MIT Han Lab researchers say they trained an open 27-billion-parameter reasoning model to operate real robots using reinforcement learning in simulation, without demonstrations of how to do the tasks. On a real dual-arm robot, the trained model succeeded in 60% of their trials; the untrained model succeeded in none.
The model, based on Qwen3.8-27B, also improved from 1.6% to 74.4% on the LIBERO robot-task benchmark. The researchers say that result approached GPT-6 Astra and GPT-6.1 Sol when those models were run at their highest reasoning settings. The results are from the researchers’ evaluations, not an independent assessment.
How does it control a robot?
The model gets camera images and information about the robot’s arm, reasons about what to do, then outputs the arm’s next position. A robot controller turns that position into movement. The researchers describe this as similar to a computer-use agent looking at a screen and issuing a click, but say their system does not use an extra layer of perception tools, motion planners or prebuilt skills.
Training used robot-use reinforcement learning: the model tried tasks in simulated environments and learned from whether it succeeded. The team’s PEAR software provides the training infrastructure, including simulations and evaluation tools. MIT Han Lab says it has released the model, code and training infrastructure.
Does simulation training transfer?
The researchers report that the model carried its skills to a real robot, a different robot and instructions it had not encountered in training. On LIBERO-Pro, which tests changed goals and object layouts, their model succeeded on 57.4% of episodes involving a new task in a familiar scene. The demonstration-trained model π0.5 succeeded on 0.75% in that setting.
That is a different kind of test from simply repeating familiar tasks. On LIBERO, the researchers evaluated models over 2,000 episodes. They also report no measurable loss across nine reasoning, coding and vision benchmarks after robot training. The paper is an arXiv preprint, and the results do not establish how the system would perform in everyday use or in other robots and settings.
Why it matters
Robotics researchers can use the released model, code and PEAR training infrastructure to investigate whether simulation-based training can reduce the need for task demonstrations. The reported real-robot result is promising, but it does not show how reliably the system would work outside the tested robot and tasks.