中国科学院自动化研究所机构知识库

Robotic skill learning suffers from the diversity and complexity of robotic tasks in continuous domains, making the learning of transferable skills one of the most challenging issues in this area, especially for the case where robots differ in terms of structure. Aiming at making the policy easier to be generalized or transferred, the graph neural networks (GNN) was previously employed to incorporate explicitly the robot structure into the policy network. In this paper, with the help of graph neural networks, we further investigate the problem of efficient learning transferable policies for robots with serial structure, which commonly appears in various robot bodies, such as robotic arms and the leg of centipede. Based on a kinematics analysis on the serial robotic structure, the policy network is improved by proposing a weighted information aggregation strategy. It is experimentally shown on different robotics structures that in a few-shot policy learning setting, the new aggregation strategy significantly improves the performance not only on the learning speed, but also on the control accuracy. (C) 2020 Elsevier B.V. All rights reserved.