Deep reinforcement learning for swarm robot autonomous behaviors

Abstract

Unmanned robots such as unmanned aerial vehicles are becoming more advanced and inexpensive due to improvement of sensors and actuators. To overcome a single robot system`s weakness, robotic system in the form of a swarm which is inspired by collective behaviors of natural social insects are emerged. However, previous swarm robot systems required direct human intervention which makes swarm robot system less flexible, so that they only work in few specific scenarios. In this paper, to improve flexibility and autonomy of swarm robot mission support, we propose deep reinforcement learning method for swarm robots` autonomous behavior. We discuss various swarm robot behavior occurred in swarm robot mission such as alignment, cohesion and separation. Then swarm MDP consisting observation space, action space, and reward function design is defined to formulate our problem. With swarm MDP defined, we provide neural network design and training algorithm. We present the experimental results of a swarm robot behavior through simulation.