Learning-Based Robust Evasion for Flexible Mobile Robots in Pursuit-Evasion Games Using Gaussian Processes

In this paper, a learning-based robust evasion method is proposed to solve the problem of evading strategies from a scalable number of pursuers. The method aims to provide more escape possibilities while overcoming the issue of sparse reward and local optima in unknown environments. The approximation of Q-functions via Gaussian process allows for an accurate online update of parameters and significantly enhances the training efficiency when dealing with high-dimensional data. Simulation and experimental results on several escape tasks for robots demonstrate the effectiveness and robustness of the method. The evading strategies can be scaled to pursuit-evasion problems involving multiple pursuers and evaders, thus providing a framework for multiplayer pursuit-evasion games.