lance-format/pusht-lerobot-lancedb-video

pusht-lerobot-lancedb-video (video blobs)

A `lerobot-lancedb`-format Lance copy of `lerobot/pusht`, using the video-blob layout: one row per frame in …video.lance (tabular only), plus one row per source mp4 in …video_videos.lance (raw mp4 bytes as Lance blob v2). Pixels are bit-exact with upstream.

Not to be confused with `lance-format/lerobot-pusht-lance`, which uses a different three-table schema and is not loadable with lerobot-lancedb. This repo is specifically the layout produced by lerobot-convert-to-lance-video.

Load

from lerobot_lancedb import LeRobotLanceVideoDataset

ds = LeRobotLanceVideoDataset(repo_id="lance-format/pusht-lerobot-lancedb-video")
print(len(ds))                                # 25_650 frames
print(ds[0]["observation.image"].shape)       # torch.Size([3, 96, 96])

Decode is on-the-fly via torchcodec against the inline mp4 bytes (no separate videos/ directory). Lance streams the segment bytes lazily through byte-range fetches from the Hub.

Repro

pip install git+https://github.com/lancedb/lerobot-lancedb.git
lerobot-convert-to-lance-video \
    --repo-id=lerobot/pusht \
    --output=./pusht-lerobot-lancedb-video \
    --table-name=pusht-lerobot-lancedb-video

Layout

pusht-lerobot-lancedb-video/
├── meta/                                   # standard LeRobotDataset metadata sidecar
├── pusht-lerobot-lancedb-video.lance/        # one row per frame, tabular only
└── pusht-lerobot-lancedb-video_videos.lance/ # one row per source mp4, raw bytes as blob v2

Frames table columns: episode_index, frame_index, index, timestamp, task_index, observation.state, action. Pixels live in the _videos.lance blob column and are decoded per-(episode, video_key) segment at sample time.

Size: ~8 MB. Companion JPEG-frame layout (~60 MB, faster single-frame decode) at `lance-format/pusht-lerobot-lancedb`.