Datasets:

---

robbyant

/

robotwin-clean-and-aug-lerobot

like 11

Follow

Robbyant 275

Summary

Some video files in this dataset fail during full decode — decoding starts successfully but fails partway through the bitstream. This suggests possible encoding/container corruption or truncation for affected files, not only a client-side decoder bug.

Example

• File (HF hub):
  lerobot_robotwin_eef_aug_500/.../episode_000003.mp4
  (observation.images.cam_left_wrist / episode_000003.mp4)

Observed behavior

• FFmpeg (full demux + decode to null) and/or PyTorch torchcodec VideoDecoder report errors when decoding beyond the first segment of frames, e.g.:
  • Invalid data found when processing input
  • Could not push packet to decoder (torchcodec / FFmpeg)

Training pipelines that decode all frames (e.g. chunked slicing over the full index range) will crash when they hit the bad segment, even if the file opens and the first frames decode fine.

Expected behavior

Each *.mp4 should decode from first to last frame without FFmpeg/torchcodec errors.

Suggested checks (for maintainers)

1. Run a full decode on the file (not only ffprobe / first frame):
   • ffmpeg -v error -i <file> -f null -
2. Optionally batch-scan the dataset with a chunked full decode to list all failing paths.

Environment (optional)

• Decoder: FFmpeg / torchcodec (VideoDecoder) with CPU decode
• Issue is data-dependent (same code succeeds on other episodes in the same layout).

Request

Could you verify this asset and, if confirmed bad, re-encode or re-upload affected videos (or publish a manifest of known-bad files)? That would unblock downstream training that assumes decodable video for every frame index.

Thank you