Learning Dexterous In-Hand Manipulation
We use reinforcement learning (RL) to learn dexterous in-hand manipulation policies which can perform vision-based object reorientation on a physical Shadow Dexterous Hand. The training is performed in a simulated environment in which we randomize many of the physical properties of the system like friction coefficients and an object's appearance.
Evidence notes
- The policies learn vision-based object reorientation entirely in simulation while randomizing physical properties such as friction and object appearance.
- The trained policies transfer to a physical Shadow Dexterous Hand without human demonstrations.
- Physical trials produce finger gaiting, multi-finger coordination, and controlled use of gravity as emergent manipulation behaviors.
Company context
Shadow Robot develops dexterous robotic hands, tactile sensing and teleoperation systems for manipulation research, embodied-AI labs and advanced robot-control work.