“NVIDIA's Eureka project trained a robot hand to perform human-level pen spinning using only reinforcement learning in massively parallel simulations, with no supervised dataset.”