Learning Dexterous In-Hand Manipulation

ABSTRACT

We demonstrate that reinforcement learning algorithms can be used to learn highly dexterous, in-hand manipulation policies that successfully transfer to the real world. We train a policy to control a physical Shadow Dexterous Hand—a robotic hand with 24 degrees of freedom—to manipulate a block into a target configuration. The policy is trained entirely in simulation using Proximal Policy Optimization (PPO) and transferred zero-shot to the physical robot. To bridge the reality gap, we utilize massive Domain Randomization, varying physical parameters like friction, gravity, and object dimensions. Our system, Dactyl, successfully discovers complex human-like grasping strategies without any human demonstrations.