Multi-adversarial Testing and Retraining for Improving the Robustness of Autonomous Driving Policies

Autonomous vehicles, especially those based on deep reinforcement learning, are known for their susceptibility to the adversarial perturbations. To enhance their robustness, it is imperative to not only detect their decision errors through testing, but also fortify their robustness against these errors. This paper proposes an iterative optimization method that trains multiple adversarial agents with varying adversarial intensities to identify decision errors in the driving vehicle and enhance its robustness by retraining to counteract these adversarial agents. The effectiveness of the method is evaluated in a lane-changing scenario and the results demonstrate improved robustness of deep reinforcement learning based autonomous driving strategies compared to the adversarial reinforcement learning with a single adversary.