A robot that's never touched a real door just learned to open one anyway.
Researchers built SMART, a system that generates synthetic manipulation demonstrations in a custom simulator called SMART-Sim. The pipeline produced over 1 million demonstrations spanning 44 atomic task types, 5 robot setups, and 2,507 articulated objects, think cabinets, drawers, and hinges, anything with moving, constrained parts. That data pretrained a vision-language-action model, the type of AI that maps camera input and instructions directly to robot motor commands. The resulting model transferred to real-world tasks with zero additional training on physical robots, a result called zero-shot sim-to-real transfer, and also scored competitively on simulation benchmarks.
Articulated objects are the stubborn case in robot learning. Doors, drawers, and hinges require precise, constraint-following contact that is slow and expensive to demonstrate by hand at scale. Prior synthetic-data efforts covered a narrow slice of object types; scaling to over 2,500 objects across dozens of task categories is the actual claim worth watching, since it targets one of the bigger bottlenecks in getting general-purpose robots out of the lab.
Simulators have promised sim-to-real breakthroughs before. The real test is whether SMART's robots can handle the doors and drawers nobody thought to model in advance.