ChaoyiPan commited on
Commit
9129406
·
verified ·
1 Parent(s): 2c61b9e

Publish SPIDER agentic pipeline materials: agentic_pipeline/2026-10-08-grasp-phases

Browse files
agentic_pipeline/2026-10-08-grasp-phases/README.md ADDED
@@ -0,0 +1,51 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # SPIDER: actual pickup, grasp establishment, placement and release
2
+
3
+ [**Watch the corrected comparison video**](pickcup_right/motion/independent_evaluation/comparison.mp4) · [Sequence contact sheet](pickcup_right/motion/independent_evaluation/contact_sheet.jpg) · [All phase checks](pickcup_right/motion/phase_evaluation/evaluation.json) · [Result summary](pickcup_right/RESULT.json)
4
+
5
+ The robot begins **open and separated** while the free cup rests on a physical tabletop measured from RGB/depth. It approaches, closes, establishes a physical grasp before lifting, carries the cup, restores table support before release, opens fully while hovering, and withdraws. The cup then remains on the table. A saved grasp is a target, not the initial simulation state.
6
+
7
+ The complete **269-frame** clip executes at **quarter speed**, covering **35,680 actual physics steps** over 35.68 seconds. In original source time, grasp establishment completes at **2.213 s**, lift begins at **2.775 s**, sustained table support returns by **7.090 s**, and the grasp is released by **7.568 s**. The hand is fully open during the release hover. All **46 independent phase checks pass**, including physical/visual table clearance and final rest. Replaying the controls reproduces every saved position and velocity exactly.
8
+
9
+ Maximum position error against the declared execution reference is **14.61 mm**. Full gravity is used throughout; no object actuator, weld, applied force or scored-state reset assists the motion. The controller blocks lift without an established grasp and blocks release without measured support load. Multiple contact detection represents the flat cup base and prevents the previous single-point rocking.
10
+
11
+ ## Download and reproduce
12
+
13
+ From an empty directory:
14
+
15
+ ```bash
16
+ BASE=https://huggingface.co/datasets/retarget/example_video_data/resolve/main/agentic_pipeline/2026-10-08-grasp-phases
17
+ for FILE in source.tar.gz datasets.tar.gz manifest.json SHA256SUMS README.md; do
18
+ curl --fail --location --retry 3 "$BASE/$FILE" --output "$FILE"
19
+ done
20
+ sha256sum --check SHA256SUMS
21
+ tar -xzf source.tar.gz
22
+ tar -xzf datasets.tar.gz
23
+ cd spider
24
+ uv sync --frozen --extra dev
25
+
26
+ MUJOCO_GL=egl OPENBLAS_NUM_THREADS=1 uv run python -m spider.video_rgbd \
27
+ --raw-dir example_datasets/video_examples/source_rgbd/pickcup_right \
28
+ --output-dir example_datasets/video_workflows/my_pickup_run \
29
+ --depth-scale 0.001 --mass-kg 0.05 \
30
+ --azimuth-deg -90 --height-fraction 0.15 --time-scale 4 \
31
+ --interaction-review-json example_datasets/video_workflows/pickcup_phased_final/interaction/review.json \
32
+ --allow-uncertain-tracking --render
33
+ ```
34
+
35
+ Use a fresh output directory. No GPU or MANO models are needed for this RGB-D example. The reviewed event file is tied to the original source hashes. For another source, inspect the generated interaction proposals and event images before supplying its own review. Both the independent phase report and object-tracking report must pass; perfect tracking alone does not establish a pickup.
36
+
37
+ Open Codex or Claude Code in the extracted checkout and invoke `$spider-video-to-motion` or `/spider-video-to-motion`. Six shared skills cover pipeline routing, video processing, conversion, retargeting, tuning and visualization. Keep Claude's symlinks. Full stage commands and input contracts are in `docs/usage/agentic-pipeline.md`.
38
+
39
+ ## Evidence and scope
40
+
41
+ `pickcup_right/` contains source interaction evidence, approximate geometry, a closing target, robot-only approach waypoints, the actual open initial state, physical trajectory/controls, measured-support provenance, independent replay, and the comparison video. `workflow.json` records a fresh execution from the original frames. The code update passed **204 tests**, and all six skills validated. Runtime logs are excluded.
42
+
43
+ The original RGB-D reconstruction still includes **36 uncertain poses**. Depth scale, registration, level-table gravity, approximate axial cup geometry, mass and friction remain declared assumptions. Supported start/end conditions require explicit reference corrections, up to **7.4 mm**, retained alongside the unchanged original estimate. Axial spin is unobserved; full quaternion errors are reported while explicit axial symmetry gates the observed cup-axis orientation. Initial settling happens with the open hand separated and is saved separately; acquisition occurs within the scored source timeline. This is simulation verification, not hardware or liquid-transfer certification.
44
+
45
+ The older always-held cup result is marked **superseded for pickup/release behavior**. Its historical air-tracking metrics did not verify these interaction phases. The earlier Whisking example remains available for an already-held source; it must not be used as a shortcut for pickup clips.
46
+
47
+ The archives share the `spider/` root and retain the prior conversion examples and valid Whisking inputs for existing commands. See `manifest.json` for exact directories, source revision, sizes and hashes. Original `pickcup_right/` RGB/depth data at the dataset root is unchanged. Logs were also removed from the previous published archives; configs, trajectories and verification reports remain.
48
+
49
+ ## Provenance and licenses
50
+
51
+ SPIDER code retains its repository license (CC BY-NC 4.0). Do-as-I-Do ports and Sharpa assets retain bundled MIT notices and upstream revision `824591b808c342b20079c3b4198a8c2bdf88c74e`. Public conversion inputs retain their dataset terms. The cup RGB/depth source is `retarget/example_video_data/pickcup_right`, with its original license. Model weights are not redistributed.
agentic_pipeline/2026-10-08-grasp-phases/SHA256SUMS ADDED
@@ -0,0 +1,4 @@
 
 
 
 
 
1
+ b51a85bb4efa569333899e5d94ac6bded55c92f9df0751f6818c7d868352e132 source.tar.gz
2
+ a7b58fa3aaba0892efab537766ef6234e94591f0314a50d9f7d6e894389912ad datasets.tar.gz
3
+ 0b8f227adcbb7fcb0a86b4d6b8adebb141027affd2baef590b38d02107e0bf6c manifest.json
4
+ da4755698dda23cd0bd538937027febff6bcf9977baa72babb1d4bd1119a8e00 README.md
agentic_pipeline/2026-10-08-grasp-phases/datasets.tar.gz ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:a7b58fa3aaba0892efab537766ef6234e94591f0314a50d9f7d6e894389912ad
3
+ size 1045145713
agentic_pipeline/2026-10-08-grasp-phases/manifest.json ADDED
@@ -0,0 +1,48 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "schema_version": 1,
3
+ "created_at": "2026-10-08T02:16:54.533793+00:00",
4
+ "source_commit": "10c4197cf5eb53e61b8d17e13aaf6b94a58284f9",
5
+ "source_has_task_changes": false,
6
+ "archives": {
7
+ "source.tar.gz": {
8
+ "sha256": "b51a85bb4efa569333899e5d94ac6bded55c92f9df0751f6818c7d868352e132",
9
+ "bytes": 183521137,
10
+ "files": 956
11
+ },
12
+ "datasets.tar.gz": {
13
+ "sha256": "a7b58fa3aaba0892efab537766ef6234e94591f0314a50d9f7d6e894389912ad",
14
+ "bytes": 1045145713,
15
+ "files": 4332
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/video_workflows/pickcup_phased_final",
25
+ "example_datasets/video_workflows/pickcup_phases/source_phases"
26
+ ],
27
+ "runtime_logs_excluded": true,
28
+ "notes": [
29
+ "Source archive includes current task changes and both agents' skills.",
30
+ "Neural model weights and MANO assets are not included; precomputed reconstruction is included.",
31
+ "Execution receipts retain original paths as provenance; use a fresh output directory for new runs.",
32
+ "Raw terminal/runtime logs, rotated logs, log directories, and stdout/stderr captures are excluded."
33
+ ],
34
+ "validation": {
35
+ "tests_passed": 204,
36
+ "skills_validated": 6,
37
+ "phase_checks_passed": 46,
38
+ "source_frames": 269,
39
+ "simulation_steps": 35680,
40
+ "source_time_scale": 4,
41
+ "open_separated_initial_state": true,
42
+ "actual_grasp_establishment": true,
43
+ "physical_placement_before_release": true,
44
+ "fully_open_release_and_withdrawal": true,
45
+ "independent_replay_exact": true,
46
+ "source_reference_is_ground_truth": false
47
+ }
48
+ }
agentic_pipeline/2026-10-08-grasp-phases/source.tar.gz ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:b51a85bb4efa569333899e5d94ac6bded55c92f9df0751f6818c7d868352e132
3
+ size 183521137