LeRobot streams 10.7TB robot dataset during training with just 0.52% idle time
RemiCadene · x · 2026-10-17
Hugging Face's LeRobot team demonstrated streaming training of the π0.5 policy on 4× H100 for 8 hours, reading a 10.7 TB dataset (finerobotics/realomni-10kh, 1M+ samples) directly from the Hub without downloading it first.
- Time spent waiting for data: only 0.52% of step time, at 42.5 samples/s per node
- Enabled by a single flag: lerobot-train --policy.path=... --dataset.repoid=... --dataset.streaming=true
- The team open-sourced a streaming-dataloader Space, claiming support for 100 TB+ robot datasets
More from Embodied
- Hobbyist models humanoid robot, hits battery space and servo torque limits — _Stocko_ · 2026-10-17
- Android robots learn to switch emotions in real time, crossing the uncanny valley — TinfoilTricorn · 2026-10-17
- LEMO agent pipeline runs on real robots, completing supervised tennis-ball-in-cup task — du_yilun · 2026-10-17
- India's Altersage unveils Cocobot, billed as first commercially viable coconut-harvesting robot — Scobleizer · 2026-10-17
- Quadruped robot tech transforms into an all-terrain wheelchair — TinfoilTricorn · 2026-10-17
- AMD acquires Fei-Fei Li's World Labs in $8.2B all-stock deal; Li becomes Chief Scientist — lukas_m_ziegler · 2026-10-17