Researchers at MIT-HAN Lab say their open 27-billion-parameter model PEAR learned robot control through reinforcement learning in simulation, without human demonstrations or a specialised control harness. The reported result is a sharp jump on a benchmark, plus a test on physical robots that the model had not encountered before.
Watch Desk analysis
What happened
The researchers call their method Robot-Use RL. They report that PEAR’s task success on the LIBERO benchmark rose from 1.6% to 74.4% after training. On real-world robot tasks, they say it scored 18 out of 30 without specialised perception tools or a harness.
The lab says the model’s simulation-trained skills transferred to real robots it had not previously encountered. That is the important step beyond a tidy simulation result: the claim is that one training route can carry over to unfamiliar physical machines.
Why it matters
Robot learning often depends on demonstrations or carefully built systems around a model. If simulation-based reinforcement learning can produce a model that transfers to real robots without those additions, it could offer researchers a more scalable route to teaching robots new tasks. The results also keep the model’s general reasoning, vision and coding capabilities, according to the researchers.
The figures are the lab’s reported results, not proof that the method will generalise across every robot, task or working environment. Still, a reported 18 successes out of 30 real-world tasks is a more useful test than asking a robot to behave itself in a virtual world and calling it a day.
Our read
This is a substantial robotics result because it connects simulation training to physical performance, rather than stopping at a benchmark chart. The key question is how much of that transfer survives independent testing and a broader range of robots and tasks. Researchers should look closely at the evaluation details before treating the headline percentage as a general measure of capability.
What to watch
- Whether independent researchers reproduce the LIBERO result.
- How performance changes across different robots, tasks and environments.
- Whether the approach works without specialised tools beyond the reported tests.
Discussion spark: Would you trust simulation-trained robot skills on unfamiliar physical machines, or should broader real-world testing come before deployment?
Sources and evidence
- Training Reasoning Models in Simulation to Natively Use Real Robots (13 November 2026, 18:39 UTC)
Watch Desk is operated by WittyWires as an independent cross-cutting AI news tracker. It does not speak for the organisations or people it covers.