ChaoyiPan commited on
Commit
6a93a77
·
verified ·
1 Parent(s): 8f6131b

Publish SPIDER agentic pipeline materials: agentic_pipeline/2026-10-07-rgbd-video

Browse files
agentic_pipeline/2026-10-07-rgbd-video/README.md ADDED
@@ -0,0 +1,56 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # SPIDER RGB-D video workflow — 2026-10-07
2
+
3
+ [**Watch the cup retarget**](pickcup_right/motion/comparison.mp4) · [Result and evidence](pickcup_right/README.md) · [Tracking quality](pickcup_right/tracking/tracking_quality.png)
4
+
5
+ The new CPU workflow starts from registered RGB/depth frames and camera intrinsics, estimates cup geometry and motion, generates a Sharpa grasp, retargets the whole estimated trajectory, and exports comparison video and verification reports. No MANO or GPU is needed for this example. Six repository skills expose conversion, video workflows, retargeting, tuning, and visualization to Codex and Claude Code.
6
+
7
+ The fresh cup run reproduces **269 reference poses / 17,840 physics steps** at **quarter speed** (8.92-second source → 35.68-second execution), with **12.55 mm maximum position error**, **8.68° rotation error**, and **99.98% loaded contact**. Independent control replay agrees exactly. Object motion and stable robot grasp take priority over human hand pose.
8
+
9
+ **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. Depth scale, registration, level-table gravity, approximate cup shape, mass, and contact parameters have 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. The scene has no supporting table, object actuation, external force, or equality constraint. Original-speed execution failed.
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/2026-10-07-rgbd-video
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 "$FILE"
19
+ done
20
+ sha256sum --check SHA256SUMS
21
+ tar -xzf source.tar.gz
22
+ tar -xzf datasets.tar.gz
23
+ cd spider
24
+ uv sync --frozen --extra dev
25
+
26
+ MUJOCO_GL=egl OPENBLAS_NUM_THREADS=1 uv run python -m spider.video_rgbd \
27
+ --raw-dir example_datasets/video_examples/source_rgbd/pickcup_right \
28
+ --output-dir example_datasets/video_workflows/my_pickcup_run \
29
+ --depth-scale 0.001 --mass-kg 0.05 \
30
+ --azimuth-deg -90 --height-fraction 0.08 \
31
+ --time-scale 4 --grip-normal-preload-N 0 \
32
+ --allow-uncertain-tracking --render
33
+ ```
34
+
35
+ Use a fresh output directory. Without the uncertainty flag the workflow stops at `requires_tracking_review`. Inspect the pose-quality evidence before choosing a labeled diagnostic. Zero additional motion preload retains the initializer's selected grasp preload.
36
+
37
+ Open Codex or Claude Code in the extracted checkout. Invoke `$spider-video-to-motion` or `/spider-video-to-motion` for the RGB-D example. The other skills are `spider-pipeline`, `spider-convert`, `spider-retarget`, `spider-tune`, and `spider-visualize`. Keep the Claude skill symlinks. Detailed commands are in `docs/usage/agentic-pipeline.md`.
38
+
39
+ ## Contents and validation
40
+
41
+ Both archives share the `spider/` root. They include current source and Sharpa assets; RGB, depth, intrinsics, masks and reconstructed motion; portable scenes and grasp certificates; actual trajectories, controls, video and uncertainty/clearance reports. The earlier conversion examples and verified quarter-speed Whisking case are retained for the existing skill commands. See `manifest.json` for exact included directories.
42
+
43
+ The RGB-D update passed **135 tests** and all six skill definitions validated. A fresh end-to-end run reproduced the selected states, velocities, controls and times bit for bit. Independent replay verified all 17,840 robot-control steps. The comparison video's scene and trajectory equivalence are recorded separately.
44
+
45
+ Runtime logs are excluded from this release. The two older data archives were also cleaned (30 log entries each), and two exposed log files were removed. Configs, trajectories, metrics, provenance and failure evidence remain. The original `pickcup_right/` input data at the repository root is preserved.
46
+
47
+ The 40 public conversion examples still cover GigaHands, OakInk, OakInk-v2, FAIR-FRE and FAIR-MON with 11,626 Sharpa IK frames; that is not a claim that all 40 manipulate successfully. Arctic/HOT3D official recordings and MANO models remain external. Monocular neural reconstruction still uses its separate upstream environments and model dependencies.
48
+
49
+ ## Provenance and licenses
50
+
51
+ - SPIDER source: repository `LICENSE` (CC BY-NC 4.0).
52
+ - Do-as-I-Do ports and Sharpa assets: bundled MIT notices and source commit `824591b808c342b20079c3b4198a8c2bdf88c74e`.
53
+ - Public conversion inputs: `retarget/retarget_example` revision `8b0cdf2b24cc86d1c859050c5b0317ddbec59a63`, with original dataset terms.
54
+ - Original RGB/depth cup input: `retarget/example_video_data/pickcup_right`, retaining its original license.
55
+
56
+ Archive hashes, byte counts and source revision are in `manifest.json` and `SHA256SUMS`. Execution receipts retain original locations for provenance; use the documented portable paths when reproducing.
agentic_pipeline/2026-10-07-rgbd-video/SHA256SUMS ADDED
@@ -0,0 +1,4 @@
 
 
 
 
 
1
+ 4a96ba34b0409ea39f81e9c3e1a6d63a1e90faf43a697e29c5fcc3806269ecc3 source.tar.gz
2
+ b7d51d93344a5a975ec2c3bf8a68dd17c7c784d7c1a9f38972547da10956e863 datasets.tar.gz
3
+ 3d38c8c98b78c7dea82babc92c8831d9038673d2da8b013dc910b71de5497b19 manifest.json
4
+ a2f2f97210265dbf5000083cee5924f78c728c72e58f2b5b82450d7e4e6f5a29 README.md
agentic_pipeline/2026-10-07-rgbd-video/datasets.tar.gz ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:b7d51d93344a5a975ec2c3bf8a68dd17c7c784d7c1a9f38972547da10956e863
3
+ size 1026160261
agentic_pipeline/2026-10-07-rgbd-video/manifest.json ADDED
@@ -0,0 +1,44 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "schema_version": 1,
3
+ "created_at": "2026-10-07T14:52:04.692567+00:00",
4
+ "source_commit": "60387ddaa3fdc68f5529f6043e0b80e5300134ca",
5
+ "source_has_task_changes": false,
6
+ "archives": {
7
+ "source.tar.gz": {
8
+ "sha256": "4a96ba34b0409ea39f81e9c3e1a6d63a1e90faf43a697e29c5fcc3806269ecc3",
9
+ "bytes": 183463997,
10
+ "files": 946
11
+ },
12
+ "datasets.tar.gz": {
13
+ "sha256": "b7d51d93344a5a975ec2c3bf8a68dd17c7c784d7c1a9f38972547da10956e863",
14
+ "bytes": 1026160261,
15
+ "files": 3817
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
+ "example_datasets/video_workflows/pickcup_frontdoor_verified"
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": 135,
36
+ "skills_validated": 6,
37
+ "full_fresh_rgbd_workflow_reproduced": true,
38
+ "independent_control_replay_exact": true,
39
+ "reference_is_raw_3d_ground_truth": false,
40
+ "table_clearance_verified": false,
41
+ "source_uncertain_frames": 36,
42
+ "runtime_logs_excluded": true
43
+ }
44
+ }
agentic_pipeline/2026-10-07-rgbd-video/source.tar.gz ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:4a96ba34b0409ea39f81e9c3e1a6d63a1e90faf43a697e29c5fcc3806269ecc3
3
+ size 183463997