ChaoyiPan commited on
Commit
1bb99e5
·
verified ·
1 Parent(s): 9129406

Feature the corrected source-timed grasp-establishment workflow

Browse files
README.md CHANGED
@@ -6,20 +6,18 @@ license: apache-2.0
6
 
7
  # SPIDER video and data workflows
8
 
9
- [**Watch the RGB-D cup retarget**](agentic_pipeline/2026-10-07-rgbd-video/pickcup_right/motion/comparison.mp4) · [Result and evidence](agentic_pipeline/2026-10-07-rgbd-video/pickcup_right/README.md) · [Download code/data and run](agentic_pipeline/2026-10-07-rgbd-video/README.md)
10
 
11
- The new workflow starts from the original `pickcup_right` RGB/depth frames and intrinsics, estimates object geometry and motion, generates a **Sharpa** grasp, retargets the complete estimated reference, and exports video and verification reports. It runs locally without a GPU or MANO models. Call `$spider-video-to-motion` in Codex or `/spider-video-to-motion` in Claude Code; six shared skills cover conversion, video processing, retargeting, tuning, and visualization.
12
 
13
- The complete **269-frame estimated cup path** executes over **17,840 simulation steps at quarter speed**: 12.55 mm maximum position error, 8.68° maximum rotation error, and 99.98% loaded contact. Fresh end-to-end execution and independent control replay reproduce the selected states exactly. Object motion and stable grasp take priority over human hand pose.
14
 
15
- **This is an air-manipulation diagnostic, not a table-safe pouring or hardware result.** Thirty-six source poses remain uncertain during occlusion and axial spin is unobserved. Registered depth, depth scale, level-table gravity, approximate cup shape and mass are explicit assumptions. Nominal table crossings reach 5.55 mm for physical hand geometry, 9.29 mm for visual meshes and 5.93 mm for the cup. There is no table support, object actuator, external force or equality assistance. Original-speed execution failed. The workflow stops for source-quality review by default; the published diagnostic uses an explicit uncertainty override.
16
 
17
- The update passed **135 tests** and all six skill definitions validated. Code, original inputs, meshes, scene, grasp certificates, actual controls/trajectories, video, uncertainty and clearance reports are included. Runtime logs are excluded; logs were also removed from the two previous data archives and the exposed example folders. Configs and verification evidence remain.
18
 
19
- The [earlier quarter-speed Whisking example](agentic_pipeline/2026-10-07-object-motion/README.md) and [previous attempts](agentic_pipeline/2026-10-06-full-retarget/) remain available. Conversion/full IK checks cover 40 public GigaHands, OakInk, OakInk-v2, FAIR-FRE and FAIR-MON examples (11,626 Sharpa frames); these checks do not claim full manipulation success for every clip. Official Arctic/HOT3D recordings and MANO models remain external.
20
 
21
- [Current manifest](agentic_pipeline/2026-10-07-rgbd-video/manifest.json) · [Checksums](agentic_pipeline/2026-10-07-rgbd-video/SHA256SUMS) · [Archive portability verification](agentic_pipeline/2026-10-07-rgbd-video/PORTABILITY.json)
22
-
23
- The original `pickcup_right/` inputs are preserved. Components retain their source licenses and provenance; see each release and bundled notices. The existing dataset license applies to the original data.
24
 
25
  <!-- /SPIDER_AGENTIC_RELEASE -->
 
6
 
7
  # SPIDER video and data workflows
8
 
9
+ [**Watch the corrected pickup and release**](agentic_pipeline/2026-10-08-grasp-phases/pickcup_right/motion/independent_evaluation/comparison.mp4) · [Sequence contact sheet](agentic_pipeline/2026-10-08-grasp-phases/pickcup_right/motion/independent_evaluation/contact_sheet.jpg) · [Code/data and reproduction command](agentic_pipeline/2026-10-08-grasp-phases/README.md)
10
 
11
+ The RGB-D cup workflow now starts with an **open, separated hand** and a free cup resting on a measured table. It approaches, closes, establishes a physical grasp before lift, transports, restores table support, opens fully during release, and withdraws. The grasp checkpoint is a target, not the initial simulation state. The previous always-held cup recording is superseded for pickup/release behavior.
12
 
13
+ All **269 source frames** are represented at **quarter speed**, with **35,680 actual simulation steps**. Grasp establishment completes at **2.213 s**, before lift at **2.775 s**; support returns before release, and the final cup remains at rest. All **46 independent phase checks pass**, including hand/table clearance. Control replay matches saved positions and velocities exactly. Maximum position error against the declared execution reference is **14.61 mm**.
14
 
15
+ The update passed **204 tests**. Six shared Codex/Claude skills expose the workflow; invoke `$spider-video-to-motion` or `/spider-video-to-motion`. The example runs on CPU without MANO models. Inputs, source phase reviews, grasp targets, approach waypoints, physical trajectories/controls, video, support provenance and verification reports are included. Runtime logs remain excluded.
16
 
17
+ Depth scale, registration, level-table gravity, approximate axial cup geometry, mass and friction remain assumptions; 36 reconstructed poses are uncertain. Support-compatible reference corrections are recorded separately (up to 7.4 mm), and the original track is retained. Axial spin is unobserved; full quaternion errors remain reported. Execution uses full gravity, real finger/table contacts, no object actuation, no welds, no applied forces and no scored-state resets. This verifies simulated interaction, not hardware or liquid transfer.
18
 
19
+ [Current manifest](agentic_pipeline/2026-10-08-grasp-phases/manifest.json) · [Checksums](agentic_pipeline/2026-10-08-grasp-phases/SHA256SUMS) · [Portability check](agentic_pipeline/2026-10-08-grasp-phases/PORTABILITY.json) · [Phase evidence](agentic_pipeline/2026-10-08-grasp-phases/pickcup_right/motion/phase_evaluation/evaluation.json)
20
 
21
+ Earlier conversion examples and the [Whisking case](agentic_pipeline/2026-10-07-object-motion/README.md) remain available. The original `pickcup_right/` inputs are unchanged. Components retain their source licenses and provenance; the dataset license applies to original data. See release instructions and bundled notices for code, robot assets and upstream inputs.
 
 
22
 
23
  <!-- /SPIDER_AGENTIC_RELEASE -->
agentic_pipeline/2026-10-08-grasp-phases/PORTABILITY.json ADDED
@@ -0,0 +1,16 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "fresh_archive_extraction": true,
3
+ "all_phase_checks_passed": true,
4
+ "steps": 35680,
5
+ "arrays_bitwise_identical": {
6
+ "qpos": true,
7
+ "qvel": true,
8
+ "ctrl": true,
9
+ "time": true,
10
+ "phase": true
11
+ },
12
+ "independent_replay_max_qpos_error": 0.0,
13
+ "source_commit": "10c4197",
14
+ "dependencies": "Existing Python environment; code/data loaded from the extracted archive",
15
+ "scope": "Fresh phased simulation and independent replay from portable targets/plans. workflow.json separately records the complete raw-frame execution."
16
+ }