Long-horizon units: both ego views on one clock, the instruction a reviewer approved, the action steps with what each hand is doing, the moments where the target and the grasp point are grounded on the frame, and the recorded end-effector state underneath.
—
—
—
—
—
—
—
—
The 53 folding episodes and both OpenArm policy rollouts, untouched: original videos, Parquet, IMU and the affordance grounding, with the LeRobot directory structure preserved. These are sample captures — the long-horizon units on the first tab come from a different export and are not in this archive. The complete captures are quoted separately.
—01 · Request the code — we email it to you
02 · Enter the code
Access is by code. Requests are kept with the name and affiliation given, and every download is recorded — time, IP address, country and user agent — so we can see how the batch is being used.
Our challenge dataset for the IROS 2026 Bimanual Manipulation Workshop: handheld
bimanual demonstrations recorded in real homes, unfolding, flattening and folding garments. Log in to
Hugging Face and accept the terms to download.
Crobotia/iUMI
Each episode is its own LeRobot dataset, packed into tar shards of about 3 GiB on whole-episode boundaries. episodes.csv lists every episode with its instruction, duration and shard, so you can fetch one shard and start.
| In every episode | Detail |
|---|---|
| Video | Left and right fisheye views, 960×960, 30 fps, H.264 |
| State / action | 16-D per frame: position, quaternion and jaw angle for each hand |
| IMU and audio | About 1 kHz six-axis IMU and 48 kHz stereo audio per side |
| Annotations | Task intervals, action-step intervals and keyframe visual grounding |
| Calibration | Per-body camera intrinsics, camera–IMU extrinsics and IMU biases |
A real report from our internal QC system on a 1,000-hour folding set, and the training loss of a π0.5 policy fine-tuned on our folding data. Both are shown as produced, not mocked up.
17 checks over structure, timing, video and motion run on every episode. Each check either blocks the episode, flags it for a person to review, or only records what it found. 9 episodes hit an execution error and are queued to re-run; until they are, the report marks itself incomplete instead of rounding up.
| What it found · examples | Where | Outcome |
|---|---|---|
| video_black_framesfold the gray shirt · 47.4 s | 0.0 – 32.3 s · cam L | blocked — not delivered |
| video_frozen_runfold the beige pants · 43.6 s | 26.0 s · cam L | same picture for over 1 s while the rig was moving — sent to review |
| video.edge_stillfold the black pants · 38.1 s | 35.0 – 38.1 s | still tail — trimmed to the valid range |
| motion.edge_idlefold the denim overalls · 76.1 s | 0.0 s → | idle head candidate — sent to review |
| Check | Rule | Flagged |
|---|---|---|
| Structure & integrity | ||
| inventory.integrity | blocks | 0 |
| episode_too_short | review | 0 |
| Timing | ||
| time.monotonic | review | 0 |
| Video | ||
| missing_camera | blocks | 0 |
| video_decode_failure | blocks | 0 |
| video_decode_frame_count_mismatch | blocks | 0 |
| video_black_frames | blocks | 8 |
| video_white_frames | blocks | 0 |
| video_frozen_run | review | 62 |
| video_duplicate_frames | review | 0 |
| video_extreme_exposure | review | 0 |
| video_overexposed_frames | review | 0 |
| video_flat_frames | review | 0 |
| video_low_contrast | record | 0 |
| video.edge_still | record | 44 |
| Motion & pipeline | ||
| motion.edge_idle | review | 488 |
| signal.events | record | 0 |
Smoothed training loss (moving average over ~130k samples), log scale. One run of 6 epochs on 56 GPUs, then two continuations of 2 epochs each; the second adds 141 real-robot episodes at double weight. Each continuation restarts the learning rate at 7.5e-5, hence the short spike; grey bands are the annealing phases.
—