A new training method lets robots practice in thousands of AI-generated simulated rooms before ever touching a real one.
Researchers fine-tuned vision-language-action models, the systems that let robots interpret instructions and act on them, using reinforcement learning inside scenes built by generative 3D world models. A language-driven scene designer produced 100 distinct interactive environments with varied objects and backgrounds, letting the team train policies in parallel instead of hand-building each scene. In simulation, task success jumped from 9.7 percent to 79.8 percent, with robots completing tasks 1.25 times faster. When those same policies moved to physical hardware, real-world success rose from 21.7 percent to 75 percent - more than tripling performance - with a 1.13 times speedup.
Robot learning has long faced an ugly tradeoff: train in the real world and get a policy that only works in that one room, or train in simulation and hit a wall because building varied simulated scenes takes too much manual labor. Generative 3D worlds sidestep that bottleneck by manufacturing scene diversity on demand, and the team's own ablation study showed that adding more generated scenes directly improved how robots handled situations they had never seen before. That is the actual constraint robotics has been stuck on, not compute or model size.
It's a lab result, not a shipping product, and a 75 percent success rate still means the robot fails one time in four - but it's a more convincing answer to the sim-to-real problem than just throwing more real-world trial and error at it.