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.

Access is provided for evaluation only. Please do not redistribute this sample dataset.

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

factory-manipulation-sample-v2

Egocentric factory hand-manipulation sample for robotics research. Three ~60-second clips, each a continuous, idle-trimmed manipulation task captured from a head-mounted camera, with timestamped 21-keypoint hand pose, head IMU, and a natural-language description. Private evaluation sample — not for redistribution.

Packaged as a single archive:

hf download daneel-data/factory-manipulation-sample-v2 factory-manipulation-sample-v2.tar.gz --repo-type dataset --local-dir .
tar -xzf factory-manipulation-sample-v2.tar.gz   # -> factory-manipulation-sample-v2/{clip_001,clip_002,clip_003}/

Per-clip files

  • video.mp4 — RGB egocentric video, 1640×1232, 30 fps (PII-safe: faces blurred where present).
  • video_handpose.mp4 — same video with the 21-keypoint hand skeleton drawn + a QC HUD.
  • handpose.jsonl — per-frame hand pose. HaMeR (MANO), one JSON object per frame. Each hand: COCO-style keypoints [x,y,v] (v: 2=visible, 1=low-confidence/inferred), keypoints_world_m + keypoints_cam_m (3D, meters), mano (axis-angle pose/shape), bbox (COCO xywh), per-hand valid/flags, and per-frame frame_flags. See handpose_README.md for the full schema and the accuracy/validity flags.
  • handpose_qc.csv — per-frame quality flags for quick filtering.
  • handpose_README.md — annotation data card (schema + how to use the flags).
  • head_imu.csv — head IMU (accel + gyro), ~200 Hz, on the same continuous timeline. Columns: t_ms (continuous ms), accel_{x,y,z} (g), gyro_{x,y,z} (deg/s). For SI: accel × 9.80665 → m/s², gyro × π/180 → rad/s. Frame: see intrinsics.jsonhead_imu.
  • intrinsics.json — shared fisheye camera calibration (1640×1232: fx/fy/cx/cy, k1–k4) and the camera + head-IMU coordinate-frame conventions. Same calibration for all clips.
  • language_annotation.json — action label + natural-language task description.
  • metadata.json — clip-level metadata (action, fps, resolution, modalities, PII status).

License

CC BY-NC-ND 4.0 — evaluation and citation only; no redistribution or commercial use. Full corpus & licensing: https://www.daneeldata.com

Downloads last month
14