ChaoyiPan commited on
Commit
01f39a6
·
verified ·
1 Parent(s): 75a2b92

Publish SPIDER agentic pipeline materials: agentic_pipeline/2026-10-07-object-motion

Browse files
agentic_pipeline/2026-10-07-object-motion/README.md ADDED
@@ -0,0 +1,60 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # Verified Sharpa object motion and agentic pipeline — 2026-10-07
2
+
3
+ [**Watch the verified retarget**](whisking_sharpa_4x/independent_evaluation/comparison.mp4) · [Results and source video](whisking_sharpa_4x/) · [Metrics](whisking_sharpa_4x/RESULT.json)
4
+
5
+ The complete reconstructed whisk path is retargeted to **Sharpa at quarter speed**: 138 source poses, 9,143 actual simulation steps, 18.285 seconds of execution. Maximum position error is **14.842 mm**, maximum rotation error **5.15°**, and thumb/index contact remains loaded throughout. Independent replay of the saved robot controls matches every saved generalized position exactly.
6
+
7
+ Object motion and stable grasp take priority over human hand pose. The object poses are unchanged and source timestamps are explicitly multiplied by four. Execution uses full gravity with no object actuation, applied forces, equality constraints, or supporting floor contacts. This verifies grasped-tool motion; the source pot/liquid is absent. Original-speed attempts failed. Hand actuator force limits are not hardware-calibrated.
8
+
9
+ ## Download and run
10
+
11
+ From an empty directory:
12
+
13
+ ```bash
14
+ BASE=https://huggingface.co/datasets/retarget/example_video_data/resolve/main/agentic_pipeline/2026-10-07-object-motion
15
+ for FILE in source.tar.gz datasets.tar.gz manifest.json SHA256SUMS; do
16
+ curl --fail --location --retry 3 "$BASE/$FILE" --output "$FILE"
17
+ done
18
+ sha256sum --check SHA256SUMS
19
+ tar -xzf source.tar.gz
20
+ tar -xzf datasets.tar.gz
21
+ cd spider
22
+ uv sync --frozen --extra dev
23
+ ```
24
+
25
+ Both archives share the `spider/` root. Open Codex or Claude Code there. Ask `$spider-retarget` or `/spider-retarget` to reproduce the verified object-motion example. The six skills cover the full pipeline, video reconstruction/import, dataset conversion, retargeting, parameter tuning, and visualization. Sharpa is the default hand. Keep the Claude symlinks when copying the checkout.
26
+
27
+ Run the fresh CPU retarget directly:
28
+
29
+ ```bash
30
+ CASE=example_datasets/full_video/whisking_sharpa_object_priority_4x
31
+ SEED=example_datasets/full_video/power_grasp/azimuth-minus-90-verified
32
+ MUJOCO_GL=egl uv run python -m spider.video_retarget.grasp_motion \
33
+ --seed-dir "$SEED" \
34
+ --reference "$CASE/source_trajectory_original_timing.npz" \
35
+ --human-keypoints "$CASE/source_human_keypoints.npz" \
36
+ --output-dir example_datasets/whisking_object_reproduction \
37
+ --time-scale 4 --grip-normal-preload-N 10 \
38
+ --source-video-alignment "$CASE/source_video_alignment_original.json" \
39
+ --render
40
+ ```
41
+
42
+ Inspect `example_datasets/whisking_object_reproduction/independent_evaluation/comparison.mp4`, `evaluation.json`, and `execution.json`. Use a fresh output directory for each run. The command needs no CUDA or neural weights. Other full MPC and tuning commands use CUDA and are documented in `docs/usage/agentic-pipeline.md`.
43
+
44
+ ## Included materials
45
+
46
+ - Current source code, Sharpa assets, six skills shared by Codex/Claude, tests, and README/workflow documentation.
47
+ - Working object-motion case: source video and reconstruction references, portable scene, grasp seed, actual trajectory and controls, comparison video, contact records, independent evaluation and exact replay evidence.
48
+ - Public raw conversion examples for GigaHands (3), OakInk (7), OakInk-v2 (7), FAIR-FRE (21), and FAIR-MON (2), with all **11,626 Sharpa IK frames**. These conversion/IK checks do not claim full manipulation success for all 40 clips.
49
+ - Whisking reconstruction inputs and the separately labeled original upstream Sharpa recording. Failed full attempts remain documented in the [earlier attempt release](../2026-10-06-full-retarget/).
50
+
51
+ Arctic/HOT3D provider boundaries and GigaHands' optional MANO path have fixture coverage; official source recordings and MANO models are not included. New-video neural reconstruction requires external environments and gated models. The supplied RGB/depth cup data has not been newly reconstructed into a successful robot trajectory.
52
+
53
+ ## Provenance and licenses
54
+
55
+ - SPIDER code: repository `LICENSE` (CC BY-NC 4.0).
56
+ - Do-as-I-Do port/assets/demo: MIT, upstream commit `824591b808c342b20079c3b4198a8c2bdf88c74e`; see bundled `third_party/do-as-i-do/LICENSE` and robot asset notices.
57
+ - Public conversion inputs: `retarget/retarget_example` revision `8b0cdf2b24cc86d1c859050c5b0317ddbec59a63`; original dataset terms apply.
58
+ - Original `pickcup_right/` RGB/depth data remains at this Hugging Face dataset root with its original license.
59
+
60
+ `manifest.json` lists the exact archived directories, source revision, byte counts, and SHA-256 hashes. Verify `SHA256SUMS` before extraction. Execution receipts preserve original paths for provenance; use the portable scenes and documented relative input paths for reproduction.
agentic_pipeline/2026-10-07-object-motion/SHA256SUMS ADDED
@@ -0,0 +1,4 @@
 
 
 
 
 
1
+ 7f9108f15753cab4c5a0edfeb5d85406817e762b718d99f47406a25651fb408e source.tar.gz
2
+ 9b8c158e752ae4edc8dc551262fa2f0f447c3c86725500dbcf39d1af73fafa9a datasets.tar.gz
3
+ bcb549af4d1002b0f7cf5478ad0e6ef6f8e2bcc5acead9702e9784bfec6f9d4f manifest.json
4
+ 5e5c38819f8ffe3b3ee9f17071d7c14f6e7b75d4fd516e2cbdec0a76a26ed08a README.md
agentic_pipeline/2026-10-07-object-motion/datasets.tar.gz ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:9b8c158e752ae4edc8dc551262fa2f0f447c3c86725500dbcf39d1af73fafa9a
3
+ size 840051297
agentic_pipeline/2026-10-07-object-motion/manifest.json ADDED
@@ -0,0 +1,43 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "schema_version": 1,
3
+ "created_at": "2026-10-07T02:25:24.118957+00:00",
4
+ "source_commit": "41477aa4958c590668b610f19bbb991fd99dfa30",
5
+ "source_has_task_changes": false,
6
+ "archives": {
7
+ "source.tar.gz": {
8
+ "sha256": "7f9108f15753cab4c5a0edfeb5d85406817e762b718d99f47406a25651fb408e",
9
+ "bytes": 183412881,
10
+ "files": 938
11
+ },
12
+ "datasets.tar.gz": {
13
+ "sha256": "9b8c158e752ae4edc8dc551262fa2f0f447c3c86725500dbcf39d1af73fafa9a",
14
+ "bytes": 840051297,
15
+ "files": 3138
16
+ }
17
+ },
18
+ "data_directories": [
19
+ "example_datasets/pipeline_verification",
20
+ "example_datasets/video_examples",
21
+ "example_datasets/full_video/whisking_sharpa_object_priority_4x",
22
+ "example_datasets/full_video/power_grasp/azimuth-minus-90-verified",
23
+ "example_datasets/full_video/source_alignment",
24
+ "example_datasets/full_video/whisking_sharpa_object_priority_cli_verified"
25
+ ],
26
+ "notes": [
27
+ "Source archive includes current task changes and both agents' skills.",
28
+ "Neural model weights and MANO assets are not included; precomputed reconstruction is included.",
29
+ "Execution receipts retain original paths as provenance; use a fresh output directory for new runs."
30
+ ],
31
+ "verification": {
32
+ "shared_tests_passed": 199,
33
+ "shared_subtests_passed": 17,
34
+ "video_backend_tests_passed": 47,
35
+ "skill_definitions_validated": 6,
36
+ "working_object_case": "whisking_sharpa_4x",
37
+ "source_time_scale": 4,
38
+ "full_source_covered": true,
39
+ "exact_control_replay": true,
40
+ "production_cli_exact_reproduction": true,
41
+ "portable_seed_hold_seconds": 30
42
+ }
43
+ }
agentic_pipeline/2026-10-07-object-motion/source.tar.gz ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:7f9108f15753cab4c5a0edfeb5d85406817e762b718d99f47406a25651fb408e
3
+ size 183412881