lerobot-lancedb

📖 Docs: https://lancedb.github.io/lerobot-lancedb/

Lance-backed datasets for LeRobot. Drop-in replacement for LeRobotDataset with two storage layouts:

• LeRobotLanceDataset — per-frame JPEG bytes (lossy, fastest at single-frame access, optional GPU NVJPEG decode).
• LeRobotLanceVideoDataset — per-file mp4 bytes stored via Lance blob v2, decoded on the fly with torchcodec. Bit-exact pixels, ~same disk size as upstream.

Both subclass LeRobotDataset so existing trainers / samplers / isinstance checks accept them transparently.

Install

pip install lerobot-lancedb

For local development:

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

Quickstart

# Convert (recommended path for dtype=video sources)
lerobot-convert-to-lance-video \
    --repo-id=lerobot/aloha_static_cups_open \
    --output=./aloha_cups_open_lance_video --overwrite

from lerobot_lancedb import LeRobotLanceVideoDataset
ds = LeRobotLanceVideoDataset(root="./aloha_cups_open_lance_video")

For the JPEG layout, use lerobot-convert-to-lance and LeRobotLanceDataset instead. See the docs for the full CLI / API reference.

Benchmark

Realistic training read pattern (delta_timestamps, 8 frames / sample, batch 32, num_workers 4, CPU decode, H100):

Reproducible via examples/benchmark_formats.py.

Training parity

convert_to_lance_video trains a DiffusionPolicy on pusht to 68.4 % gym-pusht success (seed=42, 500 rollouts) — matches the head-to-head upstream parquet+mp4 result (68.0 %) and the published lerobot/diffusion_pusht (65.4 %).

Full numbers (pusht env-eval + ALOHA cups_open held-out MSE across all storage modes) in docs/benchmarks.md. Reproducers: examples/train_and_eval_lance.py and examples/aloha_loader_parity.py.

Cloud / Hub

Both readers accept s3://, gs://, hf://datasets/..., hf://buckets/... URIs and pick up credentials from the usual env vars (AWS_*, GOOGLE_APPLICATION_CREDENTIALS, HF_TOKEN). Lance does byte-range fetches — no full-dataset download.

Pre-converted reference datasets you can paste directly:

from lerobot_lancedb import LeRobotLanceDataset, LeRobotLanceVideoDataset

LeRobotLanceDataset(repo_id="lance-format/pusht-lerobot-lancedb")        # 60 MB JPEG layout
LeRobotLanceVideoDataset(repo_id="lance-format/pusht-lerobot-lancedb-video")  # 8 MB video-blob layout

License

Apache 2.0.