Learning Human-to-Robot Handovers from Point Clouds

Sim-to-Real Transfer + Test Figure 1 . We introduce a framework to learn human-to-robot handover policies from point cloud input. Our policies take input from a wrist mounted camera and directly generate action output for the robot's end effector. We train our policies in a simulated handover environment, and evaluate on unseen handover motion and poses. We further transfer the model across physics simulators and to a real robotic platform.