Datasets:

---

BitRobot

/

HIW-500-LeRobot

Like 24

Follow

BitRobot 65

Hi, and thanks for releasing HIW-500 and its LeRobot conversion. While cataloguing the dataset I compared converted episodes with the original MCAP recordings in BitRobot/HIW-500, and found that timestamp (and fps: 30 in meta/info.json) does not match real time for many episodes.

What's wrong. Each converted row corresponds to one head-camera image, and rows are stamped 1/30 s apart. In many recordings from mid-April 2026 onward the head camera delivered images at about 11–22 Hz, so those episodes are compressed in time.

Evidence (HIW-500-LeRobot@2d32431, HIW-500@c35830b)

• Episode 17776 vs Move-The-Pillow-To-The-Sofa-From-Floor/episode_2026-05-26_12-59-54/episode_0007/episode_0007.mcap: the MCAP has 458 /camera/head/image/compressed messages over 28.31 s (median interval 62 ms, ~16 Hz); the wrist cameras have 850 images (30 Hz) over the same span. The converted episode has 458 rows, last timestamp 15.23 s. observation.state and action equal /stamped/lowstate and /wbc_lerobot nearest each head image, bit for bit, so the values are right and only the time axis is wrong.
• Episode 6598 vs Sweep-Floor/episode_2026-05-19_15-39-51/episode_0009/episode_0009.mcap: 467 rows = 15.5 s at 30 fps, against 43.3 s recorded (head camera ~10.8 Hz).
• How common: in the MCAP summaries of 160 episodes (16 per task, all 10 tasks), the head-camera rate has median 28 Hz, 10th percentile 14.7 Hz, minimum 10.8 Hz; 46 % are below 27 Hz. All sampled January–March recordings are ~30 Hz; from about 2026-04-13 on, 73 of 114 sampled episodes are slower. Some switch between ~30 and ~16 Hz partway through. Wrist cameras stay at 30 Hz, joint state at 100 Hz, /wbc_lerobot at 50 Hz.
• Every one of the 160 sampled MCAPs has a converted episode with exactly as many rows as head images; converted/recorded duration has median 0.94 and minimum 0.36.
• Release-wide, over the 10 tasks both releases share, converted rows/30 sum to 294.0 h against 385.6 h of recordings (info.json duration_sec).
• Control: a 30 Hz January episode (8865 vs Hang-Keys-On-A-Hook/episode_2026-01-30_15-48-55/episode_0007) converts correctly: 66.97 s vs 67.06 s.
• To reproduce: compare the head topic's channel_message_counts and message start/end times in an MCAP summary with that episode's length and length / 30.

Impact. In affected episodes motion appears up to ~2–3x faster than it was. Anything that uses timestamp/fps is off: per-step velocities, fixed-horizon action chunks, delta_timestamps, playback, and training that assumes a fixed control rate. The wrist videos are also one frame per row, so they are subsampled to the head rate and compressed the same way.

Suggested fix (either):

• Resample every stream onto a fixed 30 Hz grid using the recorded message times (nearest or most recent head image and state at each 1/30 s tick), so fps: 30 holds; or
• keep one row per head image but set timestamp from each image's recorded time, and document that the rate varies.

A README note that the head-camera rate varies (30 Hz early, often ~16 or ~11 Hz later) would also help. Happy to share the script I used. Thanks!