MIT Scales Robot Skills With PhysicsGen Simulations

MIT researchers have developed a system called PhysicsGen that generates thousands of training simulations from just a few virtual reality demonstrations. Designed to help robots master object manipulation, the system captures human hand movements, remaps them to robotic joints, and optimizes motions for efficient task completion.

The pipeline boosts robot learning accuracy and flexibility across both virtual and real machines. It also reduces the need for repetitive human training and adapts older datasets for new systems. The long-term goal is to build scalable, robot-specific data that serves as a foundation for general-purpose robotic skills in homes and factories.