“The Eureka project trained a robot hand to perform pen spinning using reinforcement learning in massively parallel simulations on GPUs, which can run 10,000 times faster than real time.”