In-Hand Cube Reorientation

Key Insight

Controlling multi-finger hands to perform in-hand manipulation requires managing continuous contact changes and joint coordination. By training a control policy using reinforcement learning in a physics simulator like MuJoCo, the robot can learn to rotate a cube to arbitrary target orientations. Applying sim-to-real transfer techniques ensures the learned policy remains robust against the unmodeled friction, backlash, and sensor noise of physical robotic hands.