Training Reasoning Models in Simulation to Natively Use Real Robots

Researchers trained an open 27B reasoning model to operate real-world robots entirely through reinforcement learning in simulation, without using demonstrations or a harness. The method, termed Robot-Use RL, raises task success rates on the LIBERO benchmark from 1.6% to 74.4% while preserving the base model's general reasoning, vision, and coding capabilities. Open article →