Hugging Face's logo

Datasets:
daneel-data
/
factory-manipulation-sample-v3

Tasks:
Robotics
Other
Modalities:
Video
Size:
< 1K
Tags:
egocentric
stereo
hand-pose
manipulation
footwear-assembly
robotics
Libraries:
Datasets
License:
Dataset card Data Studio Files
xet
 Community
1

You need to agree to share your contact information to access this dataset

This repository is publicly accessible, but you have to accept the conditions to access its files and content.

This is a private evaluation sample. By requesting access you agree to use it for non-commercial research only and not to redistribute it (CC BY-NC-ND 4.0).

Log in or Sign Up to review the conditions and access this dataset content.

factory-manipulation-sample-v3

Egocentric stereo factory hand-manipulation sample for robotics research. Four ~3-minute clips, each a continuous manipulation task captured from a head-mounted stereo RGB camera, with metric 3D 21-keypoint hand pose, head IMU, stereo calibration, and a natural-language task description. No faces are present (no_faces_present).

Packaged as a single archive factory-manipulation-sample-v3.zip — download and extract: hf download daneel-data/factory-manipulation-sample-v3 factory-manipulation-sample-v3.zip --repo-type dataset A 3-second preview (v3_preview_clip001_3s.mp4) is included.

Preview

3-second preview (clip_001): egocentric stereo with metric-3D 21-keypoint hand pose (blue=Left, green=Right). Inline playback requires dataset access (gated).

Per-clip files

  • left.mp4, right.mp4 — synchronized stereo RGB, 2000×1200 per eye @ 60 fps, H.264. Upright.
  • handpose.jsonl — per-frame metric-3D 21-keypoint hand pose (Left/Right), one object per frame. See handpose_README.md for the schema.
  • handpose_overlay.mp4 — left eye with the 21-keypoint skeleton drawn (blue=Left, green=Right) + depth (QC aid).
  • calibration.json — stereo calibration: per-eye intrinsics + distortion, stereo extrinsics (R/T, 60 mm baseline), and rectification (R1/R2/P1/P2/Q). Used to recover the metric 3D. Shared across clips (same rig).
  • head_imu.csv — head IMU (6-axis), ~600 Hz, on the t_ms master clock. Columns t_ms, video_frame, ax, ay, az, gx, gy, gz (accel m/s², gyro rad/s), rotated to the upright frame.
  • frame_index.jsonl — per-frame exposure timestamps (re-based to 0), the video↔IMU alignment key.
  • language_annotation.json — action label + natural-language task description.

Notes

  • Metric 3D: use keypoints_3d_m from source == "stereo" hands. mono hands (one-eye-only) carry 2D + per-eye wrist-relative 3D but no metric depth.
  • Occlusion is honest — hands occluded in both eyes are absent for that frame, never fabricated.
  • Known gaps: the camera↔IMU extrinsic is not available, so the IMU coordinate frame is the documented assumption (see head_imu); clip_004's IMU/frame_index sidecars are short on the final ~4 s (the video and hand pose cover all frames).

License: CC BY-NC-ND 4.0.
Downloads last month
5
Total file size:
2.95 GB