Physics engine (robot simulation)

Engines such as Nvidia's PhysX, DeepMind's MuJoCo and the open-source Newton run thousands of parallel environments on a GPU. Throughput, measured in environment steps per second, sets the cost of training; Skild AI's Steel solver claims a 3.8x average gain over PhysX. Speed only matters if policies trained in simulation transfer to real robots, the sim-to-real problem.