Pick Anything: Learning Human-like Grasps from Human Videos

Robots that learn should generalize: locomotion across terrain, and manipulation across object categories. Pick Anything demonstrates human-like grasps on 66 different objects, ranging in size and shape from AirPods to dumbbells to pipe wrenches, using a single policy. The system perceives the world only through a stereo camera head, with no wrist cameras, and manipulates with Sharpa Wave dexterous hands. It was trained from human video and reinforcement learning, with zero teleoperation data. We are showing live demos of the system at CoRL 2026 (Demo 12, November 9 and 10).