Hugging Face released LeRobot v0.6.0: policies built on 'world models', new VLA models, reward models to score success, a unified CLI for evaluation, and six simulation benchmarks for robot learning.
- LeRobot v0.6.0 (July 7) adds policies built on world models, new VLA models, reward models, and 6 benchmarks under one CLI.
- Plus depth sensors, VLM auto-annotation, FSDP multi-GPU, and cloud training through Hugging Face Jobs.
- Read: robotics is repeating software's path - open foundations, with the edge going to data and integration.
The open stack for robot learning keeps maturing. This release shifts the focus from 'react' to 'predict'.
The core of it is the 'world model'. The robot imagines the next state and measures itself against it, instead of learning blind, by trial and error. That's a change of direction, not another changelog-style checkbox.
But the model is open - so it isn't the edge. The winner is whoever accumulates their own recordings of real movement and knows how to tie them together. The data from your robots becomes the moat. Everyone has the architecture.