LeRobot 0.7 streams terabyte robotics datasets from the Hub, cutting data wait to 0.52% of step time
LysandreJik · x · 2026-11-11
Hugging Face's LeRobot team released v0.7 with a "train without downloading" stack:
- Streaming: each worker fetches only the video bytes it needs — data waiting is just 0.52% of step time on π0.5 with 4×H100 over a 10.7 TB dataset
- LanceDB in-place frame reads: 1.7% waiting vs 37.4% when reading from local disk after a 384 GB download
- GPU augmentations: 2.8× faster ACT training on an H100, plus 8 new robotics-specific augmentations
- Big-model training stack: FSDP2 + HSDP, checkpoints saved 3.81× faster and resumable on any GPU count, FP16, EMA and Trackio support
Enable with --dataset.streaming=true.
More from Embodied
- Sharpa Humanoid Robot Takes Polaroid Photos at Shanghai CIIE — Scobleizer · 2026-11-11
- Fable 5.2 designs a Tamagotchi across KiCad, Inventor and Blender — ZeYanjie · 2026-11-11
- Compact exoskeleton glove with encoders and tactile sensing shown at CoRL 2026 — ZeYanjie · 2026-11-11
- Wheeled bases, not legs: why production-test humanoids should skip bipedal design — AIFlow_ML · 2026-11-11
- Walden Robotics demos a robot tending a laser cutter fully autonomously — adnothing · 2026-11-11
- GEN-1.5 robot assembles engine housings 100 times in a row with Foundry Robotics — ishabytes · 2026-11-11