Add lerobot-lancedb LIBERO example: VLA training with LanceDB as the data layer

[AyushExel]

What

End-to-end example (→ blog) training SmolVLA on LIBERO twice — base LeRobot on the official HuggingFaceVLA/libero parquet dataset vs the lerobot-lancedb video format (mp4 bytes in a Lance blob v2 column, bit-exact pixels) — plus everything the Lance table can do that parquet+mp4 can't.

Highlights

Performance (4×H100, SmolVLA read pattern: 2 cams + 50-step action chunk)

• Dataloader: 2.4× vs parquet+mp4 at every worker count, 45–90× vs the published image-parquet (which is 17× larger on disk precisely to dodge slow mp4 reads)
• Identical 40k-step multi-GPU finetunes: 2h17m vs 6h05m wall-clock, 1.04 vs 2.01 kWh; GPU power traces show working vs starved GPUs
• S3 streaming: 2,458 samples/s straight from s3://, no local copy

Training parity (the point): closed-loop eval over all 4 suites, 400 episodes/model —

	base	lance
LIBERO avg success	80.8%	82.0%

identical loss curves throughout; before-finetuning baseline 0%.

Migration: a 30-line train_lance.py entrypoint on stock lerobot-train — same CLI, same accelerate multi-GPU path.

Curation on the same table: Geneva stateful GPU UDF backfills SigLIP2 embeddings (273k frames, ~10 min, zero-copy schema evolution); IVF-PQ + FTS + btree indexes; 13 ms text→frame semantic search; SQL task filter → episode list → --dataset.episodes (curated finetune: 0→77% on libero_object).

Before/after rollout videos in assets/, all reproduction scripts in scripts/.

🤖 Generated with Claude Code