Train on terabytes. Don't download them
LeRobot 0.7 enables direct training from the Hub by streaming video bytes to workers, reducing data wait time to 0.52% on H100 GPUs with LanceDB and introducing GPU augmentations for faster ACT training.
LeRobot 0.7 introduces a streaming training stack that reduces data fetch wait time to 0.52% on H100 GPUs using LanceDB and GPU augmentations, enabling direct training from the Hub without downloading terabytes of data.
Source: LeRobot — X · Read original article ↗
Article text · Original source · English
Train on terabytes. Don't download them. 📦
LeRobot 0.7 trains straight from the Hub:
- Streaming: each worker fetches only the video bytes it needs. 0.52% of step time spent waiting for data (π0.5, 4× H100, 10.7 TB dataset)
- LanceDB, built with @lancedb: frames are read in place. 1.7% waiting for data, vs 37.4% from local disk after a 384 GB download
- GPU augmentations: 2.8× faster ACT training on an H100, with 8 new robotics augmentations
Plus a training stack for big models: FSDP2 + HSDP, checkpoints saved 3.81× faster and resumable on any GPU count, FP16, EMA and Trackio.
--dataset.streaming=true
Docs in the comments 👇
What the source reports
Publisher-reported claims, with original evidence. These results have not been independently verified by RoboSignal.
Reported numbers
dataset size
10.7
View original evidence
- Streaming: each worker fetches only the video bytes it needs. 0.52% of step time spent waiting for data (π0.5, 4× H100, 10.7 TB dataset)
Open source S3GPU training speedup
2.8
View original evidence
- GPU augmentations: 2.8× faster ACT training on an H100, with 8 new robotics augmentations
Open source S5checkpoint save speedup
3.81
View original evidence
Plus a training stack for big models: FSDP2 + HSDP, checkpoints saved 3.81× faster and resumable on any GPU count, FP16, EMA and Trackio.
Open source S6dataset download size
384
View original evidence
- LanceDB, built with @lancedb: frames are read in place. 1.7% waiting for data, vs 37.4% from local disk after a 384 GB download
Open source S4
- paper: not_reported
- code: not_reported
- weights: not_reported
- dataset: not_reported
Source excerpts and review record
Automatically extracted; no manual editorial approval recorded.
- Streaming: each worker fetches only the video bytes it needs. 0.52% of step time spent waiting for data (π0.5, 4× H100, 10.7 TB dataset)
Open source S3
- GPU augmentations: 2.8× faster ACT training on an H100, with 8 new robotics augmentations
Open source S5
Plus a training stack for big models: FSDP2 + HSDP, checkpoints saved 3.81× faster and resumable on any GPU count, FP16, EMA and Trackio.
Open source S6
- LanceDB, built with @lancedb: frames are read in place. 1.7% waiting for data, vs 37.4% from local disk after a 384 GB download
Open source S4
Train on terabytes. Don't download them. 📦
Open source S1
Source：LeRobot · x.com