Research
MIT Han Lab trains an open 27B reasoning model to operate real robots using only reinforcement learning in simulation, and open-sources the PEAR training engine
A team led by Jiaming Tang, Zhuoyang Zhang, Junxian Guo and Song Han released PEAR on November 13, 2026, with arXiv paper 2611.18562. Starting from Qwen3.8-27B and using no demonstrations, simulated robot-use reinforceme…