A new framework called SimEX teaches robots real-world skills almost entirely through simulation, then needs only 10 minutes of real-robot interaction to finish the job.
SimEX pairs a coding agent with a physics simulator in two stages. First, the agent runs open-ended trial and error entirely in simulation, building a toolbox of robot skills without any human demonstrations. Then it tries that toolbox on a real robot: each physical trial is used to correct the simulator itself, and the corrected simulator helps diagnose what went wrong and test fixes. Researchers validated the approach on manipulation tasks including towel folding, barcode scanning, and plate handling, with only 10 minutes of real-robot interaction required.
Most robot-learning methods either hand-code rules that break down in messy real environments, or rely on slow, costly, occasionally risky trial-and-error on physical hardware. SimEX instead treats the simulator as a workable lab assistant for a coding agent's own reasoning, not just a dataset generator, which is why the real-world portion of training shrinks to minutes rather than hours. That reframes simulation's job: less about perfectly mirroring reality, more about giving an AI agent a cheap place to think out loud.
It's three tasks in one lab, not a general-purpose robot brain - but if 10 minutes of real-robot interaction holds up on harder jobs, building robot skills starts looking more like debugging code than training an animal.