Toward complete coverage planning using deep reinforcement learning by trapezoid-based transformable robot

Shape-shifting robots are the feasible solutions to solve the Complete Coverage Planning (CCP) problem. These robots can extend the covered areas by reconfiguring their shape to different forms according to space conditions. Since energy usage while navigating is constrained by the number of shape-shifting, it is desirable to cover the confined areas by trajectory using minimal robot actions within finite states. This paper presents a CCP method using deep reinforcement learning (DRL) for a reconfigurable robot with a trapezoid shape called Transbot. The framework derives optimal action policy for robot trajectory within the grid-based workspace. DRL model relies on Convolutional Neural Networks (CNNs) with Long Short Temporary Memory (LSTM) layers using Actor-Critic with Experience Replay (ACER) as the decision layers. The trained DRL model simultaneously generates robot shapes and directions with optimal energy cost by maximizing the cumulative reward representing the Transbot actions. By creating the trajectory with less 28.95% energy and 19.42% time in tested simulation and real-world experiments, the proposed CCP framework outperforms the existing tiling-based heuristic optimization techniques.