Balanced Map Coverage using Reinforcement Learning in Repeated Obstacle Environments

This paper demonstrates novel Complete Coverage Path Planning using reinforcement learning to enable a robot to complete map coverage at high speeds in complicated environments, like factories and airline cabins, with many repeated obstacles, such as furniture and walls. Our framework trains the robot in a simulated environment to move to uncovered areas and to avoid frequent collisions using rewards. Additionally, it encourages the robot to complete map coverage missions efficiently and quickly. We select the Machine-Learning Agent provided by Unity3D to build a fragment (sample cell) of an airline cabin environment in which to train the robot. We implement Proximal Policy Optimization as the main training network, and added curiosity functions (i.e., intrinsic rewards) to encourage the robot to explore uncovered areas during training. We use Generative Adversarial Imitation Learning to guide the training policy's convergence close to the expert data. Experimental results show that our optimal policy enables complete map coverage in complicated environments. We provide demonstrations comparing random motion methods to reinforcement learning networks to show differences in map coverage, trajectory length, and time-cost.