seungkukim/dexgys_video224_a0_lerobot

dexgysvideo224a0_lerobot

A LeRobot-format render of the DexGYS language-conditioned static dexterous-grasping dataset (Shadow Hand), packaged as a single zip.

Contents

huggingface-cli download seungkukim/dexgys_video224_a0_lerobot --repo-type dataset --local-dir .
sha256sum -c dexgys_video224_a0_lerobot.zip.sha256
unzip dexgys_video224_a0_lerobot.zip          # -> ./train/

meta/, data/ (parquet), videos/ (mp4) and conditioning/ (the frame-0 PNG per episode).

Facts a consumer needs, because none of them fails loudly

1. 1.One episode is ONE transition, not a trajectory. Each episode is a single unique grasp over 9 frames: frame 0 is the object alone (the image-to-video conditioning frame) and the remaining frames are the same static grasp with the camera orbiting 360 degrees. action and action.keypoint are byte-identical across every row of an episode. A loader that treats an episode as a trajectory and samples step_idx > 0 slides frame 0 off the front of the video window while keeping a correct-looking label.

1. 1.Nothing is rotated. Rendered views: 1 at azimuth 0, so azimuth and scene_rotation are 0 (-0.0) on every row and the canonical object frame IS the frame the video shows. There is no view arithmetic to undo anywhere.

1. 1.`action.keypoint` is 63-D = 21 keypoints x 3, row-major, in the order meta/info.json names them. The six kinematic chains are contiguous: palm [0:15] (palm centre + the four finger roots — the rigid group a wrist pose is registered from), ff [15:24], mf [24:33] (there is no mf_root), rf [33:42], lf [42:54] (four points, because of lf_mcp), th [54:63]. The 28-D action column is the original [translation(3) | axis-angle(3) | qpos(22)] grasp.

1. 1.`point_cloud_index` is a join key into `meta/point_clouds.jsonl`, which is dense (point_cloud_index == line number). It indexes objects, not episodes — many grasps share one object. The object point clouds themselves are NOT in this archive; they come from the DexGYS release.

1. 1.The episode index is the dense line number in `meta/episodes.jsonl`. Subsetting that file makes every episode resolve to a different one.

1. 1.meta/modality.json is included: it maps the six anatomical action groups onto the flat action.keypoint column and names point_cloud_index as the state group's join key.

Provenance

Built by dexter's turntable renderer; meta/dexgys_info.json carries the exact render_config (size, ssaa, frames, views, arc, elev, focus_margin, shadowsize, caption template) the clips were produced with. The caption is "<scene description>. <instruction>" — note the period, which separates the two.

stats.json complete (mean/std/min/max/q01/q99): True.