DA-SLAM: Deep Active SLAM based on Deep Reinforcement Learning

Abstract

This work presents improvements to the state-of-the-art algorithms for path planning and exploration of unknown and complex environments using Deep Reinforcement Learning. Our novel approach takes into consideration: (i) map information, built online by the robot using a Simultaneous Localization and Mapping algorithm and (ii) uncertainty of the robot's pose, which leads to active loop-closing to encourage exploration and better map generation within two agents. The results show that the map completeness-based reward function outperforms literature's results on shorter trajectories, thus, better performance; while uncertainty-based with loop-closing reward function improves map generation. Both agents showed the ability, to perform Active SLAM over complex environments and generalization to unseen maps capabilities.