Training a robot to grasp an object means learning from many attempts, viewpoints and situations. One robot with three cameras at 20 fps
LeRobot's new streaming loader is designed for training on large-scale robot datasets, allowing efficient processing of 100 TB+ data without full local video copies. It reads episodes directly from Hugging Face or Buckets, prepares samples on-the-fly, and balances memory and throughput for varied training batches.
LeRobot introduces a streaming loader for handling large-scale robot datasets, enabling efficient training without requiring full local video copies. The tool supports 100 TB+ datasets by reading selected episodes directly from Hugging Face or Buckets, preparing samples on-the-fly, and balancing memory and throughput for varied training batches.
Source: LeRobot — X · Read original article ↗
Article text · Original source · English
Training a robot to grasp an object means learning from many attempts, viewpoints and situations. One robot with three cameras at 20 fps produces 216,000 images per hour. Now that robotics is scaling, how can we efficiently train on all this data? https://x.com/i/article/2108539955777597440
What the source reports
Publisher-reported claims, with original evidence. These results have not been independently verified by RoboSignal.
Reported numbers
images per hour
216,000
View original evidence
One robot with three cameras at 20 fps produces 216,000 images per hour.
Open source S2Open source S5samples per second
607
View original evidence
With four workers, the two-camera SO-101 pick-and-place dataset reached 607 samples/s, up from 283 with one worker.
Open source S18samples per second
1,130–1,354
View original evidence
The three-camera R3 dataset reached 1,130–1,354 samples/s, around 2.3–2.5× faster in matched runs.
Open source S18samples per second
283
View original evidence
With four workers, the two-camera SO-101 pick-and-place dataset reached 607 samples/s, up from 283 with one worker.
Open source S18samples per second
2.3–2.5
View original evidence
The three-camera R3 dataset reached 1,130–1,354 samples/s, around 2.3–2.5× faster in matched runs.
Open source S18
- paper: not_reported
- code: not_reported
- dataset: not_reported
Source excerpts and review record
Automatically extracted; no manual editorial approval recorded.
Training a robot to grasp an object means learning from many attempts, viewpoints and situations. One robot with three cameras at 20 fps produces 216,000 images per hour. Now that robotics is scaling, how can we efficiently train on all this data? https://x.com/i/article/2108539955777597440
Open source S2
Training a robot to grasp an object means learning from many attempts, viewpoints and situations. One robot with three cameras at 20 fps produces 216,000 images per hour. Across a fleet, demonstrations can grow to hundreds of terabytes. Now that robotics is scaling, how can we efficiently train on all this data?
Open source S5
During the run, the model consumed more than a million samples at an average of 42.5 samples/s per node. It spent just 0.52% of step time waiting for data. Training continued through regular checkpoints, and restart checks returned the expected next samples. The loader was supplying data quickly enough for this model to keep training.
Open source S17
To see how much data the loader could prepare, three-minute multi-worker tests used two public datasets on an 8-vCPU machine. With four workers, the two-camera SO-101 pick-and-place dataset reached 607 samples/s, up from 283 with one worker. The three-camera R3 dataset reached 1,130–1,354 samples/s, around 2.3–2.5× faster in matched runs.
Open source S18
Source：LeRobot · x.com