| 33 |
+
|
| 34 |
+
## Simulator Environment Randomization
|
| 35 |
+
|
| 36 |
+
| Field | Value |
|
| 37 |
+
|---|---|
|
| 38 |
+
| Enabled | true |
|
| 39 |
+
| Source config | `sharpa/configs/sharpa_managerbased_pipeline/dr_contract_refaug_reset_only_v1.json` |
|
| 40 |
+
| Source config SHA256 | `d99efe6b27a179211fec8f3077e5605f9d804c666f0ac9a16c3e8349795982b8` |
|
| 41 |
+
| Runtime readback | 每次gate/formal另写create-only readback；必须17/17 matching |
|
| 42 |
+
| Runtime readback SHA256 | 冻结reference `21fbe1950c87fd029ce03827f204f70190845286ac7dd7d0d61165c2c307a804`；本次执行另写真实SHA |
|
| 43 |
+
| Activation | S0关闭randomization；S1全部4096 env开启；nominal fraction `0.0` |
|
| 44 |
+
| Schedule | S0独立plateau后切S1；S1重新初始化stage-local EMA；reference fade `.135→.292` |
|
| 45 |
+
| Mask | common rigid transform同时作用hand与phone；relative reset保持grasp-shaped范围 |
|
| 46 |
+
| Seed | 两阶段seed1 |
|
| 47 |
+
|
| 48 |
+
| Category | Parameter | Distribution / range | Sampling / effect |
|
| 49 |
+
|---|---|---|---|
|
| 50 |
+
| Reset | S0 reset | fixed；保留uniform RSI | No-DR nominal stage |
|
| 51 |
+
| Reset | S1 phone relative XYZ | each `U[-0.006,0.006] m` | every reset |
|
| 52 |
+
| Reset | S1 phone relative rotation | axis-angle `U[-3°,3°]` | every reset |
|
| 53 |
+
| Reset | S1 finger wrap | `U[0,0.06] rad` | 只增加包握 |
|
| 54 |
+
| Reset | S1 four-finger rub | `U[-0.025,0.025] rad` | 四指沿手机小范围扰动 |
|
| 55 |
+
| Reference augmentation | common XY | each `U[-0.05,0.05] m` | hand与phone共用 |
|
| 56 |
+
| Reference augmentation | common yaw | `U[-30°,30°]` | hand与phone共用 |
|
| 57 |
+
| Reference augmentation | phase fade | full through `.135`；zero at `.292` | 平滑退火回原reference |
|
| 58 |
+
|
| 59 |
+
### Explicitly disabled / not applicable
|
| 60 |
+
|
| 61 |
+
| Option | State | Meaning |
|
| 62 |
+
|---|---|---|
|
| 63 |
+
| Physics randomization | disabled | mass/COM/inertia/friction/drive固定；full gravity `-9.81 m/s²` |
|
| 64 |
+
| Observation/noise/lag | disabled | clean、zero-lag observation |
|
| 65 |
+
| Disturbance/push | disabled | 无external push |
|
| 66 |
+
|
| 67 |
+
## Rewards
|
| 68 |
+
|
| 69 |
+
| Reward | Weight | Description |
|
| 70 |
+
|---|---:|---|
|
| 71 |
+
| object_keypoints_pos_error_exp | 3.0 | 100 个手机表面点跟随 augmented reference |
|
| 72 |
+
| object_lin_vel_error_exp | 1.0 | 手机线速度跟随 augmented reference |
|
| 73 |
+
| object_ang_vel_error_exp | 1.0 | 手机角速度跟随 augmented reference |
|
| 74 |
+
| hand_wrist_pos_error_exp | 0.1 | root 位置软约束 |
|
| 75 |
+
| hand_wrist_rot_error_exp | 0.1 | root 朝向软约束 |
|
| 76 |
+
| action_l2_exp | 0.5 | residual 幅度约束 |
|
| 77 |
+
| action_rate_l2_exp | 1.0 | residual 平滑约束 |
|
| 78 |
+
| action_out_of_bounds_exp | 1.0 | raw action `[-1,1]` 越界惩罚 |
|
| 79 |
+
| early_terminated_penalty | -10.0 | 非 timeout 提前终止惩罚 |
|
| 80 |
+
| fingertip_object_graph_error_exp | 0.75 | 五指尖—手机相对几何跟随 augmented reference |
|
| 81 |
+
|
| 82 |
+
## Stage2 Reference Acceptance
|
| 83 |
+
|
| 84 |
+
| Field | Frozen evidence |
|
| 85 |
+
|---|---|
|
| 86 |
+
| Stage1 source | SHA256 `f80a65b2f8948315a343f2f1f056db8cdb795b03501cbc401995385007298837` |
|
| 87 |
+
| Stage1 manifest | SHA256 `1ef4843f3e0f3c1845925d4b4f58fe4eb8c543798529d694e5ff041a90f1feff` |
|
| 88 |
+
| Runtime human H5 | SHA256 `576fe847b7fb55c1436b12500aad894f101e76d46a7cc8b6f7e3d50fa2dddc26` |
|
| 89 |
+
| Canonical retarget H5 | SHA256 `ba667fef06c182f528b606ce1b3fea4865e8931b88a83b11bf750a8b6a030bed` |
|
| 90 |
+
| Runtime retarget H5 | SHA256 `400daea4bb1baad031a302371da89f0a87afc35ba7bd1bf7090ba113b87068fd` |
|
| 91 |
+
| Reference manifest | SHA256 `2cda6d35199f01b0d3697c8c8b99ffedc8a834849f9fc8e1ba8b84d0756003f1` |
|
| 92 |
+
| Handoff | SHA256 `6bc4e6e6996a784cf1b633bd4079f2c11adda086122b56dc9c6bfe999267b23e`；strict loader `STAGE2_HANDOFF_ACCEPTED` |
|
| 93 |
+
| Timeline | 696 frames @ 60 Hz → 348 policy targets @ 30 Hz |
|
| 94 |
+
| Stage2 QA | validation receipt SHA256 `e3f45f416ebc7593044234d2db012719767cb64a866197a4b6d48ae707de5068`；T0-147-aligned nonpenetration/joint-limit/coupling/continuity gate PASS |
|
| 95 |
+
| Full review | Stage1 MANO / Stage2 Sharpa 2×3 SHA256 `8ec37e98c08f30ec74c5eef159c13df9c158b27cf14139cf2c9fbaaa0cd3a4e6` |
|
| 96 |
+
|
| 97 |
+
## Auto-Curriculum and Evaluation Contract
|
| 98 |
+
|
| 99 |
+
| Field | Frozen value |
|
| 100 |
+
|---|---|
|
| 101 |
+
| Task ID | `SHARPA-AC-P04-001` |
|
| 102 |
+
| Stop rule | 每stage独立EMA `.95`；min `1500`；meaningful improvement `2%`；patience `1000`；max local `4999` |
|
| 103 |
+
| Transition | S0 selected checkpoint model-only warm start S1；fresh optimizer |
|
| 104 |
+
| GoalContract | SHA256 `e9d5af764551161d221ea233ad9cb8179fe5ecdadf7df0ae680a1235ae45f4d2` |
|
| 105 |
+
| Evaluation | selected S1；fresh Isaac；`refaug_reset_dr`；seed `0..7`；完整trajectory + 15-step hold |
|
| 106 |
+
| Videos | 每seed一条Stage1 MANO top row / policy actual bottom row的H.264 2×3 |
|
| 107 |
+
| Publication | S0/S1 selected checkpoint Artifact；evaluation bundle；W&B API下载回读与SHA/ffprobe |
|
| 108 |
+
| P2 / Promotion | `NOT_RUN / NON_PROMOTABLE` |
|
objects/phone/trajectories/phone_04/stage4/data041-dr-rollouts-v1/checksums.json
ADDED
|
@@ -0,0 +1,8 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"archive_sha256": "1fe5895ed77efa2c255a3503e679f0295493e2cb8505398cd6970576e207a599",
|
| 3 |
+
"archive_size_bytes": 983459840,
|
| 4 |
+
"data_task_id": "SHARPA-BC-DATA-041",
|
| 5 |
+
"manifest_sha256": "9f99d35e6bede9c9c6ac58f0a40163294e83d81d3f5467f3220bad9881f7c9cd",
|
| 6 |
+
"phone_id": "phone_04",
|
| 7 |
+
"schema": "sharpa_stage4_release_checksums_v1"
|
| 8 |
+
}
|
objects/phone/trajectories/phone_04/stage4/data041-dr-rollouts-v1/data_card.md
ADDED
|
@@ -0,0 +1,96 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
# Data Card: SHARPA-BC-DATA-041 / 20260927_phone_04_s1_c4096_dr_rsi_v1
|
| 2 |
+
|
| 3 |
+
| Field | Value |
|
| 4 |
+
|---|---|
|
| 5 |
+
| Data Task ID | SHARPA-BC-DATA-041 |
|
| 6 |
+
| Dataset ID | 20260927_phone_04_s1_c4096_dr_rsi_v1 |
|
| 7 |
+
| Status | READY / C4096 / INFRA0 |
|
| 8 |
+
| Purpose | 保存 phone_04 residual Teacher 在冻结 DR+RSI 分布下的 closed-loop rollout，供 Stage4 BC/Student 独立复用。 |
|
| 9 |
+
| Source mode | teacher_rollout；每个attempt独立采样DR与RSI，保存Teacher真实执行后的deployable observation和absolute applied target。 |
|
| 10 |
+
| Source lineage | SHARPA-AC-P04-001 selected model_7000.pt |
|
| 11 |
+
| Episodes / splits | terminal 4096；train 3034；validation 506；heldout 506；reflection 50；infra 0。 |
|
| 12 |
+
| Sampling | 预先冻结的DR14 + uniform RSI bank；失败保留原split身份，禁止按结果补抽。 |
|
| 13 |
+
| Observation | GRU current-frame 71D与MLP history-3 213D deployable views；不含reference、phase、Teacher ID或DR latent。 |
|
| 14 |
+
| Action / label | 31D absolute final-applied target；不是Teacher raw residual。 |
|
| 15 |
+
| Normalization | admitted train only；attempt等权、attempt内transition等权；validation/heldout transform-only；std floor `1e-8`。 |
|
| 16 |
+
| Producer | formal041_recovery_v3 finalizer |
|
| 17 |
+
| Artifact / manifest | archive 983459840 bytes / SHA `1fe5895ed77efa2c255a3503e679f0295493e2cb8505398cd6970576e207a599`；dataset manifest SHA `9f99d35e6bede9c9c6ac58f0a40163294e83d81d3f5467f3220bad9881f7c9cd`。 |
|
| 18 |
+
| Claim boundary | 仅支持 phone_04、同一Teacher与同一DR+RSI collection distribution的动作蒸馏；不证明跨trajectory、跨物体、P2 robustness或sim2real。 |
|
| 19 |
+
|
| 20 |
+
## Source and Provenance
|
| 21 |
+
|
| 22 |
+
| Artifact | Identity / value | Evidence |
|
| 23 |
+
|---|---|---|
|
| 24 |
+
| Teacher lineage | SHARPA-AC-P04-001 selected model_7000.pt | 冻结selected Teacher与Stage2 identity。 |
|
| 25 |
+
| Dataset manifest | `9f99d35e6bede9c9c6ac58f0a40163294e83d81d3f5467f3220bad9881f7c9cd` | READY finalizer输出；所有成员逐SHA绑定。 |
|
| 26 |
+
| Release archive | `1fe5895ed77efa2c255a3503e679f0295493e2cb8505398cd6970576e207a599` / 983459840 bytes | create-only `dataset_v1/` tar。 |
|
| 27 |
+
|
| 28 |
+
## Schema
|
| 29 |
+
|
| 30 |
+
| Field | Shape / type | Semantics |
|
| 31 |
+
|---|---|---|
|
| 32 |
+
| observation | `(A,71)`或`(A,213)` float32 | 当前帧GRU view或attempt-local history-3 MLP view。 |
|
| 33 |
+
| target | `(A,31)` float32 | wrist xyz3 + rotation6D6 + 22 absolute finger targets。 |
|
| 34 |
+
| transition_index | `(A,)` int64 | attempt-local连续控制步。 |
|
| 35 |
+
| state_t1_student_frame | `(A,71)` float32 | physics step后的next deployable state，仅用于审计。 |
|
| 36 |
+
|
| 37 |
+
## Temporal Alignment
|
| 38 |
+
|
| 39 |
+
| Item | Contract |
|
| 40 |
+
|---|---|
|
| 41 |
+
| Physics / control | PhysX 120Hz；policy/control 30Hz；decimation4。 |
|
| 42 |
+
| Transition | `obs_t → Teacher → final_applied_target_t → 4 physics substeps → state_t+1`。 |
|
| 43 |
+
| Boundaries | history和future-action window不得跨attempt、split或index gap。 |
|
| 44 |
+
|
| 45 |
+
## Simulator Environment Randomization
|
| 46 |
+
|
| 47 |
+
| Field | Value |
|
| 48 |
+
|---|---|
|
| 49 |
+
| Enabled | true |
|
| 50 |
+
| Profile | reset-only DR14 + reference augmentation + uniform RSI。 |
|
| 51 |
+
| Exact sampling authority | 发布的DR bank及dataset manifest。 |
|
| 52 |
+
|
| 53 |
+
| Category | Parameter | Distribution / range | Sampling / effect |
|
| 54 |
+
|---|---|---|---|
|
| 55 |
+
| Reference augmentation | common X/Y | each `U[-0.05,0.05] m` | 同时作用hand和phone。 |
|
| 56 |
+
| Reference augmentation | common yaw | `U[-30°,30°]` | 同时作用hand和phone。 |
|
| 57 |
+
| Reset | phone XYZ | each `U[-0.006,0.006] m` | reference/RSI选择后应用。 |
|
| 58 |
+
| Reset | phone rotation | signed angle `U[-3°,3°]` | 每次reset独立采样。 |
|
| 59 |
+
| Reset | finger wrap/rub | wrap `U[0,0.06] rad`；rub `U[-0.025,0.025] rad` | grasp-shaped reset。 |
|
| 60 |
+
| RSI | normalized phase | `U[0,1)` | 映射到合法reference start index。 |
|
| 61 |
+
|
| 62 |
+
| Option | State | Meaning |
|
| 63 |
+
|---|---|---|
|
| 64 |
+
| Physics randomization | disabled | mass、COM、inertia、friction和drive固定。 |
|
| 65 |
+
| Observation/noise/lag | disabled | clean observation；zero lag。 |
|
| 66 |
+
| Disturbance/push | disabled | none。 |
|
| 67 |
+
|
| 68 |
+
## Splits, Sampling, and Leakage
|
| 69 |
+
|
| 70 |
+
| Split | Episodes / rows | Purpose |
|
| 71 |
+
|---|---:|---|
|
| 72 |
+
| train | 3034 | 唯一optimization和normalization来源。 |
|
| 73 |
+
| validation | 506 | offline diagnostic；不参与normalization。 |
|
| 74 |
+
| heldout | 506 | same-source unseen-attempt audit。 |
|
| 75 |
+
| reflection | 50 | 失败attempt；不进入监督数据。 |
|
| 76 |
+
| infra | 0 | 任一infra都会阻止READY。 |
|
| 77 |
+
|
| 78 |
+
Approved exception：本批次使用冻结的C4096或DATA-019 PARTIAL2763拓扑，而不是默认96 train + 32 heldout。
|
| 79 |
+
|
| 80 |
+
## Normalization
|
| 81 |
+
|
| 82 |
+
| Field | Contract |
|
| 83 |
+
|---|---|
|
| 84 |
+
| Source | admitted train attempts only。 |
|
| 85 |
+
| Weighting | attempt等权；attempt内transition等权。 |
|
| 86 |
+
| Validation / heldout | train statistics transform-only。 |
|
| 87 |
+
|
| 88 |
+
## Validation and Readiness
|
| 89 |
+
|
| 90 |
+
| Gate | Result | Evidence |
|
| 91 |
+
|---|---|---|
|
| 92 |
+
| Completeness | PASS | terminal 4096；infra 0。 |
|
| 93 |
+
| Split leakage | PASS | 原split身份保留；reflection不进入训练。 |
|
| 94 |
+
| Member identity | PASS | manifest `9f99d35e6bede9c9c6ac58f0a40163294e83d81d3f5467f3220bad9881f7c9cd`。 |
|
| 95 |
+
| Release identity | PASS | archive `1fe5895ed77efa2c255a3503e679f0295493e2cb8505398cd6970576e207a599`。 |
|
| 96 |
+
| Dataset status | READY | READY / C4096 / INFRA0。 |
|
objects/phone/trajectories/phone_04/stage4/data041-dr-rollouts-v1/dataset_manifest.json
ADDED
|
The diff for this file is too large to render.
See raw diff
|
|
|
objects/phone/trajectories/phone_04/stage4/data041-dr-rollouts-v1/distribution_calibration.json
ADDED
|
@@ -0,0 +1,120 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"counts": {
|
| 3 |
+
"all": 4096,
|
| 4 |
+
"heldout": 512,
|
| 5 |
+
"train": 3072,
|
| 6 |
+
"validation": 512
|
| 7 |
+
},
|
| 8 |
+
"dimensions": [
|
| 9 |
+
{
|
| 10 |
+
"index": 0,
|
| 11 |
+
"max": 0.999777190387249,
|
| 12 |
+
"mean": 0.4999999995343387,
|
| 13 |
+
"min": 7.078051567077637e-08
|
| 14 |
+
},
|
| 15 |
+
{
|
| 16 |
+
"index": 1,
|
| 17 |
+
"max": 0.9997896309942007,
|
| 18 |
+
"mean": 0.4999999995343387,
|
| 19 |
+
"min": 0.00012197811156511307
|
| 20 |
+
},
|
| 21 |
+
{
|
| 22 |
+
"index": 2,
|
| 23 |
+
"max": 0.9998861188068986,
|
| 24 |
+
"mean": 0.4999999995343387,
|
| 25 |
+
"min": 0.0002353023737668991
|
| 26 |
+
},
|
| 27 |
+
{
|
| 28 |
+
"index": 3,
|
| 29 |
+
"max": 0.9999722978100181,
|
| 30 |
+
"mean": 0.4999999995343387,
|
| 31 |
+
"min": 0.0002089785411953926
|
| 32 |
+
},
|
| 33 |
+
{
|
| 34 |
+
"index": 4,
|
| 35 |
+
"max": 0.999808656051755,
|
| 36 |
+
"mean": 0.4999999995343387,
|
| 37 |
+
"min": 0.00017888471484184265
|
| 38 |
+
},
|
| 39 |
+
{
|
| 40 |
+
"index": 5,
|
| 41 |
+
"max": 0.9998929128050804,
|
| 42 |
+
"mean": 0.4999999995343387,
|
| 43 |
+
"min": 0.00017568469047546387
|
| 44 |
+
},
|
| 45 |
+
{
|
| 46 |
+
"index": 6,
|
| 47 |
+
"max": 0.9999021058902144,
|
| 48 |
+
"mean": 0.4999999995343387,
|
| 49 |
+
"min": 8.884258568286896e-05
|
| 50 |
+
},
|
| 51 |
+
{
|
| 52 |
+
"index": 7,
|
| 53 |
+
"max": 0.9998147236183286,
|
| 54 |
+
"mean": 0.4999999995343387,
|
| 55 |
+
"min": 8.64863395690918e-05
|
| 56 |
+
},
|
| 57 |
+
{
|
| 58 |
+
"index": 8,
|
| 59 |
+
"max": 0.9999467926099896,
|
| 60 |
+
"mean": 0.4999999995343387,
|
| 61 |
+
"min": 5.036313086748123e-05
|
| 62 |
+
},
|
| 63 |
+
{
|
| 64 |
+
"index": 9,
|
| 65 |
+
"max": 0.9998581157997251,
|
| 66 |
+
"mean": 0.4999999995343387,
|
| 67 |
+
"min": 0.00022254977375268936
|
| 68 |
+
},
|
| 69 |
+
{
|
| 70 |
+
"index": 10,
|
| 71 |
+
"max": 0.9998724302276969,
|
| 72 |
+
"mean": 0.4999999995343387,
|
| 73 |
+
"min": 0.0001163659617304802
|
| 74 |
+
},
|
| 75 |
+
{
|
| 76 |
+
"index": 11,
|
| 77 |
+
"max": 0.9998724861070514,
|
| 78 |
+
"mean": 0.4999999995343387,
|
| 79 |
+
"min": 3.587640821933746e-05
|
| 80 |
+
},
|
| 81 |
+
{
|
| 82 |
+
"index": 12,
|
| 83 |
+
"max": 0.9998503969982266,
|
| 84 |
+
"mean": 0.4999999995343387,
|
| 85 |
+
"min": 0.0001909751445055008
|
| 86 |
+
},
|
| 87 |
+
{
|
| 88 |
+
"index": 13,
|
| 89 |
+
"max": 0.9999571787193418,
|
| 90 |
+
"mean": 0.4999999995343387,
|
| 91 |
+
"min": 0.0002157101407647133
|
| 92 |
+
},
|
| 93 |
+
{
|
| 94 |
+
"index": 14,
|
| 95 |
+
"max": 0.9998373240232468,
|
| 96 |
+
"mean": 0.4999999995343387,
|
| 97 |
+
"min": 0.0001782132312655449
|
| 98 |
+
}
|
| 99 |
+
],
|
| 100 |
+
"gates": {
|
| 101 |
+
"G10_dimension_means": "PASS",
|
| 102 |
+
"G11_rsi_full_discrete_support": "PASS",
|
| 103 |
+
"G12_action_state_contract": "PASS",
|
| 104 |
+
"G1_exact_cardinality": "PASS",
|
| 105 |
+
"G2_dimension_15d": "PASS",
|
| 106 |
+
"G3_finite": "PASS",
|
| 107 |
+
"G4_unit_cube": "PASS",
|
| 108 |
+
"G5_unique_case_ids": "PASS",
|
| 109 |
+
"G6_unique_case_seeds": "PASS",
|
| 110 |
+
"G7_unique_dr_seeds": "PASS",
|
| 111 |
+
"G8_split_membership_frozen": "PASS",
|
| 112 |
+
"G9_dimension_extrema": "PASS"
|
| 113 |
+
},
|
| 114 |
+
"rsi_start_phase": {
|
| 115 |
+
"max": 346,
|
| 116 |
+
"min": 0
|
| 117 |
+
},
|
| 118 |
+
"schema": "sharpa_coldclone_formal_distribution_calibration_v1",
|
| 119 |
+
"status": "PASS"
|
| 120 |
+
}
|
objects/phone/trajectories/phone_04/stage4/data041-dr-rollouts-v1/dr_bank.json
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:4be2fdb9ce5452d54db9244f0a45b845f65c368121b136f911d7c7edb3d173fc
|
| 3 |
+
size 13928378
|
objects/phone/trajectories/phone_05/stage4/data031-dr-rollouts-v1/checksums.json
ADDED
|
@@ -0,0 +1,8 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"archive_sha256": "1cf4aa43da4f865a3761028e8b37502c65fc84d5033f055e88d7a4ada4b6e45e",
|
| 3 |
+
"archive_size_bytes": 935895040,
|
| 4 |
+
"data_task_id": "SHARPA-BC-DATA-031",
|
| 5 |
+
"manifest_sha256": "c14e8c19878b031a3856780fc13e006b700f9078684ba9641f29e93e8b74e5c0",
|
| 6 |
+
"phone_id": "phone_05",
|
| 7 |
+
"schema": "sharpa_stage4_release_checksums_v1"
|
| 8 |
+
}
|
objects/phone/trajectories/phone_05/stage4/data031-dr-rollouts-v1/data_card.md
ADDED
|
@@ -0,0 +1,139 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
# Data Card: SHARPA-BC-DATA-031 / 20260903_sharpa_ac_p05_001_phone_05_c4096_dr_rsi_v1
|
| 2 |
+
|
| 3 |
+
| Field | Value |
|
| 4 |
+
|---|---|
|
| 5 |
+
| Data Task ID | SHARPA-BC-DATA-031 |
|
| 6 |
+
| Dataset ID | 20260903_sharpa_ac_p05_001_phone_05_c4096_dr_rsi_v1 |
|
| 7 |
+
| Status | READY / C4096 / INFRA0 |
|
| 8 |
+
| Purpose | 从phone_05 `SHARPA-AC-P05-001` S1 selected checkpoint采集4096个与Teacher训练时RefAug/reset-only DR及uniform RSI分布一致的独立closed-loop rollouts，形成3072 train、512 validation、512 sealed heldout。 |
|
| 9 |
+
| Source mode | `teacher_rollout`；fresh Isaac/PhysX；每个attempt独立采样DR与RSI，从采样phase执行到reference terminal再加15 control-step hold。 |
|
| 10 |
+
| Source lineage | `SHARPA-AC-P05-001`；W&B S1 selected `model_7000.pt`；原始8-seed evaluation仅为diagnostic，P2=`NOT_RUN`。 |
|
| 11 |
+
| Episodes / splits | 4096 preregistered attempts=`3072 train + 512 validation + 512 sealed heldout`；失败保留原split身份并进入reflection/infra，禁止按结果补抽或搬移。 |
|
| 12 |
+
| Sampling | 15D=`DR14 + uniform RSI phase`；每row独立采样；先物化、校准并冻结4096-row bank，之后才能launch。 |
|
| 13 |
+
| Observation | 71D deployable current frame及attempt-local history-3 213D view；不含velocity、phase/reference、Teacher ID、DR latent或future state。 |
|
| 14 |
+
| Action / label | controller真实执行的31D absolute `final_applied_target`；Teacher 28D residual仅作diagnostic，不能作为Student label。 |
|
| 15 |
+
| Normalization | admitted train only；attempt等权、attempt内transition等权；validation/heldout transform-only；std floor `1e-8`。 |
|
| 16 |
+
| Producer | formal producer commit `fc39d2f03df50e8888e2fd1bc00afbfc1725f1e5`；source draft `config/data_collection/sharpa_multiteacher_c4096_v1/phone_05_c4096_dr_rsi_v1.json`；executable spec `outputs/formal031_v3/spec.json`。 |
|
| 17 |
+
| Artifact / manifest | create-only `outputs/formal031_v3`；bank SHA `c06c2f0b3ba853050dc3548a222275e89a603b7621aef79428b88b0af9318808`；dataset manifest待4096 receipts后生成。 |
|
| 18 |
+
| Claim boundary | 仅支持phone_05、同一Teacher、同RefAug/reset-only DR+uniform-RSI分布的动作蒸馏研究；不支持physics/noise/push、跨episode、跨物体、P2 robustness或真机泛化。 |
|
| 19 |
+
|
| 20 |
+
## Source and Provenance
|
| 21 |
+
|
| 22 |
+
| Artifact | Identity / value | Evidence |
|
| 23 |
+
|---|---|---|
|
| 24 |
+
| Teacher Task | `SHARPA-AC-P05-001` | Auto-Curriculum S1 RefAug/reset-only DR candidate。 |
|
| 25 |
+
| Teacher Training Card | `teacher/checkpoints/sharpa_phone20_autocurriculum_selected_v1/phone_05/inputs/training_card.md`；SHA `394c778edd2d31fcddb93c3bf26a4e30c87ff3afc0d4ada0e6570246df0df2bc` | W&B selected-checkpoint artifact原样下载。 |
|
| 26 |
+
| Selected checkpoint | `teacher/checkpoints/sharpa_phone20_autocurriculum_selected_v1/phone_05/model_7000.pt`；`model_7000.pt`；iteration `7000`；SHA `4387769c731d96cc7a26580177664b431ce8ba361d1eed7a6ba17e9c71325283` | W&B artifact `dfnet/dex-retarget/sharpa-phone_05-autocurriculum-s1-v1-s1-selected-checkpoint:v0`；local SHA gate PASS。 |
|
| 27 |
+
| Teacher code | commit `4a146fe07ef7f421745c3506ca803594f39b4fc8` | checkpoint selection manifest。 |
|
| 28 |
+
| Runtime manifest | `teacher/checkpoints/sharpa_phone20_autocurriculum_selected_v1/phone_05/inputs/runtime_manifest.json`；SHA `63272d0e2eff6af2e5857f92cb386ecf987c248476f15cb8c985919fc86be5ec` | 仅保存manifest；完整可执行runtime仍须在G13物化并回读。 |
|
| 29 |
+
| Curriculum manifest | `teacher/checkpoints/sharpa_phone20_autocurriculum_selected_v1/phone_05/inputs/curriculum_manifest.json`；SHA `a49b2f37a0b6fad60f341712161775b5637c4f7fe057dd5d3004231f53724a65` | S0 nominal → S1 RefAug/reset-only DR。 |
|
| 30 |
+
| Stage2 handoff | `teacher/checkpoints/sharpa_phone20_autocurriculum_selected_v1/phone_05/inputs/stage2_handoff.json`；SHA `1f6f486535bef9cbd0c947dc0208be42e551c0ddeb007237c8a45e501111eeb9`；status `STAGE2_HANDOFF_ACCEPTED` | W&B artifact原样下载。 |
|
| 31 |
+
| Stage1 NPZ identity | SHA `871063aabcad4abf5b13e728f7c3cb3b02e6d6c439902ddef826024a5b936e1a` | bytes尚未随本checkpoint bundle下载；G13前必须解析。 |
|
| 32 |
+
| Human H5 identity | SHA `d1489d0612041d1837a120f2d8babb4ae06550d23f3ac0bc6789e553e27a4336` | `663` frames at `60.0` Hz。 |
|
| 33 |
+
| Runtime retarget H5 identity | SHA `9fe13a0aae246b2e727fedb0c03e54a69b3ca8ad28f1982e046beaf810a3d1fd` | `331` nominal policy targets at `30.0` Hz。 |
|
| 34 |
+
| Reference manifest identity | SHA `322c7be195ad458c2bdadd0d9303fc56c46977e56e9c86089bb59c5449670f0f` | Stage2 accepted reference。 |
|
| 35 |
+
| GoalContract identity | SHA `2d2b0d99b4f93e76e1e6ca3ae1552980fa3ea807b0c63857313b48ad256746c1` | runtime artifact manifest。 |
|
| 36 |
+
| Historical evaluation | [W&B seed0–7 diagnostic](https://wandb.ai/dfnet/dex-retarget/runs/sharpa-phone_05-autocurriculum-evaluation-seed0-7-v1) | checkpoint identity matched；P2 `NOT_RUN / NON_PROMOTABLE`。 |
|
| 37 |
+
|
| 38 |
+
## Schema
|
| 39 |
+
|
| 40 |
+
| Field | Shape / type | Semantics |
|
| 41 |
+
|---|---|---|
|
| 42 |
+
| `time_s` | `(S,) float64` | 30 Hz control timestamps from the sampled RSI phase。 |
|
| 43 |
+
| `actual_q` | `(S,22) float32`, rad | PhysX actual Sharpa finger joints。 |
|
| 44 |
+
| `actual_wrist_pose_wxyz` | `(S,7) float32`, world frame | actual wrist translation and quaternion。 |
|
| 45 |
+
| `actual_object_pose_wxyz` | `(S,7) float32`, world frame | actual phone translation and quaternion。 |
|
| 46 |
+
| `raw_residual` / `processed_residual` | `(A,28) float32` | Teacher diagnostic action before/after adapter processing。 |
|
| 47 |
+
| `final_applied_target_wxyz` | `(A,29) float32` | runtime controller target representation。 |
|
| 48 |
+
| `student_frame` | `(A,71) float32` | deployable current-frame observation。 |
|
| 49 |
+
| `applied_target_31` | `(A,31) float32` | unique BC supervision label: wrist position3 + rotation6D6 + joint22。 |
|
| 50 |
+
| `state_t1_student_frame` | `(A,71) float32` | post-physics next state，audit-only。 |
|
| 51 |
+
| `pose_history3_mlp_v1` | `(A,213) float32` | three attempt-local 71D frames，oldest→current；first-frame repeat padding。 |
|
| 52 |
+
| `normalized_sampling_15d` | `(15,) float64` | DR14+RSI provenance-only；不得进入policy input。 |
|
| 53 |
+
| reset/failure/action ledgers | JSON | sampled parameters、actual runtime readback、drop/invalid/infra与action-chain evidence。 |
|
| 54 |
+
|
| 55 |
+
## Temporal Alignment
|
| 56 |
+
|
| 57 |
+
| Item | Contract |
|
| 58 |
+
|---|---|
|
| 59 |
+
| Rates | PhysX 120 Hz；policy/control 30 Hz；decimation=4。 |
|
| 60 |
+
| Transition | `obs_t → Teacher residual_t → controller final_applied_target_t → four PhysX substeps → state_t+1`。 |
|
| 61 |
+
| RSI | `u_phase ~ U[0,1)`；`start=floor(u_phase × 331)`，clamp到`0…330`。 |
|
| 62 |
+
| Horizon | `A=(331-start)+15`；`S=A+1`；即使drop/invalid也记录完整终态与ledger。 |
|
| 63 |
+
| History | strictly attempt-local；不得跨attempt、split或index gap。 |
|
| 64 |
+
| Future windows | 后续Consumer可构造`T_p` window，但window不得跨attempt；tail必须masked。 |
|
| 65 |
+
|
| 66 |
+
## Simulator Environment Randomization
|
| 67 |
+
|
| 68 |
+
| Field | Value |
|
| 69 |
+
|---|---|
|
| 70 |
+
| Enabled | true |
|
| 71 |
+
| Profile | `SHARPA-AC-P05-001` S1 RefAug/reset-only DR + uniform RSI。 |
|
| 72 |
+
| Source config SHA | `d99efe6b27a179211fec8f3077e5605f9d804c666f0ac9a16c3e8349795982b8`。 |
|
| 73 |
+
| Activation | every reset；nominal fraction `0.0`。 |
|
| 74 |
+
| Mask | common rigid transform同时作用hand与phone；relative reset保持grasp-shaped范围。 |
|
| 75 |
+
| Seed | 4096-row bank seed/namespace在distribution calibration阶段冻结；当前`NOT_YET_MATERIALIZED`。 |
|
| 76 |
+
|
| 77 |
+
| Category | Parameter | Distribution / range | Sampling / effect |
|
| 78 |
+
|---|---|---|---|
|
| 79 |
+
| Reference augmentation | common X/Y | each `U[-0.05,0.05] m` | 同时作用hand和phone。 |
|
| 80 |
+
| Reference augmentation | common yaw | `U[-30°,30°]` | 同时作用hand和phone。 |
|
| 81 |
+
| Relative reset | phone XYZ | each `U[-0.006,0.006] m` | after RSI/reference selection。 |
|
| 82 |
+
| Relative reset | phone rotation | axis uniformly sampled；signed angle `U[-3°,3°]` | every reset。 |
|
| 83 |
+
| Finger reset | shared wrap | `U[0,0.06] rad` | coupled flexion joints，只增加包握。 |
|
| 84 |
+
| Finger reset | four-finger rub | each `U[-0.025,0.025] rad` | independent rub joints。 |
|
| 85 |
+
| RSI | normalized phase | `U[0,1)` | maps to discrete start phase。 |
|
| 86 |
+
|
| 87 |
+
### Explicitly disabled / not applicable
|
| 88 |
+
|
| 89 |
+
| Option | State | Meaning |
|
| 90 |
+
|---|---|---|
|
| 91 |
+
| Physics/actuator randomization | disabled | mass、COM、inertia、friction与drive固定。 |
|
| 92 |
+
| Observation/noise/lag | disabled | clean observation；zero lag。 |
|
| 93 |
+
| Disturbance/push | disabled | none。 |
|
| 94 |
+
| Gravity curriculum | disabled | fixed full gravity `-9.81 m/s²`。 |
|
| 95 |
+
|
| 96 |
+
## Splits, Sampling, and Leakage
|
| 97 |
+
|
| 98 |
+
Approved Exception: Shengtao于2026-09-03批准每个独立Teacher使用4096 attempts并划分为3072 train、512 validation、512 sealed heldout，明确偏离默认96 train + 32 heldout topology。
|
| 99 |
+
|
| 100 |
+
| Split | Episodes / rows | Purpose |
|
| 101 |
+
|---|---:|---|
|
| 102 |
+
| train | 3072 | admitted rows的唯一optimization与normalization来源。 |
|
| 103 |
+
| validation | 512 | offline checkpoint selection；不得参与normalization。 |
|
| 104 |
+
| sealed heldout | 512 | 所有模型选择冻结后的同分布unseen-attempt audit。 |
|
| 105 |
+
| reflection | E5/E6/E9失败 | 保留预注册split身份；不得进入监督数据或被成功补抽替换。 |
|
| 106 |
+
| infra_failure | identity/runtime/schema failure | 阻止Dataset READY；不得解释成policy failure。 |
|
| 107 |
+
|
| 108 |
+
Split membership必须在rollout前由case ID确定。同一seed、trajectory window、parent/derived copy不得跨split；heldout不得反向用于collector、normalization、checkpoint、模型或超参数选择。
|
| 109 |
+
|
| 110 |
+
## Normalization
|
| 111 |
+
|
| 112 |
+
| Field | Contract |
|
| 113 |
+
|---|---|
|
| 114 |
+
| Source | admitted train attempts only。 |
|
| 115 |
+
| Weighting | each rollout equal；within each rollout，each transition equal。 |
|
| 116 |
+
| Stored arrays | MLP213、GRU71与target31 mean/second moment/std，以及exact train attempt IDs和weights。 |
|
| 117 |
+
| Floor | `std=max(std,1e-8)`。 |
|
| 118 |
+
| Validation / heldout | train statistics transform-only；禁止重算。 |
|
| 119 |
+
|
| 120 |
+
## Validation and Readiness
|
| 121 |
+
|
| 122 |
+
| Gate | Result | Evidence |
|
| 123 |
+
|---|---|---|
|
| 124 |
+
| Data Card format | PASS | unified validator对本批次7张Card报告`validated=7 failures=0`。 |
|
| 125 |
+
| Selected checkpoint identity | PASS | W&B selected artifact与local SHA一致：`4387769c731d96cc7a26580177664b431ce8ba361d1eed7a6ba17e9c71325283`。 |
|
| 126 |
+
| Stage2/H5/runtime resolvability | PASS | `outputs/prepared_v1`中clone-local H5/assets/runtime逐SHA解析通过。 |
|
| 127 |
+
| G1–G12 distribution calibration | PASS | `outputs/formal031_v3/distribution_calibration.json`；SHA `f19f6603944019023191e1f4770da52e1884bcdeb5dd3ab3f12d6389efe1707d`。 |
|
| 128 |
+
| G13 identity/runtime | PASS | producer `fc39d2f`；checkpoint、runtime、artifact、goal与spec/bank SHA preflight通过。 |
|
| 129 |
+
| G14 one-attempt RSI smoke | PASS | v3正式bank真实non-PLAY policy path运行通过。 |
|
| 130 |
+
| G15 multi-row distribution readback | PASS | 8 train + 1 validation + 1 sealed-heldout正式rows运行通过；runtime readback与DR/RSI ledger通过。 |
|
| 131 |
+
| G16 completeness | PASS | 4096 terminal receipts；infra0；reflection226。 |
|
| 132 |
+
| G17 schema/timing | PASS | MLP213/GRU71/target31/state-t+1 71；所有admitted attempts满足`S=A+1`。 |
|
| 133 |
+
| G18 split/normalization | PASS | train2899/538300 transitions；validation485/90164；heldout486/89967；normalization仅来自admitted train。 |
|
| 134 |
+
| G19 final manifest | PASS | manifest SHA `c14e8c19878b031a3856780fc13e006b700f9078684ba9641f29e93e8b74e5c0`；archive SHA `1cf4aa43da4f865a3761028e8b37502c65fc84d5033f055e88d7a4ada4b6e45e`。 |
|
| 135 |
+
| DATASET_READY | READY | 只支持phone_05 same-profile Teacher imitation。 |
|
| 136 |
+
|
| 137 |
+
## Approved Exception
|
| 138 |
+
|
| 139 |
+
Shengtao于2026-09-03批准本批次每个独立Teacher请求4096个DR+RSI rollout，固定3072/512/512三split，偏离默认128-attempt topology。该批准只授权Draft Card与checkpoint准备；正式并行collection须在Cards审核和G1–G15通过后另行启动。
|
objects/phone/trajectories/phone_05/stage4/data031-dr-rollouts-v1/dataset_manifest.json
ADDED
|
The diff for this file is too large to render.
See raw diff
|
|
|
objects/phone/trajectories/phone_05/stage4/data031-dr-rollouts-v1/rollout_dataset.tar
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:1cf4aa43da4f865a3761028e8b37502c65fc84d5033f055e88d7a4ada4b6e45e
|
| 3 |
+
size 935895040
|
objects/phone/trajectories/phone_07/stage4/data019-dr-rollouts-v1/checksums.json
ADDED
|
@@ -0,0 +1,8 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"archive_sha256": "61feae2f23dae2f16e096aa53d5d6cb24ce9062fdff8ec6c3421dfc4e9bee31d",
|
| 3 |
+
"archive_size_bytes": 691230720,
|
| 4 |
+
"data_task_id": "SHARPA-BC-DATA-019",
|
| 5 |
+
"manifest_sha256": "740ff1d0f02eab5429c97d7c6ee9054696133684d9999e57a8d1d12f57faf405",
|
| 6 |
+
"phone_id": "phone_07",
|
| 7 |
+
"schema": "sharpa_stage4_release_checksums_v1"
|
| 8 |
+
}
|
objects/phone/trajectories/phone_07/stage4/data019-dr-rollouts-v1/data_card.md
ADDED
|
@@ -0,0 +1,96 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
# Data Card: SHARPA-BC-DATA-019 / 20260831_t0149_model7800_phone07_partial2763_pow_v1
|
| 2 |
+
|
| 3 |
+
| Field | Value |
|
| 4 |
+
|---|---|
|
| 5 |
+
| Data Task ID | SHARPA-BC-DATA-019 |
|
| 6 |
+
| Dataset ID | 20260831_t0149_model7800_phone07_partial2763_pow_v1 |
|
| 7 |
+
| Status | READY / PROOF-OF-WORK SNAPSHOT / PARENT COVERAGE GAP |
|
| 8 |
+
| Purpose | 将用户停止DATA-017采集时已经具备完整terminal receipt的2763个phone07 Teacher rollouts冻结为可训练的proof-of-work BC Dataset。 |
|
| 9 |
+
| Source mode | derived_view；从DATA-017原始Teacher rollout attempts选择完整terminal cases，label和rollout bytes不变。 |
|
| 10 |
+
| Source lineage | SHARPA-BC-DATA-017；SHARPA-T0-149 model7800；同一PPO-matched DR14+uniform-RSI bank。 |
|
| 11 |
+
| Episodes / splits | 2763 terminal attempts=`2073 train +359 validation +331 heldout`；admitted=`1936/340/311`；reflection=176。 |
|
| 12 |
+
| Sampling | 选择规则是2026-08-31用户停止采集边界时已存在完整terminal receipt；这是post-hoc availability snapshot，不宣称近似完整4096-row分布。 |
|
| 13 |
+
| Observation | MLP history-3 213D与GRU current-frame 71D deployable views；不含velocity、phase、reference、Teacher ID或DR latent。 |
|
| 14 |
+
| Action / label | Teacher/controller真实执行的31D absolute target；28D residual仅保留在parent action-chain diagnostic。 |
|
| 15 |
+
| Normalization | 仅1936个admitted train attempts；rollout等权、rollout内transition等权；validation/heldout transform-only；std floor`1e-8`。 |
|
| 16 |
+
| Producer | commit`73698f2`；`student.sharpa.teacher_rollout_finalize` snapshot selector模式。 |
|
| 17 |
+
| Artifact / manifest | `outputs/sharpa_bc_data019/20260831_t0149_model7800_phone07_partial2763_pow_v1/dataset_v1`；manifest SHA`740ff1d0f02eab5429c97d7c6ee9054696133684d9999e57a8d1d12f57faf405`。 |
|
| 18 |
+
| Claim boundary | 只证明phone07同Teacher partial snapshot可用于BC训练；禁止宣称C4096 coverage、无偏PPO分布近似、跨demo、跨物体、闭环成功或真机泛化。 |
|
| 19 |
+
|
| 20 |
+
## Source and Provenance
|
| 21 |
+
|
| 22 |
+
| Artifact | Identity / value | Evidence |
|
| 23 |
+
|---|---|---|
|
| 24 |
+
| Parent Data Card | SHARPA-BC-DATA-017 | 原C4096采集合同保持NOT_READY/STOPPED，不被DATA-019倒改。 |
|
| 25 |
+
| Teacher Task | SHARPA-T0-149 | phone07 RefAug/reset-only DR Auto-Curriculum Teacher。 |
|
| 26 |
+
| Teacher checkpoint | model7800 SHA`6a364bd96390372787f2a2518fc85be44f399263fd90f1c615daf33671713ff2` | 与parent receipts一致。 |
|
| 27 |
+
| Parent CollectionSpec | SHA`1275eddd65a5ea17c5424ff17abf51ca8c375637975585dbcaac5e3939557adf` | 冻结runtime、controller、DR/RSI和codec。 |
|
| 28 |
+
| Parent source bank | SHA`7861b4b1af651b30cddde1d2455eaa15a0b7da8f61c4c1dc6e4639ef7027d833` | 原4096-row phone07 source bank。 |
|
| 29 |
+
| Snapshot selector | SHA`430b10f833a48321fa8f8f5d7dd4c7cfab1603de73c389c25ed8351c1003dcc6` | 2763个唯一case IDs及原split membership。 |
|
| 30 |
+
| Dataset manifest | SHA`740ff1d0f02eab5429c97d7c6ee9054696133684d9999e57a8d1d12f57faf405` | 5186个create-only members。 |
|
| 31 |
+
| Normalization | SHA`597501b789d6075cb2580c5c8b620c36ce05412b190e137f29d804a59ad07e58` | attempt IDs严格等于admitted train index。 |
|
| 32 |
+
|
| 33 |
+
## Schema
|
| 34 |
+
|
| 35 |
+
| Field | Shape / type | Semantics |
|
| 36 |
+
|---|---|---|
|
| 37 |
+
| MLP observation | `(A,213)` float32 | 三个71D deployable frames，旧帧到当前帧；attempt起点使用首帧重复padding。 |
|
| 38 |
+
| GRU observation | `(A,71)` float32 | 当前deployable state frame。 |
|
| 39 |
+
| target | `(A,31)` float32 | 9D wrist absolute pose target +22D hand joint absolute target。 |
|
| 40 |
+
| transition_index | `(A,)` int64 | attempt-local action/control index。 |
|
| 41 |
+
| state_t1_student_frame | `(A,71)` float32 | 执行target并完成physics substeps后的下一状态，仅用于时间审计。 |
|
| 42 |
+
| indexes | JSON rows | case ID、source、原split、RSI/start phase、view path/SHA、transition count。 |
|
| 43 |
+
| normalization | NPZ | MLP/GRU/target mean、second moment、std、train attempt IDs和weights。 |
|
| 44 |
+
|
| 45 |
+
所有view都按attempt独立存储；窗口不得跨attempt边界。坐标、单位、field ordering和action codec继承parent CollectionSpec，不做重排或label变换。
|
| 46 |
+
|
| 47 |
+
## Temporal Alignment
|
| 48 |
+
|
| 49 |
+
| Item | Contract |
|
| 50 |
+
|---|---|
|
| 51 |
+
| Physics / control | PhysX 120Hz；policy/control 30Hz；decimation=4。 |
|
| 52 |
+
| Transition | `obs_t -> Teacher 28D residual -> controller 31D final_applied_target_t -> 4 physics substeps -> state_t+1`。 |
|
| 53 |
+
| Stored label | 31D `final_applied_target_t`，不是raw/processed 28D residual。 |
|
| 54 |
+
| State/action count | 每个admitted attempt满足`S=A+1`。 |
|
| 55 |
+
| History | attempt-local history-3；起始padding不读取其他attempt。 |
|
| 56 |
+
| Terminal | 从冻结RSI phase执行到reference end+15 hold；reflection保留但不进入监督view。 |
|
| 57 |
+
|
| 58 |
+
## Splits, Sampling, and Leakage
|
| 59 |
+
|
| 60 |
+
| Split | Episodes / rows | Purpose |
|
| 61 |
+
|---|---:|---|
|
| 62 |
+
| train | 2073 attempts；1936 admitted | 唯一训练和normalization来源。 |
|
| 63 |
+
| validation | 359 attempts；340 admitted | checkpoint选择与离线误差；不得参与normalization。 |
|
| 64 |
+
| heldout | 331 attempts；311 admitted | 所有训练决策冻结后的同snapshot审计；不得参与checkpoint选择或normalization。 |
|
| 65 |
+
| reflection | 176 attempts；0 admitted | E5/E6/E9未全部通过；保留原split身份，不训练。 |
|
| 66 |
+
| infra | 0 | 任一infra会阻止READY。 |
|
| 67 |
+
|
| 68 |
+
Train、validation和heldout case IDs两两不相交。DATA-019 selection是按停止时receipt availability形成，而非重新随机抽样；因此只能称为partial snapshot，不能把split比例或统计量外推为完整C4096分布。
|
| 69 |
+
|
| 70 |
+
## Normalization
|
| 71 |
+
|
| 72 |
+
| Field | Contract |
|
| 73 |
+
|---|---|
|
| 74 |
+
| Source | admitted train 1936 attempts only。 |
|
| 75 |
+
| Weighting | 每个rollout等权；每个rollout内部transition等权。 |
|
| 76 |
+
| Stored arrays | `mlp_observation_*`、`gru_frame_*`、`target_*`、attempt IDs/counts/weights。 |
|
| 77 |
+
| Numerical floor | std最低`1e-8`。 |
|
| 78 |
+
| Validation / heldout | 使用train statistics transform-only。 |
|
| 79 |
+
|
| 80 |
+
## Validation and Readiness
|
| 81 |
+
|
| 82 |
+
| Gate | Result | Evidence |
|
| 83 |
+
|---|---|---|
|
| 84 |
+
| Selector identity | PASS | 2763 unique case IDs；parent bank SHA匹配。 |
|
| 85 |
+
| Terminal completeness | PASS | selector中的2763 cases全部具有terminal receipt。 |
|
| 86 |
+
| Infra | PASS | 0 infra。 |
|
| 87 |
+
| Admission | PASS | 2587 admitted；176 reflection独立索引。 |
|
| 88 |
+
| Schema/timing | PASS | 71/213/31/71；admitted attempts均`S=A+1`。 |
|
| 89 |
+
| Split leakage | PASS | 三split indexes不相交。 |
|
| 90 |
+
| Normalization leakage | PASS | normalization attempt IDs严格等于1936 train IDs。 |
|
| 91 |
+
| Member identity | PASS | 5186 members写入manifest。 |
|
| 92 |
+
| Dataset status | READY | manifest status=`READY`，claim仅限proof-of-work snapshot。 |
|
| 93 |
+
|
| 94 |
+
## Approved Exception
|
| 95 |
+
|
| 96 |
+
Shengtao于2026-08-31明确批准停止DATA-017 C4096采集，并将当时完成的2763条另立为SHARPA-BC-DATA-019，用于验证BC能否work。该批准允许偏离默认128/96/32 topology和原DATA-017 C4096目标，但不允许把DATA-017本身标记READY，也不允许扩大DATA-019的coverage或泛化宣称。
|
objects/phone/trajectories/phone_07/stage4/data019-dr-rollouts-v1/dataset_manifest.json
ADDED
|
The diff for this file is too large to render.
See raw diff
|
|
|
objects/phone/trajectories/phone_07/stage4/data019-dr-rollouts-v1/rollout_dataset.tar
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:61feae2f23dae2f16e096aa53d5d6cb24ce9062fdff8ec6c3421dfc4e9bee31d
|
| 3 |
+
size 691230720
|
objects/phone/trajectories/phone_09/stage3/selected-teacher-v1/checkpoint_receipt.json
ADDED
|
@@ -0,0 +1,17 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"checkpoint": {
|
| 3 |
+
"path": "model_8500.pt",
|
| 4 |
+
"sha256": "597263648cd50b1410afca20e48e7d2f44a472f6ef4c35974d923a742b6b4159",
|
| 5 |
+
"size_bytes": 21074741
|
| 6 |
+
},
|
| 7 |
+
"claim_boundary": "source identity only; no new evaluation or promotion claim",
|
| 8 |
+
"phone_id": "phone_09",
|
| 9 |
+
"schema": "sharpa_stage3_selected_teacher_receipt_v1",
|
| 10 |
+
"selection_basis": "frozen source Teacher for SHARPA-BC-DATA-043",
|
| 11 |
+
"task_id": "SHARPA-AC-P09-003",
|
| 12 |
+
"training_card": {
|
| 13 |
+
"path": "training_card.md",
|
| 14 |
+
"sha256": "7cd6a9af698e31e4ef186cfdc2dcaa6541f07a572cc8b63ac7f9387ba490235f",
|
| 15 |
+
"size_bytes": 8250
|
| 16 |
+
}
|
| 17 |
+
}
|
objects/phone/trajectories/phone_09/stage3/selected-teacher-v1/model_8500.pt
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:597263648cd50b1410afca20e48e7d2f44a472f6ef4c35974d923a742b6b4159
|
| 3 |
+
size 21074741
|
objects/phone/trajectories/phone_09/stage3/selected-teacher-v1/training_card.md
ADDED
|
@@ -0,0 +1,118 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
# Training Card: phone_09_autocurriculum_retry_20260924_v3
|
| 2 |
+
|
| 3 |
+
| Field | Value |
|
| 4 |
+
|---|---|
|
| 5 |
+
| Description | [SHARPA-AC-P09-003] accepted phone_09 Stage2 reference的create-only基础设施修复重试：`No-DR plateau → RefAug/reset-only DR plateau → 8-seed 2×3 → W&B artifact/readback`正式teacher候选。原P09-001/P09-002失败证据保持不可变。 |
|
| 6 |
+
| Role | T0 residual-teacher Auto-Curriculum mentor batch candidate |
|
| 7 |
+
| Policy | Asymmetric MLP residual PPO；actor/critic `[1024,512,256,128]`；ELU；S0 fresh；S1 model-only warm start、fresh optimizer |
|
| 8 |
+
| Policy input | 162D actor；199D asymmetric critic；history total 3 frames |
|
| 9 |
+
| Hidden state | false |
|
| 10 |
+
| Action | 28D normalized residual：6D floating root + 22D Sharpa fingers；no chunk `(1/1)` |
|
| 11 |
+
| Data | accepted phone_09 handoff SHA256 `0f26cfe1ceff5b625bf8cd74e0592e97ac186f4f2a9d58a2243c3c86986599f7`；moving28 H5 SHA256 `0118bc159c33cd69247a01628e4280eb83d90ee63b088be863483de4a53da635` |
|
| 12 |
+
| Training | S0 `Regrind-Sharpa-Wuji2ActuatorParity-RouteA-v0`后S1 `Regrind-Sharpa-Wuji2ActuatorParity-RefAugDR-RouteA-v1`；各4096 env、24 steps/env、seed1、最多5000 iterations；checkpoint interval 500 |
|
| 13 |
+
| Reward profile | exact T0-153/WUJI2 RouteA reward、termination与PPO；S1只开启冻结RefAug/reset-only DR |
|
| 14 |
+
| Claim | 训练和发布完成只证明单episode remote reproduction；8-seed为diagnostic；P2仍`NOT_RUN`，不得声称robust teacher或真机成功。 |
|
| 15 |
+
| Metric | matching 16-env gate；S0/S1 plateau selected checkpoint；8/8 rollout与2×3；W&B checkpoint/video artifact下载SHA回读 |
|
| 16 |
+
| W&B | entity `dfnet`；project `dex-retarget`；group `sharpa-phone_09-autocurriculum-retry-20260924-v3`；run IDs `sharpa-phone_09-autocurriculum-s0-retry-20260924-v3`、`sharpa-phone_09-autocurriculum-s1-retry-20260924-v3`、`sharpa-phone_09-autocurriculum-evaluation-seed0-7-retry-20260924-v3` |
|
| 17 |
+
|
| 18 |
+
## Policy Inputs
|
| 19 |
+
|
| 20 |
+
| Field | Dim | Description |
|
| 21 |
+
|---|---:|---|
|
| 22 |
+
| object_position_world | 3 | 当前手机世界位置 |
|
| 23 |
+
| object_rotation_6d_world | 6 | 当前手机世界朝向连续 6D 表示 |
|
| 24 |
+
| wrist_position_history_3 | 9 | 当前及前两控制步 root 位置 |
|
| 25 |
+
| wrist_rotation_6d_history_3 | 18 | 当前及前两控制步 root 朝向 |
|
| 26 |
+
| joint_position_history_3 | 66 | 当前及前两控制步 22D Sharpa 关节角 |
|
| 27 |
+
| previous_raw_action | 28 | 上一控制步 raw residual action |
|
| 28 |
+
| demonstration_phase | 1 | 当前 reference index 的归一化 phase |
|
| 29 |
+
| action_base_wrist_position_rotation_6d | 9 | 当前 reference root target |
|
| 30 |
+
| action_base_joint_position | 22 | 当前 22D reference joint target |
|
| 31 |
+
|
| 32 |
+
Critic 在 actor 162D 上增加五个语义 fingertip world positions（15D）和 22D joint velocity，共 199D；这些 privileged 量不进入 actor。
|
| 33 |
+
|
| 34 |
+
## Simulator Environment Randomization
|
| 35 |
+
|
| 36 |
+
| Field | Value |
|
| 37 |
+
|---|---|
|
| 38 |
+
| Enabled | true |
|
| 39 |
+
| Source config | `sharpa/configs/sharpa_managerbased_pipeline/dr_contract_refaug_reset_only_v1.json` |
|
| 40 |
+
| Source config SHA256 | `d99efe6b27a179211fec8f3077e5605f9d804c666f0ac9a16c3e8349795982b8` |
|
| 41 |
+
| Runtime readback | 每次gate/formal另写create-only readback；必须17/17 matching |
|
| 42 |
+
| Runtime readback SHA256 | 冻结reference `21fbe1950c87fd029ce03827f204f70190845286ac7dd7d0d61165c2c307a804`；本次执行另写真实SHA |
|
| 43 |
+
| Activation | S0关闭randomization；S1全部4096 env开启；nominal fraction `0.0` |
|
| 44 |
+
| Schedule | S0独立plateau后切S1；S1重新初始化stage-local EMA；reference fade `.135→.292` |
|
| 45 |
+
| Mask | common rigid transform同时作用hand与phone；relative reset保持grasp-shaped范围 |
|
| 46 |
+
| Seed | 两阶段seed1 |
|
| 47 |
+
|
| 48 |
+
| Category | Parameter | Distribution / range | Sampling / effect |
|
| 49 |
+
|---|---|---|---|
|
| 50 |
+
| Reset | S0 reset | fixed；保留uniform RSI | No-DR nominal stage |
|
| 51 |
+
| Reset | S1 phone relative XYZ | each `U[-0.006,0.006] m` | every reset |
|
| 52 |
+
| Reset | S1 phone relative rotation | axis-angle `U[-3°,3°]` | every reset |
|
| 53 |
+
| Reset | S1 finger wrap | `U[0,0.06] rad` | 只增加包握 |
|
| 54 |
+
| Reset | S1 four-finger rub | `U[-0.025,0.025] rad` | 四指沿手机小范围扰动 |
|
| 55 |
+
| Reference augmentation | common XY | each `U[-0.05,0.05] m` | hand与phone共用 |
|
| 56 |
+
| Reference augmentation | common yaw | `U[-30°,30°]` | hand与phone共用 |
|
| 57 |
+
| Reference augmentation | phase fade | full through `.135`；zero at `.292` | 平滑退火回原reference |
|
| 58 |
+
|
| 59 |
+
### Explicitly disabled / not applicable
|
| 60 |
+
|
| 61 |
+
| Option | State | Meaning |
|
| 62 |
+
|---|---|---|
|
| 63 |
+
| Physics randomization | disabled | mass/COM/inertia/friction/drive固定；full gravity `-9.81 m/s²` |
|
| 64 |
+
| Observation/noise/lag | disabled | clean、zero-lag observation |
|
| 65 |
+
| Disturbance/push | disabled | 无external push |
|
| 66 |
+
|
| 67 |
+
## Rewards
|
| 68 |
+
|
| 69 |
+
| Reward | Weight | Description |
|
| 70 |
+
|---|---:|---|
|
| 71 |
+
| object_keypoints_pos_error_exp | 3.0 | 100 个手机表面点跟随 augmented reference |
|
| 72 |
+
| object_lin_vel_error_exp | 1.0 | 手机线速度跟随 augmented reference |
|
| 73 |
+
| object_ang_vel_error_exp | 1.0 | 手机角速度跟随 augmented reference |
|
| 74 |
+
| hand_wrist_pos_error_exp | 0.1 | root 位置软约束 |
|
| 75 |
+
| hand_wrist_rot_error_exp | 0.1 | root 朝向软约束 |
|
| 76 |
+
| action_l2_exp | 0.5 | residual 幅度约束 |
|
| 77 |
+
| action_rate_l2_exp | 1.0 | residual 平滑约束 |
|
| 78 |
+
| action_out_of_bounds_exp | 1.0 | raw action `[-1,1]` 越界惩罚 |
|
| 79 |
+
| early_terminated_penalty | -10.0 | 非 timeout 提前终止惩罚 |
|
| 80 |
+
| fingertip_object_graph_error_exp | 0.75 | 五指尖—手机相对几何跟随 augmented reference |
|
| 81 |
+
|
| 82 |
+
## Stage2 Reference Acceptance
|
| 83 |
+
|
| 84 |
+
| Field | Frozen evidence |
|
| 85 |
+
|---|---|
|
| 86 |
+
| Stage1 source | SHA256 `536a87d5853b4e7b090906f5d51df65b3fcb5cc9cb6c3a85d071a9b44beb5d76` |
|
| 87 |
+
| Stage1 manifest | SHA256 `17a6e91994a90c33f8e8c6b09f53a63f0424c2bd7533f202538d1239b30e2954` |
|
| 88 |
+
| Runtime human H5 | SHA256 `83bbdc431fc6fa7149ba403ab5de99c9c0b207f74bce94880643bf0aad46f078` |
|
| 89 |
+
| Canonical retarget H5 | SHA256 `2f46b49b427f1e938b19cda6cb58f3382033006dfc0072eb7c4c72656e8aa1dc` |
|
| 90 |
+
| Runtime retarget H5 | SHA256 `0118bc159c33cd69247a01628e4280eb83d90ee63b088be863483de4a53da635` |
|
| 91 |
+
| Reference manifest | SHA256 `5066ec4a4df8287e6b5f1998c57c2fe2b5e56401c97dcff28e549fcd669928ae` |
|
| 92 |
+
| Handoff | SHA256 `0f26cfe1ceff5b625bf8cd74e0592e97ac186f4f2a9d58a2243c3c86986599f7`；strict loader `STAGE2_HANDOFF_ACCEPTED` |
|
| 93 |
+
| Timeline | 1025 frames @ 60 Hz → 513 policy targets @ 30 Hz |
|
| 94 |
+
| Stage2 QA | validation receipt SHA256 `3f4cf26fcad534a749c887d04fc606bb54044e7f04e91b8f33f9be3e09f43b40`；T0-147-aligned nonpenetration/joint-limit/coupling/continuity gate PASS |
|
| 95 |
+
| Full review | Stage1 MANO / Stage2 Sharpa 2×3 SHA256 `77b6a046db26109eda44e70e5fd3ac54e1f0460803cd14a2abeaff10487b5170` |
|
| 96 |
+
|
| 97 |
+
## Auto-Curriculum and Evaluation Contract
|
| 98 |
+
|
| 99 |
+
| Field | Frozen value |
|
| 100 |
+
|---|---|
|
| 101 |
+
| Task ID | `SHARPA-AC-P09-003` |
|
| 102 |
+
| Stop rule | 每stage独立EMA `.95`；min `1500`；meaningful improvement `2%`；patience `1000`；max local `4999` |
|
| 103 |
+
| Transition | S0 selected checkpoint model-only warm start S1；fresh optimizer |
|
| 104 |
+
| GoalContract | SHA256 `5617b10a4c8b9140aeb5ab2ae350b603b5624465c63e460ac71bccbab2e0f2e7` |
|
| 105 |
+
| Evaluation | selected S1；fresh Isaac；`refaug_reset_dr`；seed `0..7`；完整trajectory + 15-step hold |
|
| 106 |
+
| Videos | 每seed一条Stage1 MANO top row / policy actual bottom row的H.264 2×3 |
|
| 107 |
+
| Publication | S0/S1 selected checkpoint Artifact；evaluation bundle；W&B API下载回读与SHA/ffprobe |
|
| 108 |
+
| P2 / Promotion | `NOT_RUN / NON_PROMOTABLE` |
|
| 109 |
+
|
| 110 |
+
## Execution Update
|
| 111 |
+
|
| 112 |
+
| Field | Value |
|
| 113 |
+
|---|---|
|
| 114 |
+
| Status | `RETRY_PLANNED / CREATE_ONLY` |
|
| 115 |
+
| Reason | P09-001未产出checkpoint；P09-002在首个formal step因W&B 0.28拒绝`demo_len=513 > 512`个histogram bins而`INFRA_FAILURE`，未产生scalar/checkpoint。两条历史均不得覆盖或复用。 |
|
| 116 |
+
| Scientific contract | Policy、observation/action、PPO、reward、randomization、Stage2 handoff、H5、seed、budget与stop rule均与P09-001相同。 |
|
| 117 |
+
| Code identity | 科学代码基线 `sharpa-main@ff687dd925b5b5b4e5944c5c470ef8e86728615b`；仅日志修复 commit `b2eef7c90fa8ff46619f5f8391e7d64f3fe07580` 将W&B sampled-timestep histogram bins限制为`min(demo_len,512)`，不改变样本、训练或评测合同。 |
|
| 118 |
+
| Claim boundary | 本Card只授权create-only gate与训练重试；checkpoint仍是候选，evaluation与人工审片未完成前不得称为可用expert。 |
|
objects/phone/trajectories/phone_09/stage4/data043-dr-rollouts-v1/checksums.json
ADDED
|
@@ -0,0 +1,8 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"archive_sha256": "c472015a1dc6705ccf7ab132daa6af0753340fb3ee8c244fe99998858c5d087b",
|
| 3 |
+
"archive_size_bytes": 1307750400,
|
| 4 |
+
"data_task_id": "SHARPA-BC-DATA-043",
|
| 5 |
+
"manifest_sha256": "08e5f98f85fec52935c70fad2ab1a8d8910f19f2973a4cf8f5b33010f65bd859",
|
| 6 |
+
"phone_id": "phone_09",
|
| 7 |
+
"schema": "sharpa_stage4_release_checksums_v1"
|
| 8 |
+
}
|
objects/phone/trajectories/phone_09/stage4/data043-dr-rollouts-v1/data_card.md
ADDED
|
@@ -0,0 +1,96 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
# Data Card: SHARPA-BC-DATA-043 / 20260927_phone_09_s1_c4096_dr_rsi_v1
|
| 2 |
+
|
| 3 |
+
| Field | Value |
|
| 4 |
+
|---|---|
|
| 5 |
+
| Data Task ID | SHARPA-BC-DATA-043 |
|
| 6 |
+
| Dataset ID | 20260927_phone_09_s1_c4096_dr_rsi_v1 |
|
| 7 |
+
| Status | READY / C4096 / INFRA0 |
|
| 8 |
+
| Purpose | 保存 phone_09 residual Teacher 在冻结 DR+RSI 分布下的 closed-loop rollout，供 Stage4 BC/Student 独立复用。 |
|
| 9 |
+
| Source mode | teacher_rollout；每个attempt独立采样DR与RSI，保存Teacher真实执行后的deployable observation和absolute applied target。 |
|
| 10 |
+
| Source lineage | SHARPA-AC-P09-003 selected model_8500.pt |
|
| 11 |
+
| Episodes / splits | terminal 4096；train 2884；validation 479；heldout 480；reflection 253；infra 0。 |
|
| 12 |
+
| Sampling | 预先冻结的DR14 + uniform RSI bank；失败保留原split身份，禁止按结果补抽。 |
|
| 13 |
+
| Observation | GRU current-frame 71D与MLP history-3 213D deployable views；不含reference、phase、Teacher ID或DR latent。 |
|
| 14 |
+
| Action / label | 31D absolute final-applied target；不是Teacher raw residual。 |
|
| 15 |
+
| Normalization | admitted train only；attempt等权、attempt内transition等权；validation/heldout transform-only；std floor `1e-8`。 |
|
| 16 |
+
| Producer | formal043_recovery_v3 finalizer |
|
| 17 |
+
| Artifact / manifest | archive 1307750400 bytes / SHA `c472015a1dc6705ccf7ab132daa6af0753340fb3ee8c244fe99998858c5d087b`；dataset manifest SHA `08e5f98f85fec52935c70fad2ab1a8d8910f19f2973a4cf8f5b33010f65bd859`。 |
|
| 18 |
+
| Claim boundary | 仅支持 phone_09、同一Teacher与同一DR+RSI collection distribution的动作蒸馏；不证明跨trajectory、跨物体、P2 robustness或sim2real。 |
|
| 19 |
+
|
| 20 |
+
## Source and Provenance
|
| 21 |
+
|
| 22 |
+
| Artifact | Identity / value | Evidence |
|
| 23 |
+
|---|---|---|
|
| 24 |
+
| Teacher lineage | SHARPA-AC-P09-003 selected model_8500.pt | 冻结selected Teacher与Stage2 identity。 |
|
| 25 |
+
| Dataset manifest | `08e5f98f85fec52935c70fad2ab1a8d8910f19f2973a4cf8f5b33010f65bd859` | READY finalizer输出；所有成员逐SHA绑定。 |
|
| 26 |
+
| Release archive | `c472015a1dc6705ccf7ab132daa6af0753340fb3ee8c244fe99998858c5d087b` / 1307750400 bytes | create-only `dataset_v1/` tar。 |
|
| 27 |
+
|
| 28 |
+
## Schema
|
| 29 |
+
|
| 30 |
+
| Field | Shape / type | Semantics |
|
| 31 |
+
|---|---|---|
|
| 32 |
+
| observation | `(A,71)`或`(A,213)` float32 | 当前帧GRU view或attempt-local history-3 MLP view。 |
|
| 33 |
+
| target | `(A,31)` float32 | wrist xyz3 + rotation6D6 + 22 absolute finger targets。 |
|
| 34 |
+
| transition_index | `(A,)` int64 | attempt-local连续控制步。 |
|
| 35 |
+
| state_t1_student_frame | `(A,71)` float32 | physics step后的next deployable state，仅用于审计。 |
|
| 36 |
+
|
| 37 |
+
## Temporal Alignment
|
| 38 |
+
|
| 39 |
+
| Item | Contract |
|
| 40 |
+
|---|---|
|
| 41 |
+
| Physics / control | PhysX 120Hz；policy/control 30Hz；decimation4。 |
|
| 42 |
+
| Transition | `obs_t → Teacher → final_applied_target_t → 4 physics substeps → state_t+1`。 |
|
| 43 |
+
| Boundaries | history和future-action window不得跨attempt、split或index gap。 |
|
| 44 |
+
|
| 45 |
+
## Simulator Environment Randomization
|
| 46 |
+
|
| 47 |
+
| Field | Value |
|
| 48 |
+
|---|---|
|
| 49 |
+
| Enabled | true |
|
| 50 |
+
| Profile | reset-only DR14 + reference augmentation + uniform RSI。 |
|
| 51 |
+
| Exact sampling authority | 发布的DR bank及dataset manifest。 |
|
| 52 |
+
|
| 53 |
+
| Category | Parameter | Distribution / range | Sampling / effect |
|
| 54 |
+
|---|---|---|---|
|
| 55 |
+
| Reference augmentation | common X/Y | each `U[-0.05,0.05] m` | 同时作用hand和phone。 |
|
| 56 |
+
| Reference augmentation | common yaw | `U[-30°,30°]` | 同时作用hand和phone。 |
|
| 57 |
+
| Reset | phone XYZ | each `U[-0.006,0.006] m` | reference/RSI选择后应用。 |
|
| 58 |
+
| Reset | phone rotation | signed angle `U[-3°,3°]` | 每次reset独立采样。 |
|
| 59 |
+
| Reset | finger wrap/rub | wrap `U[0,0.06] rad`；rub `U[-0.025,0.025] rad` | grasp-shaped reset。 |
|
| 60 |
+
| RSI | normalized phase | `U[0,1)` | 映射到合法reference start index。 |
|
| 61 |
+
|
| 62 |
+
| Option | State | Meaning |
|
| 63 |
+
|---|---|---|
|
| 64 |
+
| Physics randomization | disabled | mass、COM、inertia、friction和drive固定。 |
|
| 65 |
+
| Observation/noise/lag | disabled | clean observation；zero lag。 |
|
| 66 |
+
| Disturbance/push | disabled | none。 |
|
| 67 |
+
|
| 68 |
+
## Splits, Sampling, and Leakage
|
| 69 |
+
|
| 70 |
+
| Split | Episodes / rows | Purpose |
|
| 71 |
+
|---|---:|---|
|
| 72 |
+
| train | 2884 | 唯一optimization和normalization来源。 |
|
| 73 |
+
| validation | 479 | offline diagnostic；不参与normalization。 |
|
| 74 |
+
| heldout | 480 | same-source unseen-attempt audit。 |
|
| 75 |
+
| reflection | 253 | 失败attempt；不进入监督数据。 |
|
| 76 |
+
| infra | 0 | 任一infra都会阻止READY。 |
|
| 77 |
+
|
| 78 |
+
Approved exception：本批次使用冻结的C4096或DATA-019 PARTIAL2763拓扑，而不是默认96 train + 32 heldout。
|
| 79 |
+
|
| 80 |
+
## Normalization
|
| 81 |
+
|
| 82 |
+
| Field | Contract |
|
| 83 |
+
|---|---|
|
| 84 |
+
| Source | admitted train attempts only。 |
|
| 85 |
+
| Weighting | attempt等权；attempt内transition等权。 |
|
| 86 |
+
| Validation / heldout | train statistics transform-only。 |
|
| 87 |
+
|
| 88 |
+
## Validation and Readiness
|
| 89 |
+
|
| 90 |
+
| Gate | Result | Evidence |
|
| 91 |
+
|---|---|---|
|
| 92 |
+
| Completeness | PASS | terminal 4096；infra 0。 |
|
| 93 |
+
| Split leakage | PASS | 原split身份保留；reflection不进入训练。 |
|
| 94 |
+
| Member identity | PASS | manifest `08e5f98f85fec52935c70fad2ab1a8d8910f19f2973a4cf8f5b33010f65bd859`。 |
|
| 95 |
+
| Release identity | PASS | archive `c472015a1dc6705ccf7ab132daa6af0753340fb3ee8c244fe99998858c5d087b`。 |
|
| 96 |
+
| Dataset status | READY | READY / C4096 / INFRA0。 |
|
objects/phone/trajectories/phone_09/stage4/data043-dr-rollouts-v1/dataset_manifest.json
ADDED
|
The diff for this file is too large to render.
See raw diff
|
|
|
objects/phone/trajectories/phone_09/stage4/data043-dr-rollouts-v1/distribution_calibration.json
ADDED
|
@@ -0,0 +1,120 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"counts": {
|
| 3 |
+
"all": 4096,
|
| 4 |
+
"heldout": 512,
|
| 5 |
+
"train": 3072,
|
| 6 |
+
"validation": 512
|
| 7 |
+
},
|
| 8 |
+
"dimensions": [
|
| 9 |
+
{
|
| 10 |
+
"index": 0,
|
| 11 |
+
"max": 0.9997870242223144,
|
| 12 |
+
"mean": 0.4999999995343387,
|
| 13 |
+
"min": 7.827673107385635e-05
|
| 14 |
+
},
|
| 15 |
+
{
|
| 16 |
+
"index": 1,
|
| 17 |
+
"max": 0.9999859221279621,
|
| 18 |
+
"mean": 0.4999999995343387,
|
| 19 |
+
"min": 2.716667950153351e-06
|
| 20 |
+
},
|
| 21 |
+
{
|
| 22 |
+
"index": 2,
|
| 23 |
+
"max": 0.9998920541256666,
|
| 24 |
+
"mean": 0.4999999995343387,
|
| 25 |
+
"min": 1.5076249837875366e-05
|
| 26 |
+
},
|
| 27 |
+
{
|
| 28 |
+
"index": 3,
|
| 29 |
+
"max": 0.9999815504997969,
|
| 30 |
+
"mean": 0.4999999995343387,
|
| 31 |
+
"min": 2.8776004910469055e-05
|
| 32 |
+
},
|
| 33 |
+
{
|
| 34 |
+
"index": 4,
|
| 35 |
+
"max": 0.9999641375616193,
|
| 36 |
+
"mean": 0.4999999995343387,
|
| 37 |
+
"min": 4.5709311962127686e-05
|
| 38 |
+
},
|
| 39 |
+
{
|
| 40 |
+
"index": 5,
|
| 41 |
+
"max": 0.9999046549201012,
|
| 42 |
+
"mean": 0.4999999995343387,
|
| 43 |
+
"min": 0.00011997856199741364
|
| 44 |
+
},
|
| 45 |
+
{
|
| 46 |
+
"index": 6,
|
| 47 |
+
"max": 0.9999609068036079,
|
| 48 |
+
"mean": 0.4999999995343387,
|
| 49 |
+
"min": 0.00023153051733970642
|
| 50 |
+
},
|
| 51 |
+
{
|
| 52 |
+
"index": 7,
|
| 53 |
+
"max": 0.9998109750449657,
|
| 54 |
+
"mean": 0.4999999995343387,
|
| 55 |
+
"min": 0.00014089234173297882
|
| 56 |
+
},
|
| 57 |
+
{
|
| 58 |
+
"index": 8,
|
| 59 |
+
"max": 0.9999717762693763,
|
| 60 |
+
"mean": 0.4999999995343387,
|
| 61 |
+
"min": 0.00015771202743053436
|
| 62 |
+
},
|
| 63 |
+
{
|
| 64 |
+
"index": 9,
|
| 65 |
+
"max": 0.999759191647172,
|
| 66 |
+
"mean": 0.4999999995343387,
|
| 67 |
+
"min": 0.00020562484860420227
|
| 68 |
+
},
|
| 69 |
+
{
|
| 70 |
+
"index": 10,
|
| 71 |
+
"max": 0.9998323954641819,
|
| 72 |
+
"mean": 0.4999999995343387,
|
| 73 |
+
"min": 2.3738481104373932e-05
|
| 74 |
+
},
|
| 75 |
+
{
|
| 76 |
+
"index": 11,
|
| 77 |
+
"max": 0.9999181777238846,
|
| 78 |
+
"mean": 0.4999999995343387,
|
| 79 |
+
"min": 0.00013154372572898865
|
| 80 |
+
},
|
| 81 |
+
{
|
| 82 |
+
"index": 12,
|
| 83 |
+
"max": 0.9997697519138455,
|
| 84 |
+
"mean": 0.4999999995343387,
|
| 85 |
+
"min": 5.544256418943405e-05
|
| 86 |
+
},
|
| 87 |
+
{
|
| 88 |
+
"index": 13,
|
| 89 |
+
"max": 0.9999812860041857,
|
| 90 |
+
"mean": 0.4999999995343387,
|
| 91 |
+
"min": 0.0001823054626584053
|
| 92 |
+
},
|
| 93 |
+
{
|
| 94 |
+
"index": 14,
|
| 95 |
+
"max": 0.999907054938376,
|
| 96 |
+
"mean": 0.4999999995343387,
|
| 97 |
+
"min": 0.00018940307199954987
|
| 98 |
+
}
|
| 99 |
+
],
|
| 100 |
+
"gates": {
|
| 101 |
+
"G10_dimension_means": "PASS",
|
| 102 |
+
"G11_rsi_full_discrete_support": "PASS",
|
| 103 |
+
"G12_action_state_contract": "PASS",
|
| 104 |
+
"G1_exact_cardinality": "PASS",
|
| 105 |
+
"G2_dimension_15d": "PASS",
|
| 106 |
+
"G3_finite": "PASS",
|
| 107 |
+
"G4_unit_cube": "PASS",
|
| 108 |
+
"G5_unique_case_ids": "PASS",
|
| 109 |
+
"G6_unique_case_seeds": "PASS",
|
| 110 |
+
"G7_unique_dr_seeds": "PASS",
|
| 111 |
+
"G8_split_membership_frozen": "PASS",
|
| 112 |
+
"G9_dimension_extrema": "PASS"
|
| 113 |
+
},
|
| 114 |
+
"rsi_start_phase": {
|
| 115 |
+
"max": 511,
|
| 116 |
+
"min": 0
|
| 117 |
+
},
|
| 118 |
+
"schema": "sharpa_coldclone_formal_distribution_calibration_v1",
|
| 119 |
+
"status": "PASS"
|
| 120 |
+
}
|
objects/phone/trajectories/phone_09/stage4/data043-dr-rollouts-v1/dr_bank.json
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:5236bb4ef60101dac15b523a0780411398b8b64046e9d161979a8dcf158a256e
|
| 3 |
+
size 13929154
|
objects/phone/trajectories/phone_10/stage4/data044-dr-rollouts-v1/checksums.json
ADDED
|
@@ -0,0 +1,8 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"archive_sha256": "05322cf52ee41a14b558660da4c3baa0d1f78e8fdd2525bae6914a31d34fc5e0",
|
| 3 |
+
"archive_size_bytes": 1308344320,
|
| 4 |
+
"data_task_id": "SHARPA-BC-DATA-044",
|
| 5 |
+
"manifest_sha256": "c7cac9e30db290c0751bb43948958191007cb34eba015cf4f6ef96fd7bf8cae2",
|
| 6 |
+
"phone_id": "phone_10",
|
| 7 |
+
"schema": "sharpa_stage4_release_checksums_v1"
|
| 8 |
+
}
|
objects/phone/trajectories/phone_10/stage4/data044-dr-rollouts-v1/data_card.md
ADDED
|
@@ -0,0 +1,96 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
# Data Card: SHARPA-BC-DATA-044 / 20260927_phone_10_s1_c4096_dr_rsi_v1
|
| 2 |
+
|
| 3 |
+
| Field | Value |
|
| 4 |
+
|---|---|
|
| 5 |
+
| Data Task ID | SHARPA-BC-DATA-044 |
|
| 6 |
+
| Dataset ID | 20260927_phone_10_s1_c4096_dr_rsi_v1 |
|
| 7 |
+
| Status | READY / C4096 / INFRA0 |
|
| 8 |
+
| Purpose | 保存 phone_10 residual Teacher 在冻结 DR+RSI 分布下的 closed-loop rollout，供 Stage4 BC/Student 独立复用。 |
|
| 9 |
+
| Source mode | teacher_rollout；每个attempt独立采样DR与RSI，保存Teacher真实执行后的deployable observation和absolute applied target。 |
|
| 10 |
+
| Source lineage | SHARPA-AC-P10-001 selected model_9000.pt |
|
| 11 |
+
| Episodes / splits | terminal 4096；train 2693；validation 447；heldout 445；reflection 511；infra 0。 |
|
| 12 |
+
| Sampling | 预先冻结的DR14 + uniform RSI bank；失败保留原split身份，禁止按结果补抽。 |
|
| 13 |
+
| Observation | GRU current-frame 71D与MLP history-3 213D deployable views；不含reference、phase、Teacher ID或DR latent。 |
|
| 14 |
+
| Action / label | 31D absolute final-applied target；不是Teacher raw residual。 |
|
| 15 |
+
| Normalization | admitted train only；attempt等权、attempt内transition等权；validation/heldout transform-only；std floor `1e-8`。 |
|
| 16 |
+
| Producer | formal044_recovery_v3 finalizer |
|
| 17 |
+
| Artifact / manifest | archive 1308344320 bytes / SHA `05322cf52ee41a14b558660da4c3baa0d1f78e8fdd2525bae6914a31d34fc5e0`；dataset manifest SHA `c7cac9e30db290c0751bb43948958191007cb34eba015cf4f6ef96fd7bf8cae2`。 |
|
| 18 |
+
| Claim boundary | 仅支持 phone_10、同一Teacher与同一DR+RSI collection distribution的动作蒸馏；不证明跨trajectory、跨物体、P2 robustness或sim2real。 |
|
| 19 |
+
|
| 20 |
+
## Source and Provenance
|
| 21 |
+
|
| 22 |
+
| Artifact | Identity / value | Evidence |
|
| 23 |
+
|---|---|---|
|
| 24 |
+
| Teacher lineage | SHARPA-AC-P10-001 selected model_9000.pt | 冻结selected Teacher与Stage2 identity。 |
|
| 25 |
+
| Dataset manifest | `c7cac9e30db290c0751bb43948958191007cb34eba015cf4f6ef96fd7bf8cae2` | READY finalizer输出；所有成员逐SHA绑定。 |
|
| 26 |
+
| Release archive | `05322cf52ee41a14b558660da4c3baa0d1f78e8fdd2525bae6914a31d34fc5e0` / 1308344320 bytes | create-only `dataset_v1/` tar。 |
|
| 27 |
+
|
| 28 |
+
## Schema
|
| 29 |
+
|
| 30 |
+
| Field | Shape / type | Semantics |
|
| 31 |
+
|---|---|---|
|
| 32 |
+
| observation | `(A,71)`或`(A,213)` float32 | 当前帧GRU view或attempt-local history-3 MLP view。 |
|
| 33 |
+
| target | `(A,31)` float32 | wrist xyz3 + rotation6D6 + 22 absolute finger targets。 |
|
| 34 |
+
| transition_index | `(A,)` int64 | attempt-local连续控制步。 |
|
| 35 |
+
| state_t1_student_frame | `(A,71)` float32 | physics step后的next deployable state，仅用于审计。 |
|
| 36 |
+
|
| 37 |
+
## Temporal Alignment
|
| 38 |
+
|
| 39 |
+
| Item | Contract |
|
| 40 |
+
|---|---|
|
| 41 |
+
| Physics / control | PhysX 120Hz；policy/control 30Hz；decimation4。 |
|
| 42 |
+
| Transition | `obs_t → Teacher → final_applied_target_t → 4 physics substeps → state_t+1`。 |
|
| 43 |
+
| Boundaries | history和future-action window不得跨attempt、split或index gap。 |
|
| 44 |
+
|
| 45 |
+
## Simulator Environment Randomization
|
| 46 |
+
|
| 47 |
+
| Field | Value |
|
| 48 |
+
|---|---|
|
| 49 |
+
| Enabled | true |
|
| 50 |
+
| Profile | reset-only DR14 + reference augmentation + uniform RSI。 |
|
| 51 |
+
| Exact sampling authority | 发布的DR bank及dataset manifest。 |
|
| 52 |
+
|
| 53 |
+
| Category | Parameter | Distribution / range | Sampling / effect |
|
| 54 |
+
|---|---|---|---|
|
| 55 |
+
| Reference augmentation | common X/Y | each `U[-0.05,0.05] m` | 同时作用hand和phone。 |
|
| 56 |
+
| Reference augmentation | common yaw | `U[-30°,30°]` | 同时作用hand和phone。 |
|
| 57 |
+
| Reset | phone XYZ | each `U[-0.006,0.006] m` | reference/RSI选择后应用。 |
|
| 58 |
+
| Reset | phone rotation | signed angle `U[-3°,3°]` | 每次reset独立采样。 |
|
| 59 |
+
| Reset | finger wrap/rub | wrap `U[0,0.06] rad`；rub `U[-0.025,0.025] rad` | grasp-shaped reset。 |
|
| 60 |
+
| RSI | normalized phase | `U[0,1)` | 映射到合法reference start index。 |
|
| 61 |
+
|
| 62 |
+
| Option | State | Meaning |
|
| 63 |
+
|---|---|---|
|
| 64 |
+
| Physics randomization | disabled | mass、COM、inertia、friction和drive固定。 |
|
| 65 |
+
| Observation/noise/lag | disabled | clean observation；zero lag。 |
|
| 66 |
+
| Disturbance/push | disabled | none。 |
|
| 67 |
+
|
| 68 |
+
## Splits, Sampling, and Leakage
|
| 69 |
+
|
| 70 |
+
| Split | Episodes / rows | Purpose |
|
| 71 |
+
|---|---:|---|
|
| 72 |
+
| train | 2693 | 唯一optimization和normalization来源。 |
|
| 73 |
+
| validation | 447 | offline diagnostic；不参与normalization。 |
|
| 74 |
+
| heldout | 445 | same-source unseen-attempt audit。 |
|
| 75 |
+
| reflection | 511 | 失败attempt；不进入监督数据。 |
|
| 76 |
+
| infra | 0 | 任一infra都会阻止READY。 |
|
| 77 |
+
|
| 78 |
+
Approved exception：本批次使用冻结的C4096或DATA-019 PARTIAL2763拓扑，而不是默认96 train + 32 heldout。
|
| 79 |
+
|
| 80 |
+
## Normalization
|
| 81 |
+
|
| 82 |
+
| Field | Contract |
|
| 83 |
+
|---|---|
|
| 84 |
+
| Source | admitted train attempts only。 |
|
| 85 |
+
| Weighting | attempt等权；attempt内transition等权。 |
|
| 86 |
+
| Validation / heldout | train statistics transform-only。 |
|
| 87 |
+
|
| 88 |
+
## Validation and Readiness
|
| 89 |
+
|
| 90 |
+
| Gate | Result | Evidence |
|
| 91 |
+
|---|---|---|
|
| 92 |
+
| Completeness | PASS | terminal 4096；infra 0。 |
|
| 93 |
+
| Split leakage | PASS | 原split身份保留；reflection不进入训练。 |
|
| 94 |
+
| Member identity | PASS | manifest `c7cac9e30db290c0751bb43948958191007cb34eba015cf4f6ef96fd7bf8cae2`。 |
|
| 95 |
+
| Release identity | PASS | archive `05322cf52ee41a14b558660da4c3baa0d1f78e8fdd2525bae6914a31d34fc5e0`。 |
|
| 96 |
+
| Dataset status | READY | READY / C4096 / INFRA0。 |
|
objects/phone/trajectories/phone_10/stage4/data044-dr-rollouts-v1/dataset_manifest.json
ADDED
|
The diff for this file is too large to render.
See raw diff
|
|
|
objects/phone/trajectories/phone_10/stage4/data044-dr-rollouts-v1/distribution_calibration.json
ADDED
|
@@ -0,0 +1,120 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"counts": {
|
| 3 |
+
"all": 4096,
|
| 4 |
+
"heldout": 512,
|
| 5 |
+
"train": 3072,
|
| 6 |
+
"validation": 512
|
| 7 |
+
},
|
| 8 |
+
"dimensions": [
|
| 9 |
+
{
|
| 10 |
+
"index": 0,
|
| 11 |
+
"max": 0.9999933401122689,
|
| 12 |
+
"mean": 0.4999999995343387,
|
| 13 |
+
"min": 0.0001247953623533249
|
| 14 |
+
},
|
| 15 |
+
{
|
| 16 |
+
"index": 1,
|
| 17 |
+
"max": 0.9999042144045234,
|
| 18 |
+
"mean": 0.4999999995343387,
|
| 19 |
+
"min": 0.00020977947860956192
|
| 20 |
+
},
|
| 21 |
+
{
|
| 22 |
+
"index": 2,
|
| 23 |
+
"max": 0.9997602133080363,
|
| 24 |
+
"mean": 0.4999999995343387,
|
| 25 |
+
"min": 5.2724964916706085e-05
|
| 26 |
+
},
|
| 27 |
+
{
|
| 28 |
+
"index": 3,
|
| 29 |
+
"max": 0.9998234314844012,
|
| 30 |
+
"mean": 0.4999999995343387,
|
| 31 |
+
"min": 0.00018991343677043915
|
| 32 |
+
},
|
| 33 |
+
{
|
| 34 |
+
"index": 4,
|
| 35 |
+
"max": 0.9999875612556934,
|
| 36 |
+
"mean": 0.4999999995343387,
|
| 37 |
+
"min": 0.0002152658998966217
|
| 38 |
+
},
|
| 39 |
+
{
|
| 40 |
+
"index": 5,
|
| 41 |
+
"max": 0.9998151753097773,
|
| 42 |
+
"mean": 0.4999999995343387,
|
| 43 |
+
"min": 0.00019729789346456528
|
| 44 |
+
},
|
| 45 |
+
{
|
| 46 |
+
"index": 6,
|
| 47 |
+
"max": 0.9999903086572886,
|
| 48 |
+
"mean": 0.4999999995343387,
|
| 49 |
+
"min": 9.817536920309067e-05
|
| 50 |
+
},
|
| 51 |
+
{
|
| 52 |
+
"index": 7,
|
| 53 |
+
"max": 0.9997988818213344,
|
| 54 |
+
"mean": 0.4999999995343387,
|
| 55 |
+
"min": 0.00014535430818796158
|
| 56 |
+
},
|
| 57 |
+
{
|
| 58 |
+
"index": 8,
|
| 59 |
+
"max": 0.9997803028672934,
|
| 60 |
+
"mean": 0.4999999995343387,
|
| 61 |
+
"min": 6.829667836427689e-05
|
| 62 |
+
},
|
| 63 |
+
{
|
| 64 |
+
"index": 9,
|
| 65 |
+
"max": 0.9998542927205563,
|
| 66 |
+
"mean": 0.4999999995343387,
|
| 67 |
+
"min": 0.00012292712926864624
|
| 68 |
+
},
|
| 69 |
+
{
|
| 70 |
+
"index": 10,
|
| 71 |
+
"max": 0.9998900685459375,
|
| 72 |
+
"mean": 0.4999999995343387,
|
| 73 |
+
"min": 0.00016637425869703293
|
| 74 |
+
},
|
| 75 |
+
{
|
| 76 |
+
"index": 11,
|
| 77 |
+
"max": 0.9999030902981758,
|
| 78 |
+
"mean": 0.4999999995343387,
|
| 79 |
+
"min": 4.359148442745209e-05
|
| 80 |
+
},
|
| 81 |
+
{
|
| 82 |
+
"index": 12,
|
| 83 |
+
"max": 0.9999511381611228,
|
| 84 |
+
"mean": 0.4999999995343387,
|
| 85 |
+
"min": 3.752671182155609e-05
|
| 86 |
+
},
|
| 87 |
+
{
|
| 88 |
+
"index": 13,
|
| 89 |
+
"max": 0.9998934250324965,
|
| 90 |
+
"mean": 0.4999999995343387,
|
| 91 |
+
"min": 0.00015564821660518646
|
| 92 |
+
},
|
| 93 |
+
{
|
| 94 |
+
"index": 14,
|
| 95 |
+
"max": 0.9999322220683098,
|
| 96 |
+
"mean": 0.4999999995343387,
|
| 97 |
+
"min": 1.5500932931900024e-05
|
| 98 |
+
}
|
| 99 |
+
],
|
| 100 |
+
"gates": {
|
| 101 |
+
"G10_dimension_means": "PASS",
|
| 102 |
+
"G11_rsi_full_discrete_support": "PASS",
|
| 103 |
+
"G12_action_state_contract": "PASS",
|
| 104 |
+
"G1_exact_cardinality": "PASS",
|
| 105 |
+
"G2_dimension_15d": "PASS",
|
| 106 |
+
"G3_finite": "PASS",
|
| 107 |
+
"G4_unit_cube": "PASS",
|
| 108 |
+
"G5_unique_case_ids": "PASS",
|
| 109 |
+
"G6_unique_case_seeds": "PASS",
|
| 110 |
+
"G7_unique_dr_seeds": "PASS",
|
| 111 |
+
"G8_split_membership_frozen": "PASS",
|
| 112 |
+
"G9_dimension_extrema": "PASS"
|
| 113 |
+
},
|
| 114 |
+
"rsi_start_phase": {
|
| 115 |
+
"max": 510,
|
| 116 |
+
"min": 0
|
| 117 |
+
},
|
| 118 |
+
"schema": "sharpa_coldclone_formal_distribution_calibration_v1",
|
| 119 |
+
"status": "PASS"
|
| 120 |
+
}
|
objects/phone/trajectories/phone_10/stage4/data044-dr-rollouts-v1/dr_bank.json
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:7e853bb4fb4c08f2676c8d1e6a0383a6d24c239235a566d8aa27f7a9a5b00956
|
| 3 |
+
size 13929328
|
objects/phone/trajectories/phone_14/stage4/data033-dr-rollouts-v1/checksums.json
ADDED
|
@@ -0,0 +1,8 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"archive_sha256": "f27c1c269f33b85127d0ae3a6b173c1188f1b5e44d2fb136ab09d9c864aa4f1c",
|
| 3 |
+
"archive_size_bytes": 1039912960,
|
| 4 |
+
"data_task_id": "SHARPA-BC-DATA-033",
|
| 5 |
+
"manifest_sha256": "1237116a7ed7f004fe0ed6d74edacf178089371f438a6147161d7d564af760d5",
|
| 6 |
+
"phone_id": "phone_14",
|
| 7 |
+
"schema": "sharpa_stage4_release_checksums_v1"
|
| 8 |
+
}
|
objects/phone/trajectories/phone_14/stage4/data033-dr-rollouts-v1/data_card.md
ADDED
|
@@ -0,0 +1,139 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
# Data Card: SHARPA-BC-DATA-033 / 20260903_sharpa_ac_p14_001_phone_14_c4096_dr_rsi_v1
|
| 2 |
+
|
| 3 |
+
| Field | Value |
|
| 4 |
+
|---|---|
|
| 5 |
+
| Data Task ID | SHARPA-BC-DATA-033 |
|
| 6 |
+
| Dataset ID | 20260903_sharpa_ac_p14_001_phone_14_c4096_dr_rsi_v1 |
|
| 7 |
+
| Status | READY / C4096 / INFRA0 |
|
| 8 |
+
| Purpose | 从phone_14 `SHARPA-AC-P14-001` S1 selected checkpoint采集4096个与Teacher训练时RefAug/reset-only DR及uniform RSI分布一致的独立closed-loop rollouts，形成3072 train、512 validation、512 sealed heldout。 |
|
| 9 |
+
| Source mode | `teacher_rollout`；fresh Isaac/PhysX；每个attempt独立采样DR与RSI，从采样phase执行到reference terminal再加15 control-step hold。 |
|
| 10 |
+
| Source lineage | `SHARPA-AC-P14-001`；W&B S1 selected `model_8500.pt`；原始8-seed evaluation仅为diagnostic，P2=`NOT_RUN`。 |
|
| 11 |
+
| Episodes / splits | 4096 preregistered attempts=`3072 train + 512 validation + 512 sealed heldout`；失败保留原split身份并进入reflection/infra，禁止按结果补抽或搬移。 |
|
| 12 |
+
| Sampling | 15D=`DR14 + uniform RSI phase`；每row独立采样；先物化、校准并冻结4096-row bank，之后才能launch。 |
|
| 13 |
+
| Observation | 71D deployable current frame及attempt-local history-3 213D view；不含velocity、phase/reference、Teacher ID、DR latent或future state。 |
|
| 14 |
+
| Action / label | controller真实执行的31D absolute `final_applied_target`；Teacher 28D residual仅作diagnostic，不能作为Student label。 |
|
| 15 |
+
| Normalization | admitted train only；attempt等权、attempt内transition等权；validation/heldout transform-only；std floor `1e-8`。 |
|
| 16 |
+
| Producer | formal producer commit `fc39d2f03df50e8888e2fd1bc00afbfc1725f1e5`；executable spec `outputs/formal033_v3/spec.json`。 |
|
| 17 |
+
| Artifact / manifest | create-only `outputs/formal033_v3`；bank SHA `23504867795883852ac5acddd4e690ac4c602821933e895772e77fbd070029e5`；dataset manifest待4096 receipts后生成。 |
|
| 18 |
+
| Claim boundary | 仅支持phone_14、同一Teacher、同RefAug/reset-only DR+uniform-RSI分布的动作蒸馏研究；不支持physics/noise/push、跨episode、跨物体、P2 robustness或真机泛化。 |
|
| 19 |
+
|
| 20 |
+
## Source and Provenance
|
| 21 |
+
|
| 22 |
+
| Artifact | Identity / value | Evidence |
|
| 23 |
+
|---|---|---|
|
| 24 |
+
| Teacher Task | `SHARPA-AC-P14-001` | Auto-Curriculum S1 RefAug/reset-only DR candidate。 |
|
| 25 |
+
| Teacher Training Card | `teacher/checkpoints/sharpa_phone20_autocurriculum_selected_v1/phone_14/inputs/training_card.md`；SHA `74c2259856114f7453b8efc1cabf8c923ff4e0bf04e19868493235fc762160d6` | W&B selected-checkpoint artifact原样下载。 |
|
| 26 |
+
| Selected checkpoint | `teacher/checkpoints/sharpa_phone20_autocurriculum_selected_v1/phone_14/model_8500.pt`；`model_8500.pt`；iteration `8500`；SHA `644117cb1a26eb145e5598a8b3148fd0a2261c02a9e2bf728f414a5ee57684fa` | W&B artifact `dfnet/dex-retarget/sharpa-phone_14-autocurriculum-s1-v1-s1-selected-checkpoint:v0`；local SHA gate PASS。 |
|
| 27 |
+
| Teacher code | commit `4a146fe07ef7f421745c3506ca803594f39b4fc8` | checkpoint selection manifest。 |
|
| 28 |
+
| Runtime manifest | `teacher/checkpoints/sharpa_phone20_autocurriculum_selected_v1/phone_14/inputs/runtime_manifest.json`；SHA `edb7361a4e2788aff9966a36e09132755f64c3318ebde086c93026d6c9f1af9e` | 仅保存manifest；完整可执行runtime仍须在G13物化并回读。 |
|
| 29 |
+
| Curriculum manifest | `teacher/checkpoints/sharpa_phone20_autocurriculum_selected_v1/phone_14/inputs/curriculum_manifest.json`；SHA `41ee49343ecd9c957b79823297e5f13280c66d22ee4361a598213cb8f7d04a52` | S0 nominal → S1 RefAug/reset-only DR。 |
|
| 30 |
+
| Stage2 handoff | `teacher/checkpoints/sharpa_phone20_autocurriculum_selected_v1/phone_14/inputs/stage2_handoff.json`；SHA `1ac48397b8a998847567fbbe2b63c7e4ca47a898c796c27680a0ba1e01726fc0`；status `STAGE2_HANDOFF_ACCEPTED` | W&B artifact原样下载。 |
|
| 31 |
+
| Stage1 NPZ identity | SHA `23e3c361420a5bc791ba9985d0b4a9123d6035d1c4684aaf1e6cf1003b8e6137` | bytes尚未随本checkpoint bundle下载；G13前必须解析。 |
|
| 32 |
+
| Human H5 identity | SHA `70c630916707dc64be6cd16fa8118f27047c2564e225c44affe49d62d258f67c` | `761` frames at `60.0` Hz。 |
|
| 33 |
+
| Runtime retarget H5 identity | SHA `60930f5f686c41ab295ae15ea43f566d9fe6e3da47e02435349eb6e6756177d3` | `380` nominal policy targets at `30.0` Hz。 |
|
| 34 |
+
| Reference manifest identity | SHA `844ebeb308da5f95ba9d4c349f3fd3ddc57cae538463deea119d977ed58cf770` | Stage2 accepted reference。 |
|
| 35 |
+
| GoalContract identity | SHA `e56d52dfeec1eefa341981543ce0f4b3a6dc25e1b01179f00bf12383aaf6e14d` | runtime artifact manifest。 |
|
| 36 |
+
| Historical evaluation | [W&B seed0–7 diagnostic](https://wandb.ai/dfnet/dex-retarget/runs/sharpa-phone_14-autocurriculum-evaluation-seed0-7-v1) | checkpoint identity matched；P2 `NOT_RUN / NON_PROMOTABLE`。 |
|
| 37 |
+
|
| 38 |
+
## Schema
|
| 39 |
+
|
| 40 |
+
| Field | Shape / type | Semantics |
|
| 41 |
+
|---|---|---|
|
| 42 |
+
| `time_s` | `(S,) float64` | 30 Hz control timestamps from the sampled RSI phase。 |
|
| 43 |
+
| `actual_q` | `(S,22) float32`, rad | PhysX actual Sharpa finger joints。 |
|
| 44 |
+
| `actual_wrist_pose_wxyz` | `(S,7) float32`, world frame | actual wrist translation and quaternion。 |
|
| 45 |
+
| `actual_object_pose_wxyz` | `(S,7) float32`, world frame | actual phone translation and quaternion。 |
|
| 46 |
+
| `raw_residual` / `processed_residual` | `(A,28) float32` | Teacher diagnostic action before/after adapter processing。 |
|
| 47 |
+
| `final_applied_target_wxyz` | `(A,29) float32` | runtime controller target representation。 |
|
| 48 |
+
| `student_frame` | `(A,71) float32` | deployable current-frame observation。 |
|
| 49 |
+
| `applied_target_31` | `(A,31) float32` | unique BC supervision label: wrist position3 + rotation6D6 + joint22。 |
|
| 50 |
+
| `state_t1_student_frame` | `(A,71) float32` | post-physics next state，audit-only。 |
|
| 51 |
+
| `pose_history3_mlp_v1` | `(A,213) float32` | three attempt-local 71D frames，oldest→current；first-frame repeat padding。 |
|
| 52 |
+
| `normalized_sampling_15d` | `(15,) float64` | DR14+RSI provenance-only；不得进入policy input。 |
|
| 53 |
+
| reset/failure/action ledgers | JSON | sampled parameters、actual runtime readback、drop/invalid/infra与action-chain evidence。 |
|
| 54 |
+
|
| 55 |
+
## Temporal Alignment
|
| 56 |
+
|
| 57 |
+
| Item | Contract |
|
| 58 |
+
|---|---|
|
| 59 |
+
| Rates | PhysX 120 Hz；policy/control 30 Hz；decimation=4。 |
|
| 60 |
+
| Transition | `obs_t → Teacher residual_t → controller final_applied_target_t → four PhysX substeps → state_t+1`。 |
|
| 61 |
+
| RSI | `u_phase ~ U[0,1)`；`start=floor(u_phase × 380)`，clamp到`0…379`。 |
|
| 62 |
+
| Horizon | `A=(380-start)+15`；`S=A+1`；即使drop/invalid也记录完整终态与ledger。 |
|
| 63 |
+
| History | strictly attempt-local；不得跨attempt、split或index gap。 |
|
| 64 |
+
| Future windows | 后续Consumer可构造`T_p` window，但window不得跨attempt；tail必须masked。 |
|
| 65 |
+
|
| 66 |
+
## Simulator Environment Randomization
|
| 67 |
+
|
| 68 |
+
| Field | Value |
|
| 69 |
+
|---|---|
|
| 70 |
+
| Enabled | true |
|
| 71 |
+
| Profile | `SHARPA-AC-P14-001` S1 RefAug/reset-only DR + uniform RSI。 |
|
| 72 |
+
| Source config SHA | `d99efe6b27a179211fec8f3077e5605f9d804c666f0ac9a16c3e8349795982b8`。 |
|
| 73 |
+
| Activation | every reset；nominal fraction `0.0`。 |
|
| 74 |
+
| Mask | common rigid transform同时作用hand与phone；relative reset保持grasp-shaped范围。 |
|
| 75 |
+
| Seed | 4096-row bank seed/namespace在distribution calibration阶段冻结；当前`NOT_YET_MATERIALIZED`。 |
|
| 76 |
+
|
| 77 |
+
| Category | Parameter | Distribution / range | Sampling / effect |
|
| 78 |
+
|---|---|---|---|
|
| 79 |
+
| Reference augmentation | common X/Y | each `U[-0.05,0.05] m` | 同时作用hand和phone。 |
|
| 80 |
+
| Reference augmentation | common yaw | `U[-30°,30°]` | 同时作用hand和phone。 |
|
| 81 |
+
| Relative reset | phone XYZ | each `U[-0.006,0.006] m` | after RSI/reference selection。 |
|
| 82 |
+
| Relative reset | phone rotation | axis uniformly sampled；signed angle `U[-3°,3°]` | every reset。 |
|
| 83 |
+
| Finger reset | shared wrap | `U[0,0.06] rad` | coupled flexion joints，只增加包握。 |
|
| 84 |
+
| Finger reset | four-finger rub | each `U[-0.025,0.025] rad` | independent rub joints。 |
|
| 85 |
+
| RSI | normalized phase | `U[0,1)` | maps to discrete start phase。 |
|
| 86 |
+
|
| 87 |
+
### Explicitly disabled / not applicable
|
| 88 |
+
|
| 89 |
+
| Option | State | Meaning |
|
| 90 |
+
|---|---|---|
|
| 91 |
+
| Physics/actuator randomization | disabled | mass、COM、inertia、friction与drive固定。 |
|
| 92 |
+
| Observation/noise/lag | disabled | clean observation；zero lag。 |
|
| 93 |
+
| Disturbance/push | disabled | none。 |
|
| 94 |
+
| Gravity curriculum | disabled | fixed full gravity `-9.81 m/s²`。 |
|
| 95 |
+
|
| 96 |
+
## Splits, Sampling, and Leakage
|
| 97 |
+
|
| 98 |
+
Approved Exception: Shengtao于2026-09-03批准每个独立Teacher使用4096 attempts并划分为3072 train、512 validation、512 sealed heldout，明确偏离默认96 train + 32 heldout topology。
|
| 99 |
+
|
| 100 |
+
| Split | Episodes / rows | Purpose |
|
| 101 |
+
|---|---:|---|
|
| 102 |
+
| train | 3072 | admitted rows的唯一optimization与normalization来源。 |
|
| 103 |
+
| validation | 512 | offline checkpoint selection；不得参与normalization。 |
|
| 104 |
+
| sealed heldout | 512 | 所有模型选择冻结后的同分布unseen-attempt audit。 |
|
| 105 |
+
| reflection | E5/E6/E9失败 | 保留预注册split身份；不得进入监督数据或被成功补抽替换。 |
|
| 106 |
+
| infra_failure | identity/runtime/schema failure | 阻止Dataset READY；不得解释成policy failure。 |
|
| 107 |
+
|
| 108 |
+
Split membership必须在rollout前由case ID确定。同一seed、trajectory window、parent/derived copy不得跨split；heldout不得反向用于collector、normalization、checkpoint、模型或超参数选择。
|
| 109 |
+
|
| 110 |
+
## Normalization
|
| 111 |
+
|
| 112 |
+
| Field | Contract |
|
| 113 |
+
|---|---|
|
| 114 |
+
| Source | admitted train attempts only。 |
|
| 115 |
+
| Weighting | each rollout equal；within each rollout，each transition equal。 |
|
| 116 |
+
| Stored arrays | MLP213、GRU71与target31 mean/second moment/std，以及exact train attempt IDs和weights。 |
|
| 117 |
+
| Floor | `std=max(std,1e-8)`。 |
|
| 118 |
+
| Validation / heldout | train statistics transform-only；禁止重算。 |
|
| 119 |
+
|
| 120 |
+
## Validation and Readiness
|
| 121 |
+
|
| 122 |
+
| Gate | Result | Evidence |
|
| 123 |
+
|---|---|---|
|
| 124 |
+
| Data Card format | PASS | unified validator对本批次7张Card报告`validated=7 failures=0`。 |
|
| 125 |
+
| Selected checkpoint identity | PASS | W&B selected artifact与local SHA一致：`644117cb1a26eb145e5598a8b3148fd0a2261c02a9e2bf728f414a5ee57684fa`。 |
|
| 126 |
+
| Stage2/H5/runtime resolvability | PASS | `outputs/prepared_v1`中clone-local H5/assets/runtime逐SHA解析通过。 |
|
| 127 |
+
| G1–G12 distribution calibration | PASS | `outputs/formal033_v3/distribution_calibration.json`；SHA `a1bc11f156d3007d7c7b8a9dfa8ff70ed10e2e8a798799baa5113149b18b11d7`。 |
|
| 128 |
+
| G13 identity/runtime | PASS | producer `fc39d2f`；checkpoint、runtime、artifact、goal与spec/bank SHA preflight通过。 |
|
| 129 |
+
| G14 one-attempt RSI smoke | PASS | v3正式bank真实non-PLAY policy path运行通过。 |
|
| 130 |
+
| G15 multi-row distribution readback | PASS | 8 train + 1 validation + 1 sealed-heldout正式rows运行通过。 |
|
| 131 |
+
| G16 completeness | PASS | 4096 terminal receipts；infra0；reflection120。 |
|
| 132 |
+
| G17 schema/timing | PASS | MLP213/GRU71/target31/state-t+1 71；所有admitted attempts满足`S=A+1`。 |
|
| 133 |
+
| G18 split/normalization | PASS | train2982/604314 transitions；validation498/100834；heldout496/100383；normalization仅来自admitted train。 |
|
| 134 |
+
| G19 final manifest | PASS | manifest SHA `1237116a7ed7f004fe0ed6d74edacf178089371f438a6147161d7d564af760d5`；archive SHA `f27c1c269f33b85127d0ae3a6b173c1188f1b5e44d2fb136ab09d9c864aa4f1c`。 |
|
| 135 |
+
| DATASET_READY | READY | 只支持phone_14 same-profile Teacher imitation。 |
|
| 136 |
+
|
| 137 |
+
## Approved Exception
|
| 138 |
+
|
| 139 |
+
Shengtao于2026-09-03批准本批次每个独立Teacher请求4096个DR+RSI rollout，固定3072/512/512三split，偏离默认128-attempt topology。该批准只授权Draft Card与checkpoint准备；正式并行collection须在Cards审核和G1–G15通过后另行启动。
|
objects/phone/trajectories/phone_14/stage4/data033-dr-rollouts-v1/dataset_manifest.json
ADDED
|
The diff for this file is too large to render.
See raw diff
|
|
|
objects/phone/trajectories/phone_14/stage4/data033-dr-rollouts-v1/rollout_dataset.tar
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:f27c1c269f33b85127d0ae3a6b173c1188f1b5e44d2fb136ab09d9c864aa4f1c
|
| 3 |
+
size 1039912960
|
objects/phone/trajectories/phone_17/stage4/data034-dr-rollouts-v1/checksums.json
ADDED
|
@@ -0,0 +1,8 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"archive_sha256": "fef1362e9b106d41a8256cd4acf2be04066dff3c8ce2e0d9b7ad6f3e9565b5e0",
|
| 3 |
+
"archive_size_bytes": 1056522240,
|
| 4 |
+
"data_task_id": "SHARPA-BC-DATA-034",
|
| 5 |
+
"manifest_sha256": "7ca167eaf9c6396785dfaabed6dcf0f90816026af576c2d06dfc7581c23b0caa",
|
| 6 |
+
"phone_id": "phone_17",
|
| 7 |
+
"schema": "sharpa_stage4_release_checksums_v1"
|
| 8 |
+
}
|
objects/phone/trajectories/phone_17/stage4/data034-dr-rollouts-v1/data_card.md
ADDED
|
@@ -0,0 +1,139 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
# Data Card: SHARPA-BC-DATA-034 / 20260903_sharpa_ac_p17_001_phone_17_c4096_dr_rsi_v1
|
| 2 |
+
|
| 3 |
+
| Field | Value |
|
| 4 |
+
|---|---|
|
| 5 |
+
| Data Task ID | SHARPA-BC-DATA-034 |
|
| 6 |
+
| Dataset ID | 20260903_sharpa_ac_p17_001_phone_17_c4096_dr_rsi_v1 |
|
| 7 |
+
| Status | READY / C4096 / INFRA0 |
|
| 8 |
+
| Purpose | 从phone_17 `SHARPA-AC-P17-001` S1 selected checkpoint采集4096个与Teacher训练时RefAug/reset-only DR及uniform RSI分布一致的独立closed-loop rollouts，形成3072 train、512 validation、512 sealed heldout。 |
|
| 9 |
+
| Source mode | `teacher_rollout`；fresh Isaac/PhysX；每个attempt独立采样DR与RSI，从采样phase执行到reference terminal再加15 control-step hold。 |
|
| 10 |
+
| Source lineage | `SHARPA-AC-P17-001`；W&B S1 selected `model_7999.pt`；原始8-seed evaluation仅为diagnostic，P2=`NOT_RUN`。 |
|
| 11 |
+
| Episodes / splits | 4096 preregistered attempts=`3072 train + 512 validation + 512 sealed heldout`；失败保留原split身份并进入reflection/infra，禁止按结果补抽或搬移。 |
|
| 12 |
+
| Sampling | 15D=`DR14 + uniform RSI phase`；每row独立采样；先物化、校准并冻结4096-row bank，之后才能launch。 |
|
| 13 |
+
| Observation | 71D deployable current frame及attempt-local history-3 213D view；不含velocity、phase/reference、Teacher ID、DR latent或future state。 |
|
| 14 |
+
| Action / label | controller真实执行的31D absolute `final_applied_target`；Teacher 28D residual仅作diagnostic，不能作为Student label。 |
|
| 15 |
+
| Normalization | admitted train only；attempt等权、attempt内transition等权；validation/heldout transform-only；std floor `1e-8`。 |
|
| 16 |
+
| Producer | formal producer commit `fc39d2f03df50e8888e2fd1bc00afbfc1725f1e5`；executable spec `outputs/formal034_v3/spec.json`。 |
|
| 17 |
+
| Artifact / manifest | create-only `outputs/formal034_v3`；bank SHA `bb8b3f18c0e56e540639530eb5d80aced6e960d32b193abd52e5e2c92cb5c8cf`；dataset manifest待4096 receipts后生成。 |
|
| 18 |
+
| Claim boundary | 仅支持phone_17、同一Teacher、同RefAug/reset-only DR+uniform-RSI分布的动作蒸馏研究；不支持physics/noise/push、跨episode、跨物体、P2 robustness或真机泛化。 |
|
| 19 |
+
|
| 20 |
+
## Source and Provenance
|
| 21 |
+
|
| 22 |
+
| Artifact | Identity / value | Evidence |
|
| 23 |
+
|---|---|---|
|
| 24 |
+
| Teacher Task | `SHARPA-AC-P17-001` | Auto-Curriculum S1 RefAug/reset-only DR candidate。 |
|
| 25 |
+
| Teacher Training Card | `teacher/checkpoints/sharpa_phone20_autocurriculum_selected_v1/phone_17/inputs/training_card.md`；SHA `04fdca8ab41229a84ad3f73887486b9dd63f9aad30b1b8c1e1509fdb0d445fae` | W&B selected-checkpoint artifact原样下载。 |
|
| 26 |
+
| Selected checkpoint | `teacher/checkpoints/sharpa_phone20_autocurriculum_selected_v1/phone_17/model_7999.pt`；`model_7999.pt`；iteration `7999`；SHA `675b27da543631c11da1c8b1b7736ee6cb12dcd89893b4d7ac9867fa2dfcf5d5` | W&B artifact `dfnet/dex-retarget/sharpa-phone_17-autocurriculum-s1-v1-s1-selected-checkpoint:v0`；local SHA gate PASS。 |
|
| 27 |
+
| Teacher code | commit `4a146fe07ef7f421745c3506ca803594f39b4fc8` | checkpoint selection manifest。 |
|
| 28 |
+
| Runtime manifest | `teacher/checkpoints/sharpa_phone20_autocurriculum_selected_v1/phone_17/inputs/runtime_manifest.json`；SHA `4be61b57c42fb3d69ff80ffa294ebb291ec9722575790545caad615462754d5a` | 仅保存manifest；完整可执行runtime仍须在G13物化并回读。 |
|
| 29 |
+
| Curriculum manifest | `teacher/checkpoints/sharpa_phone20_autocurriculum_selected_v1/phone_17/inputs/curriculum_manifest.json`；SHA `318a9b690e6e423a3ae48efcd621a2355b48f9724788cc3f35d9ec2ae388d77a` | S0 nominal → S1 RefAug/reset-only DR。 |
|
| 30 |
+
| Stage2 handoff | `teacher/checkpoints/sharpa_phone20_autocurriculum_selected_v1/phone_17/inputs/stage2_handoff.json`；SHA `d2ff7390e22f8701b38f739bcb1333d398123a872349e6fbb17a09baee69e4eb`；status `STAGE2_HANDOFF_ACCEPTED` | W&B artifact原样下载。 |
|
| 31 |
+
| Stage1 NPZ identity | SHA `06b1358a53b9d8ec1e17e7c48ce60e8782344e65e8443be4721c721bbff68176` | bytes尚未随本checkpoint bundle下载；G13前必须解析。 |
|
| 32 |
+
| Human H5 identity | SHA `48cb345eaaa2e3a2ab0fe6c3b6addf6010aba5b8953ac0f5028964dbd4937632` | `781` frames at `60.0` Hz。 |
|
| 33 |
+
| Runtime retarget H5 identity | SHA `75db34aeacfb38d934a5a83d2ee4d7742fc67ddfef27ac66e394154641e5f643` | `390` nominal policy targets at `30.0` Hz。 |
|
| 34 |
+
| Reference manifest identity | SHA `985f6aae457e8ee60f27ac08b6c60c88cde20d5bd191857489648bba57eb5084` | Stage2 accepted reference。 |
|
| 35 |
+
| GoalContract identity | SHA `f5f7ac14999b655413e910a151a1fc159cc2b851506f2e6250e82e68f188a342` | runtime artifact manifest。 |
|
| 36 |
+
| Historical evaluation | [W&B seed0–7 diagnostic](https://wandb.ai/dfnet/dex-retarget/runs/sharpa-phone_17-autocurriculum-evaluation-seed0-7-v1) | checkpoint identity matched；P2 `NOT_RUN / NON_PROMOTABLE`。 |
|
| 37 |
+
|
| 38 |
+
## Schema
|
| 39 |
+
|
| 40 |
+
| Field | Shape / type | Semantics |
|
| 41 |
+
|---|---|---|
|
| 42 |
+
| `time_s` | `(S,) float64` | 30 Hz control timestamps from the sampled RSI phase。 |
|
| 43 |
+
| `actual_q` | `(S,22) float32`, rad | PhysX actual Sharpa finger joints。 |
|
| 44 |
+
| `actual_wrist_pose_wxyz` | `(S,7) float32`, world frame | actual wrist translation and quaternion。 |
|
| 45 |
+
| `actual_object_pose_wxyz` | `(S,7) float32`, world frame | actual phone translation and quaternion。 |
|
| 46 |
+
| `raw_residual` / `processed_residual` | `(A,28) float32` | Teacher diagnostic action before/after adapter processing。 |
|
| 47 |
+
| `final_applied_target_wxyz` | `(A,29) float32` | runtime controller target representation。 |
|
| 48 |
+
| `student_frame` | `(A,71) float32` | deployable current-frame observation。 |
|
| 49 |
+
| `applied_target_31` | `(A,31) float32` | unique BC supervision label: wrist position3 + rotation6D6 + joint22。 |
|
| 50 |
+
| `state_t1_student_frame` | `(A,71) float32` | post-physics next state，audit-only。 |
|
| 51 |
+
| `pose_history3_mlp_v1` | `(A,213) float32` | three attempt-local 71D frames，oldest→current；first-frame repeat padding。 |
|
| 52 |
+
| `normalized_sampling_15d` | `(15,) float64` | DR14+RSI provenance-only；不得进入policy input。 |
|
| 53 |
+
| reset/failure/action ledgers | JSON | sampled parameters、actual runtime readback、drop/invalid/infra与action-chain evidence。 |
|
| 54 |
+
|
| 55 |
+
## Temporal Alignment
|
| 56 |
+
|
| 57 |
+
| Item | Contract |
|
| 58 |
+
|---|---|
|
| 59 |
+
| Rates | PhysX 120 Hz；policy/control 30 Hz；decimation=4。 |
|
| 60 |
+
| Transition | `obs_t → Teacher residual_t → controller final_applied_target_t → four PhysX substeps → state_t+1`。 |
|
| 61 |
+
| RSI | `u_phase ~ U[0,1)`；`start=floor(u_phase × 390)`，clamp到`0…389`。 |
|
| 62 |
+
| Horizon | `A=(390-start)+15`；`S=A+1`；即使drop/invalid也记录完整终态与ledger。 |
|
| 63 |
+
| History | strictly attempt-local；不得跨attempt、split或index gap。 |
|
| 64 |
+
| Future windows | 后续Consumer可构造`T_p` window，但window不得跨attempt；tail必须masked。 |
|
| 65 |
+
|
| 66 |
+
## Simulator Environment Randomization
|
| 67 |
+
|
| 68 |
+
| Field | Value |
|
| 69 |
+
|---|---|
|
| 70 |
+
| Enabled | true |
|
| 71 |
+
| Profile | `SHARPA-AC-P17-001` S1 RefAug/reset-only DR + uniform RSI。 |
|
| 72 |
+
| Source config SHA | `d99efe6b27a179211fec8f3077e5605f9d804c666f0ac9a16c3e8349795982b8`。 |
|
| 73 |
+
| Activation | every reset；nominal fraction `0.0`。 |
|
| 74 |
+
| Mask | common rigid transform同时作用hand与phone；relative reset保持grasp-shaped范围。 |
|
| 75 |
+
| Seed | 4096-row bank seed/namespace在distribution calibration阶段冻结；当前`NOT_YET_MATERIALIZED`。 |
|
| 76 |
+
|
| 77 |
+
| Category | Parameter | Distribution / range | Sampling / effect |
|
| 78 |
+
|---|---|---|---|
|
| 79 |
+
| Reference augmentation | common X/Y | each `U[-0.05,0.05] m` | 同时作用hand和phone。 |
|
| 80 |
+
| Reference augmentation | common yaw | `U[-30°,30°]` | 同时作用hand和phone。 |
|
| 81 |
+
| Relative reset | phone XYZ | each `U[-0.006,0.006] m` | after RSI/reference selection。 |
|
| 82 |
+
| Relative reset | phone rotation | axis uniformly sampled；signed angle `U[-3°,3°]` | every reset。 |
|
| 83 |
+
| Finger reset | shared wrap | `U[0,0.06] rad` | coupled flexion joints，只增加包握。 |
|
| 84 |
+
| Finger reset | four-finger rub | each `U[-0.025,0.025] rad` | independent rub joints。 |
|
| 85 |
+
| RSI | normalized phase | `U[0,1)` | maps to discrete start phase。 |
|
| 86 |
+
|
| 87 |
+
### Explicitly disabled / not applicable
|
| 88 |
+
|
| 89 |
+
| Option | State | Meaning |
|
| 90 |
+
|---|---|---|
|
| 91 |
+
| Physics/actuator randomization | disabled | mass、COM、inertia、friction与drive固定。 |
|
| 92 |
+
| Observation/noise/lag | disabled | clean observation；zero lag。 |
|
| 93 |
+
| Disturbance/push | disabled | none。 |
|
| 94 |
+
| Gravity curriculum | disabled | fixed full gravity `-9.81 m/s²`。 |
|
| 95 |
+
|
| 96 |
+
## Splits, Sampling, and Leakage
|
| 97 |
+
|
| 98 |
+
Approved Exception: Shengtao于2026-09-03批准每个独立Teacher使用4096 attempts并划分为3072 train、512 validation、512 sealed heldout，明确偏离默认96 train + 32 heldout topology。
|
| 99 |
+
|
| 100 |
+
| Split | Episodes / rows | Purpose |
|
| 101 |
+
|---|---:|---|
|
| 102 |
+
| train | 3072 | admitted rows的唯一optimization与normalization来源。 |
|
| 103 |
+
| validation | 512 | offline checkpoint selection；不得参与normalization。 |
|
| 104 |
+
| sealed heldout | 512 | 所有模型选择冻结后的同分布unseen-attempt audit。 |
|
| 105 |
+
| reflection | E5/E6/E9失败 | 保留预注册split身份；不得进入监督数据或被成功补抽替换。 |
|
| 106 |
+
| infra_failure | identity/runtime/schema failure | 阻止Dataset READY；不得解释成policy failure。 |
|
| 107 |
+
|
| 108 |
+
Split membership必须在rollout前由case ID确定。同一seed、trajectory window、parent/derived copy不得跨split；heldout不得反向用于collector、normalization、checkpoint、模型或超参数选择。
|
| 109 |
+
|
| 110 |
+
## Normalization
|
| 111 |
+
|
| 112 |
+
| Field | Contract |
|
| 113 |
+
|---|---|
|
| 114 |
+
| Source | admitted train attempts only。 |
|
| 115 |
+
| Weighting | each rollout equal；within each rollout，each transition equal。 |
|
| 116 |
+
| Stored arrays | MLP213、GRU71与target31 mean/second moment/std，以及exact train attempt IDs和weights。 |
|
| 117 |
+
| Floor | `std=max(std,1e-8)`。 |
|
| 118 |
+
| Validation / heldout | train statistics transform-only；禁止重算。 |
|
| 119 |
+
|
| 120 |
+
## Validation and Readiness
|
| 121 |
+
|
| 122 |
+
| Gate | Result | Evidence |
|
| 123 |
+
|---|---|---|
|
| 124 |
+
| Data Card format | PASS | unified validator对本批次7张Card报告`validated=7 failures=0`。 |
|
| 125 |
+
| Selected checkpoint identity | PASS | W&B selected artifact与local SHA一致：`675b27da543631c11da1c8b1b7736ee6cb12dcd89893b4d7ac9867fa2dfcf5d5`。 |
|
| 126 |
+
| Stage2/H5/runtime resolvability | PASS | `outputs/prepared_v1`中clone-local H5/assets/runtime逐SHA解析通过。 |
|
| 127 |
+
| G1–G12 distribution calibration | PASS | `outputs/formal034_v3/distribution_calibration.json`；SHA `c8e51a6ff2f0a308336a21489818ac382b1c5aed54b40da3248eb0b38afed429`。 |
|
| 128 |
+
| G13 identity/runtime | PASS | producer `fc39d2f`；checkpoint、runtime、artifact、goal与spec/bank SHA preflight通过。 |
|
| 129 |
+
| G14 one-attempt RSI smoke | PASS | v3正式bank真实non-PLAY policy path运行通过。 |
|
| 130 |
+
| G15 multi-row distribution readback | PASS | 8 train + 1 validation + 1 sealed-heldout正式rows运行通过。 |
|
| 131 |
+
| G16 completeness | PASS | 4096 terminal receipts；infra0；reflection139。 |
|
| 132 |
+
| G17 schema/timing | PASS | MLP213/GRU71/target31/state-t+1 71；所有admitted attempts满足`S=A+1`。 |
|
| 133 |
+
| G18 split/normalization | PASS | train2966/614057 transitions；validation498/103259；heldout493/101718；normalization仅来自admitted train。 |
|
| 134 |
+
| G19 final manifest | PASS | manifest SHA `7ca167eaf9c6396785dfaabed6dcf0f90816026af576c2d06dfc7581c23b0caa`；archive SHA `fef1362e9b106d41a8256cd4acf2be04066dff3c8ce2e0d9b7ad6f3e9565b5e0`。 |
|
| 135 |
+
| DATASET_READY | READY | 只支持phone_17 same-profile Teacher imitation。 |
|
| 136 |
+
|
| 137 |
+
## Approved Exception
|
| 138 |
+
|
| 139 |
+
Shengtao于2026-09-03批准本批次每个独立Teacher请求4096个DR+RSI rollout，固定3072/512/512三split，偏离默认128-attempt topology。该批准只授权Draft Card与checkpoint准备；正式并行collection须在Cards审核和G1–G15通过后另行启动。
|
objects/phone/trajectories/phone_17/stage4/data034-dr-rollouts-v1/dataset_manifest.json
ADDED
|
The diff for this file is too large to render.
See raw diff
|
|
|
objects/phone/trajectories/phone_17/stage4/data034-dr-rollouts-v1/rollout_dataset.tar
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:fef1362e9b106d41a8256cd4acf2be04066dff3c8ce2e0d9b7ad6f3e9565b5e0
|
| 3 |
+
size 1056522240
|
objects/phone/trajectories/phone_18/stage4/data018-dr-rollouts-v1/checksums.json
ADDED
|
@@ -0,0 +1,8 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"archive_sha256": "87b68630a80467e3fa40642171478f4bade5a4f6d2b7143809a006899c3a57e1",
|
| 3 |
+
"archive_size_bytes": 1199267840,
|
| 4 |
+
"data_task_id": "SHARPA-BC-DATA-018",
|
| 5 |
+
"manifest_sha256": "449ed73295963a97d25b7f5cbd7b230038493ad38ed2fb3b2b92383742266076",
|
| 6 |
+
"phone_id": "phone_18",
|
| 7 |
+
"schema": "sharpa_stage4_release_checksums_v1"
|
| 8 |
+
}
|
objects/phone/trajectories/phone_18/stage4/data018-dr-rollouts-v1/data_card.md
ADDED
|
@@ -0,0 +1,154 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
# Data Card: SHARPA-BC-DATA-018 / 20260829_t0148_model7050_phone18_ppo_rsi_c3072_v1
|
| 2 |
+
|
| 3 |
+
| Field | Value |
|
| 4 |
+
|---|---|
|
| 5 |
+
| Data Task ID | SHARPA-BC-DATA-018 |
|
| 6 |
+
| Dataset ID | 20260829_t0148_model7050_phone18_ppo_rsi_c3072_v1 |
|
| 7 |
+
| Status | DRAFT / G1-G12 DISTRIBUTION_GATES_PASS / G13-G19 NOT_RUN / COLLECTION NOT_LAUNCHED / NOT_READY |
|
| 8 |
+
| Purpose | 从phone18 T0-148 model7050采集4096个与PPO训练DR14+uniform-RSI分布一致的独立Teacher rollouts，形成3072 train、512 validation、512 sealed heldout。 |
|
| 9 |
+
| Source mode | teacher_rollout；fresh Isaac/PhysX；从预注册RSI phase运行到terminal+15 hold。 |
|
| 10 |
+
| Source lineage | SHARPA-T0-148 S1 RefAug/reset-only DR model7050；历史DATA-012为小样本基线，不复用其rows或normalization。 |
|
| 11 |
+
| Episodes / splits | 4096 preregistered attempts=`3072 train +512 validation +512 heldout`；Shengtao批准偏离默认96/32规则；失败保留原split进入reflection/infra。 |
|
| 12 |
+
| Sampling | 与DATA-017共享同一4096-row normalized15D bank；15D=`DR14 + uniform RSI phase`；phone18按480 motion actions映射离散start phase。 |
|
| 13 |
+
| Observation | 71D deployable current frame +213D attempt-local history3；不含velocity、phase/reference、Teacher ID、DR latent或future state。 |
|
| 14 |
+
| Action / label | controller实际`final_applied_target`转为31D absolute target；28D Teacher residual仅diagnostic。 |
|
| 15 |
+
| Normalization | admitted train only；rollout等权、rollout内transition等权；validation/heldout transform-only；std floor`1e-8`。 |
|
| 16 |
+
| Producer | PPO-matched RSI bank/calibration commit`d475d272d683eb8a2859761cfde4966e5ed67c53`；collector/finalizer revision尚未通过G13。 |
|
| 17 |
+
| Artifact / manifest | create-only `outputs/sharpa_bc_data018/20260829_t0148_model7050_phone18_ppo_rsi_c3072_v1`；manifest`NOT_YET_MATERIALIZED`。 |
|
| 18 |
+
| Claim boundary | 只支持phone18、T0-148 model7050、同DR14+uniform-RSI分布中的Teacher imitation；不支持physics/noise/push、phase0完整任务、跨物体或真机泛化。 |
|
| 19 |
+
|
| 20 |
+
## Source and Provenance
|
| 21 |
+
|
| 22 |
+
| Artifact | Identity / value | Evidence |
|
| 23 |
+
|---|---|---|
|
| 24 |
+
| Teacher Task/Card | SHARPA-T0-148；Card SHA`4e47cedfc99f46cd6d694bc2dc8cf1fdcc0a31d0027e31559529041794d0fd96` | phone18 Auto-Curriculum DR Teacher。 |
|
| 25 |
+
| Teacher checkpoint | model7050 SHA`8bdef80bb104e64c40f3a0bbdd33c7d7cedec759ec0a7bb9091db4c2fbb27989` | Same Teacher as DATA-012。 |
|
| 26 |
+
| Runtime manifest | SHA`1c68935470489e073254ff3cef5091598aff2bafa092670a221e2b8885faa4df` | matching T0-148 runtime。 |
|
| 27 |
+
| Runtime readback | SHA`2ae174c762c8548286d23610e106db193457a1166d2d960d91d2a675d5b314b4` | uniform RSI、RefAug enabled、fixed physics、zero lag。 |
|
| 28 |
+
| Artifact manifest | SHA`7670bba805e0448d13d137931cf186fbc749e379c795c9de89b0620f261c6eea` | create-only portable phone18 assets/reference；避免绑定可变父工程outputs。 |
|
| 29 |
+
| Stage1 NPZ | SHA`3d2be9aa9701c654f1ed85cba3af593858e2df02eb4d40bb710b67266b4c2682` | phone18 object-centric source。 |
|
| 30 |
+
| Runtime human H5 | SHA`e8dc5e5b03a3b3c49bae9c7c130760e823163438cfa08b87e3650a04915d5801` | 961 frames@60Hz。 |
|
| 31 |
+
| Runtime retarget H5 | SHA`346012a9b5a55ec3570a1f8cc8ad6985ae9805de12a860d3695e4c9ce3a00c66` | 481 control states/480 motion actions。 |
|
| 32 |
+
| Reference manifest | SHA`0ebcb16b68baf202a44227ac228dcf9e5e43d96547cb34aec6638be3b2419af5` | portable、路径重绑定后的phone18 reference receipt；Stage1/H5 bytes不变。 |
|
| 33 |
+
| Shared15D bank | paired with DATA-017 | `NOT_YET_MATERIALIZED`。 |
|
| 34 |
+
|
| 35 |
+
## Schema
|
| 36 |
+
|
| 37 |
+
| Field | Shape / type | Semantics |
|
| 38 |
+
|---|---|---|
|
| 39 |
+
| time_s | `(S,)` float64 | 30Hz timestamps from sampled RSI phase。 |
|
| 40 |
+
| actual_q | `(S,22)` float32 rad | actual joints。 |
|
| 41 |
+
| actual_wrist_pose_wxyz | `(S,7)` float32 | actual world wrist pose。 |
|
| 42 |
+
| actual_object_pose_wxyz | `(S,7)` float32 | actual world phone pose。 |
|
| 43 |
+
| raw/processed_residual | `(A,28)` float32 | Teacher diagnostic。 |
|
| 44 |
+
| final_applied_target_wxyz | `(A,29)` float32 | executed runtime target。 |
|
| 45 |
+
| student_frame | `(A,71)` float32 | deployable pose/previous-target frame。 |
|
| 46 |
+
| applied_target_31 | `(A,31)` float32 | unique BC label。 |
|
| 47 |
+
| state_t1_student_frame | `(A,71)` float32 | audit-only。 |
|
| 48 |
+
| pose_history3_mlp_v1 | `(A,213)` float32 | attempt-local3×71。 |
|
| 49 |
+
| normalized_sampling_15d | `(15,)` float64 | DR14+RSI provenance-only。 |
|
| 50 |
+
| reset/failure ledgers | JSON | physical/runtime reset及失败证据。 |
|
| 51 |
+
|
| 52 |
+
71D slices和codec与DATA-017一致；member bytes与manifest identity独立。
|
| 53 |
+
|
| 54 |
+
## Temporal Alignment
|
| 55 |
+
|
| 56 |
+
| Item | Contract |
|
| 57 |
+
|---|---|
|
| 58 |
+
| Rates | PhysX120Hz；control30Hz；decimation4。 |
|
| 59 |
+
| Transition | `obs_t → Teacher residual_t → final_applied_target_t → four substeps → state_t+1`。 |
|
| 60 |
+
| RSI | `u_phase~U[0,1)`；481 control states下`start=floor(u_phase×480)`，范围0…479；真实PPO RSI路径。 |
|
| 61 |
+
| Horizon | `A=(480-start)+15`；`S=A+1`。 |
|
| 62 |
+
| History | no cross-attempt/split/source/index gap。 |
|
| 63 |
+
| Termination | no auto-reset；失败保存完整ledger。 |
|
| 64 |
+
|
| 65 |
+
## Simulator Environment Randomization
|
| 66 |
+
|
| 67 |
+
| Field | Value |
|
| 68 |
+
|---|---|
|
| 69 |
+
| Enabled | true |
|
| 70 |
+
| Profile | T0-148/T0-149 RefAug/reset-only DR + uniform RSI。 |
|
| 71 |
+
| Source config SHA | `14f894ff771ebdeed6c6778aa1e18a286d6a4980a1141b400e6df02c58c1ef37`。 |
|
| 72 |
+
| Runtime readback SHA | `2ae174c762c8548286d23610e106db193457a1166d2d960d91d2a675d5b314b4`。 |
|
| 73 |
+
| Activation | every reset；nominal fraction0。 |
|
| 74 |
+
| Seed | paired DATA-017/018 namespace；`NOT_YET_MATERIALIZED`。 |
|
| 75 |
+
|
| 76 |
+
| Category | Parameter | Distribution / range | Sampling / effect |
|
| 77 |
+
|---|---|---|---|
|
| 78 |
+
| Reference augmentation | common X/Y | each`U[-0.05,0.05] m` | shared hand/object。 |
|
| 79 |
+
| Reference augmentation | common yaw | `U[-30°,30°]` | shared hand/object。 |
|
| 80 |
+
| Relative reset | phone xyz | each`U[-0.006,0.006] m` | after RSI/reference selection。 |
|
| 81 |
+
| Relative reset | phone rotation | uniform sphere axis；signed`U[-3°,3°]` | exact Teacher distribution。 |
|
| 82 |
+
| Finger reset | shared wrap | `U[0,0.06] rad` | coupled16 flexion joints。 |
|
| 83 |
+
| Finger reset | four rub joints | each`U[-0.025,0.025] rad` | independent。 |
|
| 84 |
+
| RSI | normalized phase | `U[0,1)` | uniform timestep。 |
|
| 85 |
+
|
| 86 |
+
### Explicitly disabled / not applicable
|
| 87 |
+
|
| 88 |
+
| Option | State | Meaning |
|
| 89 |
+
|---|---|---|
|
| 90 |
+
| Physics/actuator randomization | disabled | fixed。 |
|
| 91 |
+
| Observation/noise/lag | disabled | clean、zero lag。 |
|
| 92 |
+
| Disturbance/push | disabled | none。 |
|
| 93 |
+
| Gravity curriculum | disabled | fixed full gravity。 |
|
| 94 |
+
|
| 95 |
+
## Splits, Sampling, and Leakage
|
| 96 |
+
|
| 97 |
+
| Split | Episodes / rows | Purpose |
|
| 98 |
+
|---|---:|---|
|
| 99 |
+
| train | 3072 | optimization/normalization only。 |
|
| 100 |
+
| validation | 512 | checkpoint selection；no normalization。 |
|
| 101 |
+
| sealed heldout | 512 | final audit。 |
|
| 102 |
+
| reflection | Teacher E5/E6/E9 failure | original split retained。 |
|
| 103 |
+
| infra_failure | identity/runtime/schema failure | blocks READY。 |
|
| 104 |
+
|
| 105 |
+
DATA-017/018逐row paired；evaluation bank与all splits完全无交集。
|
| 106 |
+
|
| 107 |
+
## Normalization
|
| 108 |
+
|
| 109 |
+
| Field | Contract |
|
| 110 |
+
|---|---|
|
| 111 |
+
| Source | admitted train only。 |
|
| 112 |
+
| Weighting | rollout equal；within-rollout transitions equal。 |
|
| 113 |
+
| Stored | MLP213/GRU71/target31 moments与IDs。 |
|
| 114 |
+
| Floor | `max(std,1e-8)`。 |
|
| 115 |
+
| Validation/heldout | transform-only。 |
|
| 116 |
+
|
| 117 |
+
## Statistical Distribution Calibration Gates
|
| 118 |
+
|
| 119 |
+
train3072、validation512、heldout512与full4096分别运行G1～G12：唯一性、mean/variance、KS、bin occupancy、correlation、15D discrepancy、corner occupancy、RSI uniformity、split parity、source pairing、physical coupling与no-leakage。
|
| 120 |
+
|
| 121 |
+
每项与同样本量目标15D uniform Monte Carlo reference的99%包络比较；calibration只证明参数分布，不证明Teacher execution或Student可学。
|
| 122 |
+
|
| 123 |
+
## Validation and Readiness
|
| 124 |
+
|
| 125 |
+
| Gate | Result | Evidence |
|
| 126 |
+
|---|---|---|
|
| 127 |
+
| Data Card format | PENDING | validator待运行。 |
|
| 128 |
+
| G1–G12 distribution calibration | NOT_RUN | 15D bank尚未物化。 |
|
| 129 |
+
| G13 identity | NOT_RUN | immutable identities待冻结。 |
|
| 130 |
+
| G14 one-attempt RSI | NOT_RUN | non-PLAY真实路径。 |
|
| 131 |
+
| G15 multi-row distribution | NOT_RUN | center/tail/corner/三split。 |
|
| 132 |
+
| G16 completeness | NOT_LAUNCHED | 4096 receipts。 |
|
| 133 |
+
| G17 schema/timing | NOT_RUN | codec/cardinality/action parity。 |
|
| 134 |
+
| G18 split/normalization | NOT_RUN | train-only rollout weighting。 |
|
| 135 |
+
| G19 final manifest | NOT_RUN | member SHA/indexes。 |
|
| 136 |
+
| DATASET_READY | NOT_READY | no downstream formal training。 |
|
| 137 |
+
|
| 138 |
+
## Approved Coverage Exception
|
| 139 |
+
|
| 140 |
+
Shengtao批准DATA-018采集4096个PPO-matched DR+RSI Teacher rollouts，split3072/512/512；不创建smaller coverage views，不覆盖历史DATA-012或未采集phase0 DATA-016。
|
| 141 |
+
|
| 142 |
+
## Execution Update — 2026-08-29 G1–G12
|
| 143 |
+
|
| 144 |
+
| Artifact / gate | Result |
|
| 145 |
+
|---|---|
|
| 146 |
+
| Coverage spec | SHA`f06fab4e2bbe7288c3a8009b796af91728b3bce94902d79af3a726d4a713b69b`。 |
|
| 147 |
+
| Shared4096-row15D bank | SHA`4cc240c37251d68c217384d7a1961fcc353037a3cfee70a838d55c547cd88a19`；3072/512/512。 |
|
| 148 |
+
| G1–G9 statistical report | PASS；512 Monte Carlo reference ensembles；SHA`6719519ce43cd5c450dcff8f4889bdd38e7f9481e41dcff32288b1b762f5342b`。 |
|
| 149 |
+
| phone18 source bank | 4096 rows；RSI range0…479；SHA`ea6f9ab28fbf627eb90ae80162d641adf0d0ee357bf90afcb2e5a4abf315eb08`。 |
|
| 150 |
+
| G10/G11 pairing/physical | 4096/4096 PASS；SHA`32a787f4b47c0595ce265a37fd869277395ecfc5b06bc130407da15b95661e47`。 |
|
| 151 |
+
| G12 evaluation leakage | PASS；独立256-row bank SHA`b52468c53de80ca7c148738e41c33f3e65eadc6ea0e37604c1d5e78ddfae0394`。 |
|
| 152 |
+
| Code tests | 240 PASS；旧128-row与phase0历史回归均保留。 |
|
| 153 |
+
|
| 154 |
+
G1–G12只放行G13 identity/collector实现；尚未运行Isaac或产生Teacher rollout，DATA-018保持NOT_READY。
|