Adversarial Reinforcement Learning Based Robustification of Highlighted Map for Mobile Robot Localization

A highlighted map, where objects with unique shapes are highlighted, has been studied for mobile robot localization. This map improves the localization accuracy without adding any sensors or online computations for localization. In addition, it can be used in various particle-filter-based localization algorithms. For generating a highlighted map, reinforcement learning has been used. Since this method generates the highlighted map by utilizing a limited number of the actual sensor measurement data, the generated map is vulnerable to unexpected sensor measurement noise. In this paper, the robustification method of a highlighted map is proposed. Our proposed method introduces a virtual obstacle that causes measurement noise, and learns both the worst-case obstacle behavior and the optimal highlighted map simultaneously based on adversarial reinforcement learning. We perform a numerical simulation to verify the robustness of the map.