LeRobot: an open-source library for end-to-end robot learning, at ICLR 2026
Robotics has lacked the shared substrate that language modelling got years ago. An open end-to-end library that multiple labs actually use changes what results mean, because two papers can finally be compared.
The reproducibility problem in robot learning is worse than almost anywhere else in machine learning. Results depend on hardware, calibration, controller tuning and physical conditions that never make it into a paper. A shared library does not eliminate that, but it removes the software half of the variance.
The wider consequence is dataset compatibility. Once labs share a stack, they can share demonstrations, and the data bottleneck that every robotics group names as its binding constraint becomes at least partially poolable rather than duplicated.
Whether it becomes the default is an adoption question and not a technical one. Shared infrastructure only works if the people with the most data have a reason to publish in the common format.
arXiv — LeRobot: an open-source library for end-to-end robot learning → · arXiv — OpenVLA: an open-source vision-language-action model → · arXiv — Vision-language-action models: concepts, progress, applications and challenges →