HumanConQuad: Human Motion Control of Quadrupedal Robots using Deep Reinforcement Learning

Abstract

Robotic creatures are capable of entering hazardous environments instead of human workers, but it is challenging to develop a fully autonomous agent that can work independently in unstructured scenes. We propose a human motion-based control interface for quadrupedal robots that promises adaptable robot operations by reflecting the user’s intuition directly to the robot’s movements. Designing motion interface for different morphologies conveys tricky problems in solving dynamics and control strategies. We first retarget the captured human motion into the corresponding robot’s kinematic space with proper semantics using supervised learning and post-processing techniques. Second, we build the motion imitation controller to track the given retargeted motion using deep reinforcement learning with task-based curriculums. Finally, we apply domain randomization during training for real-world deployment. (Video1)