Multirobot coordination with deep reinforcement learning in complex environments

Abstract

In the multiple autonomous robot system, it is very important to complete path planning coordinately and effectively in the processes of interference avoidance, resource allocation and information sharing. In traditional multirobot coordination algorithms, most of the solutions are in known environments, the target position that each robot needs to move to and the robot priority are set, which limits the autonomy of the robot. Only using visual information to solve the problem of multirobot coordination is still less. This paper proposes a multi-robot cooperative algorithm based on deep reinforcement learning to make the robot more autonomous in the process of selecting target positions and moving. We use the end-to-end approach, using only the top view, that is, a robot-centered top view, and the first-person view, that is, the image information collected from the first-person perspective of the robot, as input. The proposed algorithm, which includes a dueling neural network structure, can solve task allocation and path planning; we call the algorithm TFDueling. Through its perception and understanding of the environment, the robot can reach the target position without collision, and the robot can move to any target position. We compare the proposed algorithm, TFDueling, with different input structure algorithms, TDueling and FDueling, and with different neural network structures, TFDQN and TFDDQN. Experiments show that the proposed TFDueling algorithm has the highest accuracy and robustness.