Robust Adversarial Training for Sequential Decision Making in Safety-Critical Cyber-Physical Systems