Imitation and Adaptation Based on Consistency: A Quadruped Robot Imitates Animals from Videos Using Deep Reinforcement Learning

Qingfeng Yao, Jilong Wang, Shuyu Yang, Cong Wang, Hongyin Zhang, Qifeng Zhang, Donglin Wang

Abstract: The essence of quadruped movements is the move-ment of the center of gravity, which has a pattern in the movement of quadrupeds. However, planning the gait motion of the quadruped robot is time consuming. Animals in nature can provide a large amount of gait information for robots to imitate skills. Common methods learn the posture of animals with a motion capture system or numerous motion data points. In this paper, we propose a video imitation adaptation network that can imitate the action of animals from a few seconds of video and adapt skills to the robot. The deep learning model extracts key points of animal motion from videos. A motion adaptor that eliminates noise and extracts key information of motion is proposed; the information of the extracted movements was used as a motion pattern to help deep reinforcement learning complete the imitation of skills. To ensure similarity between the learning result and the animal motion in the video, we introduce rewards that are based on the consistency of the motion. The results show that our framework can help robots learn periodic and aperiodic skills from several seconds of videos of different types of animals. And we deploy the trained model to the real robot to complete the walking and backflip tasks without fine-tuning.