MIT-led researchers have developed a reinforcement learning method that teaches AI models to operate physical robots using only simulated training.
MIT-led researchers have developed a reinforcement learning method that teaches AI models to operate physical robots using only simulated training.
Called Robot-Use RL, the approach enables a 27-billion-parameter reasoning model to interpret camera observations, plan actions, and control robotic arms directly. The system achieved a 60% success rate across 30 real-world trials, with its model weights, code, and training infrastructure publicly released.