⬢github Python · 756 ★ +40 since we first saw it · pushed 11 h ago · Apache-2.0
mit-han-lab/pear
Training Reasoning Models in Simulation to Natively Use Real Robots
PEAR is a post-training engine from MIT Han Lab that trains open reasoning models (e.g. Qwen3.8-27B) to control real robots using reinforcement learning entirely in simulation — no demonstrations. It bundles asynchronous RL training, simulators (MuJoCo, Isaac Sim), robotic and general benchmarks, and deployment tooling including a vLLM model server and agent loop for YAM and SO-101 robots.
Why now: Freshly released alongside a blog post and arXiv paper, it's trending on GitHub with 700+ stars; a recent news item highlights a fast DFlash2 checkpoint decoding at 921 tokens/s on a GB300.
Who it is for: Robotics and RL researchers or engineers who want to train and deploy LLM-based robot controllers.
Stars over our 119 snapshots: 716 to 756, since 1 d ago.
Where people talked about it
- ⬢github new repos, most starred 13 min ago
- ⬢github new repos, most starred 20 h ago
API: https://socialmediatrends-api.osmike.com/v1/repos/mit-han-lab/pear