Dexterous manipulation policies can be trained in simulation and deployed on real robots with high success rates by combining diverse training data, adaptive curricula, and learned safety filters that trade off performance for reliability.
This paper presents a method for training dexterous robotic hands to perform complex in-hand object manipulation tasks like repositioning objects to arbitrary poses. The approach combines simulation training with a learned safety filter that reduces object drops by 75% while maintaining 91.5% task success.