Five hours of first-person bimanual manipulation with per-frame 3D hand pose (MANO), wrist pose, camera intrinsics and extrinsics, packaged as a LeRobot v2.1 dataset. Ready to load into a training script.
Visualization renders: the delivered 3D hand data projected back onto the ego video. If the overlay tracks the hands, the data is right.
Episode 0 — wrist axes, fingertips and task HUD projected onto the ego video
Episode 16 — second camera configuration (fx≈871)
| Episodes | 390 |
|---|---|
| Frames | 541,192 at 30 fps (5.0 h) |
| Tasks | 231 distinct task strings |
| Video | 1440×1920 egocentric RGB, H.264, one mp4 per episode |
| Episode length | 150 – 21,045 frames, median 874 (~29 s) |
| Size | 22.3 GiB · 793 objects |
| Format | LeRobotDataset v2.1: data/*.parquet · videos/ · meta/ · calibration/ · visualization/ |
231 task strings across the 390 episodes. Verb families: wipe (99), peel (35), thread (26), assemble (24), apply (12), fold (7), attach (4), pack (4).
Per-frame columns in data/*.parquet. Lengths in metres, angles in radians, matrices row-major. Full semantics and coordinate frames are in ACTION_DATA_DESCRIPTION.md inside the dataset.
| observation.images.ego | video [1440,1920,3] egocentric RGB |
|---|---|
| observation.state | float64 [122] left 61 + right 61: wrist pos (cam), root rot, 15 joint eulers, MANO betas |
| left/right_transl_world | float64 [3] wrist translation, world frame, metres |
| left/right_orient_world | float64 [9] wrist rotation 3×3 row-major |
| left/right_hand_pose | float64 [135] 15 MANO joint rotations, 3×3 each |
| left/right_per_frame_validity | float64 [1] tracking validity per frame |
| state_mask | bool [2] which hands present |
| intrinsics / extrinsics_w2c / fov | float64 [9] / [16] / [2] per-frame camera model, OpenCV convention |
Delivered from an S3-compatible bucket. Credentials are issued per client, sent separately, and expire; one command pulls everything and resumes if interrupted.
aws s3 sync s3://<bucket>/<prefix>/ ./ego-hand-5h/ --endpoint-url <endpoint> --profile realset
This is a 5-hour sample. Formal batches follow the same LeRobot layout at the task mix, environment and volume you specify.
Three questions. We reply within one business day with a sample set and a quote.