Learning Decentralized Control Policies for Multi-Robot Formation

摘要

Decentralized formation control has been extensively studied using model-based methods, which rely on model accuracy and communication reliability. Motivated by recent advances in deep learning techniques whereby an intelligent agent is trained to compute its actions directly from highdimensional raw sensory inputs using end-to-end decisionmaking policies, we consider the problem of learning decentralized control policies for multi-robot formation. A deep neural network is designed to model the control policy that maps the robot's local observations to control commands. We propose to use a centralized training framework based on supervised learning for control policy learning. The learned policy is then deployed on each robot in a decentralized manner for online formation control. Our proposed approach is verified and evaluated in experiments using a robotic simulator. Simulation results demonstrate satisfactory performance of formation control. Compared with existing methods for formation control, the proposed approach does not need inter-robot communication, and avoids hand-engineering the components of perception and control separately.