Skip to content

lerobot-lancedb

Lance-backed datasets for LeRobot.

Two storage layouts, both subclasses of LeRobotDataset:

  • Frames format — per-frame JPEG bytes (LeRobotLanceDataset).
  • Video format — per-file mp4 bytes via Lance blob v2 (LeRobotLanceVideoDataset).

Both readers expose the same API. Pick by source dtype (see Conversion).

Install

Until the first PyPI release, install from GitHub:

pip install git+https://github.com/lancedb/lerobot-lancedb.git

For local development:

git clone https://github.com/lancedb/lerobot-lancedb.git
cd lerobot-lancedb
pip install -e '.[dev]'

Either path pulls in:

  • lerobot[dataset]
  • lancedb / pylance
  • torchcodec (used by the video format)

30-second tour

Convert a video-stored dataset to the recommended (bit-exact) layout:

lerobot-convert-to-lance-video \
    --repo-id=lerobot/aloha_static_cups_open \
    --output=./aloha_cups_open_lance_video \
    --overwrite

Use it as a regular LeRobotDataset:

from lerobot_lancedb import LeRobotLanceVideoDataset

ds = LeRobotLanceVideoDataset(root="./aloha_cups_open_lance_video")

Plug it into any code that expects a LeRobotDataset:

  • the upstream training factory
  • EpisodeAwareSampler
  • third-party trainers that do isinstance(ds, LeRobotDataset)

Headline benchmark

Realistic training read pattern (delta_timestamps, 8 frames per sample, batch_size=32, num_workers=4, CPU decode, H100) on lerobot/aloha_static_cups_open (480×640, 4-cam bimanual):

format size MB fps speedup bit-exact?
upstream parquet+mp4 485.6 18.7 1.00× ✓
convert_to_lance (JPEG-95) 3 626 46.0 2.46× ✗
convert_to_lance --jpeg-quality=100 --jpeg-subsampling=0 8 735 32.5 1.74× ✗
convert_to_lance_video 487.4 45.6 2.44× ✓

Full numbers across three datasets (pusht, ALOHA, Koch): Benchmarks.

Next