Learning Vision-based Pursuit-Evasion Robot Policies