Solving Rubik's Cube with a Robot Hand

ABSTRACT

We demonstrate that models trained only in simulation can be used to solve a manipulation problem of unprecedented complexity on a real robot. We use reinforcement learning to train a policy to solve a Rubik's Cube with a highly dexterous human-like robot hand. To enable zero-shot transfer from simulation to the real world, we introduce Automatic Domain Randomization (ADR), which dynamically generates progressively more difficult environments in simulation. The resulting policy is highly robust to unmodeled dynamics and perturbations, allowing the physical robot hand to successfully solve the Rubik's Cube despite wearing a rubber glove or having individual fingers tied together.