1. Issues

16 video files contain invalid AV1 packets. They are partially readable but decoding eventually fails mid-stream on an invalid packet. Affected files are shown as follow. decoded/total is the number of frames TorchCodec 0.11.1 read before failing.

Dataset version:

• repo: lerobot/droid_1.0.1 (dataset), codebase_version v3.0
• revision: bd92a2c4d6c85f9e16bd7c1f9729f5496d8be672
• video features: 180×320, fps 15, cameras observation.images.{wrist_left, exterior_1_left, exterior_2_left}

2. Reproduction

The failure reproduces across PyAV, TorchCodec and the FFmpeg CLI, on both the FFmpeg 7 and FFmpeg 8 library series:

Reference reproduction environment (PyAV path): Python 3.12.12, torch==2.11.0+cu128, torchvision==0.26.0+cu128, torchcodec==0.11.1, av==15.1.0, imageio-ffmpeg==0.6.0 (PyAV-bundled libavcodec 61.19.101 / libavformat 61.7.100).

Minimal PyAV reproduction:

import av
from huggingface_hub import hf_hub_download

repo_id = "lerobot/droid_1.0.1"
revision = "bd92a2c4d6c85f9e16bd7c1f9729f5496d8be672"
filename = "videos/observation.images.wrist_left/chunk-000/file-033.mp4"

path = hf_hub_download(repo_id=repo_id, repo_type="dataset", revision=revision, filename=filename)

with av.open(path) as container:
    stream = container.streams.video[0]
    for _ in container.decode(stream):   # fails partway through
        pass

Observed error:

av.error.InvalidDataError: [Errno 1094995529] Invalid data found when processing input: 'avcodec_send_packet()'

Minimal TorchCodec reproduction:

from huggingface_hub import hf_hub_download
from torchcodec.decoders import VideoDecoder

repo_id = "lerobot/droid_1.0.1"
revision = "bd92a2c4d6c85f9e16bd7c1f9729f5496d8be672"
filename = "videos/observation.images.wrist_left/chunk-000/file-033.mp4"

path = hf_hub_download(repo_id=repo_id, repo_type="dataset", revision=revision, filename=filename)

decoder = VideoDecoder(path, device="cpu", seek_mode="exact", dimension_order="NHWC")
for start in range(0, len(decoder), 512):
    end = min(start + 512, len(decoder))
    decoder.get_frames_at(list(range(start, end)))   # fails on the bad GOP

Observed error:

Could not push packet to decoder: Invalid data found when processing input

3. Fix

3.1 Source of replacement frames. The original RLDS DROID 1.0.1 dataset (gs://gresearch/robotics/droid/1.0.1, TFDS, 2048 tfrecord shards, 95 658 episodes).

3.2 Index the corresponding episode in the RLDS.

the LeRobot episode_index does not match the RLDS canonical (shard) order — the dataset build reordered episodes. So a naive index → index lookup is wrong.

We instead match episodes with the lossless actions: every LeRobot frame's action.joint_position equals exactly one RLDS step's action_dict/joint_position in the same episode. For each target episode, we read the RLDS in canonical order, match a distinctive middle-frame joint key, and confirm with a full subsequence check:

• action: per-frame joint_position max-abs-diff ≈ 0;
• episode length: RLDS step count T == LeRobot episode length L;
• pixels: decode the original LeRobot video and compare to the RLDS frames — aligned PSNR ≈ 37.5 dB (the expected AV1 loss)

All 7753 episodes matched (~2.26 M frames total), every one with an identity step mapping (L == T, contiguous)

3.3 Encode the correct frames back to video
For each file, in episode order, the matched RLDS JPEG frames are decoded to RGB and re-encoded into a single .mp4 with libsvtav1, yuv420p, fps 15, g=2, crf=30, matching the rest of the dataset. Each episode's exact frame count and order is preserved, so the existing meta/episodes parquet from/to timestamps stay valid and no metadata is changed. Re-encode quality is PSNR 36.3–41.0 dB, on par with the originals (35.5–40.0 dB).

4. Verification

Passes on all 16 updated files (no InvalidDataError, exact frame count):

import av

def full_decode(path: str) -> int:
    """Return the number of decoded frames; raises on any invalid packet."""
    n = 0
    with av.open(path) as container:
        for _ in container.decode(container.streams.video[0]):
            n += 1
    return n

# Each of the 16 updated files now decodes completely, e.g.:
#   videos/observation.images.wrist_left/chunk-000/file-033.mp4 -> 147049 (== total)
#   videos/observation.images.wrist_left/chunk-000/file-114.mp4 -> 150721 (== total)
for rel, total in AFFECTED_FILES:           # (path, expected frame count)
    assert full_decode(local_path(rel)) == total