{"name":"pear","repo":"mit-han-lab/pear","url":"https://midorreal.com/project/mit-han-lab-pear","repository":"https://github.com/mit-han-lab/pear","description":"Training Reasoning Models in Simulation to Natively Use Real Robots","facts":{"license":{"value":"Apache-2.0","source":"https://github.com/mit-han-lab/pear","checked":"22 Nov 2026"},"language":{"value":"Python","source":"https://github.com/mit-han-lab/pear","checked":"22 Nov 2026"},"releases":{"value":"0","source":"https://github.com/mit-han-lab/pear/releases","checked":"19 Nov 2026"},"contributors":{"value":"3","source":"https://github.com/mit-han-lab/pear/graphs/contributors","checked":"19 Nov 2026"},"archived":{"value":"no","source":"https://github.com/mit-han-lab/pear","checked":"22 Nov 2026"}},"stars":779.0,"writeup":{"what_it_is":"PEAR is a post-training engine for agentic robotics. It trains reasoning models to use robots with reinforcement learning in simulation, and it also covers evaluation and deployment to real robots.","audience":"Researchers and engineers training reasoning models to control robots","claims":[{"kind":"specific","claim":"Trains an open 27B reasoning model to use real robots entirely in simulation, with zero demonstrations","excerpt":"an open 27B reasoning model learns to use real robots entirely in simulation, with zero demonstrations","status":"not_checked"},{"kind":"specific","claim":"Uses fully asynchronous RL built on veRL, with GRPO and dynamic sampling","excerpt":"Fully asynchronous RL on [veRL]","status":"not_checked"},{"kind":"specific","claim":"Includes optimized kernels for H100, H200 and GB300 hardware","excerpt":"Optimized kernels for H100, H200 and GB300","status":"not_checked"},{"kind":"specific","claim":"Evaluates on robotic benchmarks including LIBERO, LIBERO-Plus, LIBERO-Pro, RoboDojo and RoboLab-120","excerpt":"Robotic benchmarks: LIBERO, LIBERO-Plus, LIBERO-Pro, RoboDojo and RoboLab-120","status":"not_checked"},{"kind":"specific","claim":"Deploys a vLLM model server on one RTX 5090 or GB300 with decode megakernels","excerpt":"vLLM server on one RTX 5090 or GB300, with decode megakernels","status":"not_checked"}],"alternatives":[],"written_by":"AI, from the project's README","date":"2026-11-17"},"why_now":null,"owner_supplied":null,"score":{"verdict":"mid","hype":42,"reality":50,"gap":-8,"momentum":null,"confidence":60,"version":"score-2.0","rule":"verdict-2.0","calculated_at":"2026-11-22T08:00:00+00:00","verdict_source":"rule"},"early_read":null,"changes":[{"date":"2026-11-13T17:43:06+00:00","text":"Repository created","source":"https://github.com/mit-han-lab/pear"}],"cite":"pear is a GitHub project at https://github.com/mit-han-lab/pear written mainly in Python, described there as \"Training Reasoning Models in Simulation to Natively Use Real Robots\".","attribution":"Data from Mid or Real, https://midorreal.com/project/mit-han-lab-pear, checked 22 November 2026. Attribution with a link is required.","still_real":null}