Neural Dynamic Policies for End-to-End Sensorimotor Learning

The current dominant paradigm in sensorimotor control, whether imitation or reinforcement learning, is to train policies directly in raw action spaces such as torque, joint angle, or end-effector position.