FactoryBench

Robots are moving from demos into factories, hospitals, warehouses, and homes. But the question buyers actually care about is not “Can this robot do the task?” It is: “Can I trust it on my floor, around my people, with my hazards?”

FactoryBench turns a real site into a deterministic proving ground. A model can propose actions, but it never judges itself. A fixed oracle scores whether the robot should finish the job, escalate to a human, or refuse. Every failure becomes a training row, so each new robot brain has to re-earn permission before deployment