Adversarial Trajectories Generation for Automotive Applications

The development of Advanced Driver Assistance Systems (ADAS) with a high level of autonomy requires immense testing efforts to ensure the safety and robustness of developed algorithms in critical situations. Unfortunately, exploration of difficult situations through test drives in natural traffic is ineffective due to the rarity of such events. While scenario-based testing in a virtual environment is often proposed as an effective method that helps to evaluate system performance in difficult situations, the manual definition of virtual test scenarios poses a significant challenge itself. Performance drops in tested systems, especially ones containing machine learning components, may be related to situations that are not necessarily considered challenging for a human driver and thus are difficult to predict in a test design. In this paper, we propose a method that allows to generate a variety of virtual test scenarios for ADAS through an adversarial trajectories generation. The method generates scenarios by finding trajectories of the road users in the proximity of the vehicle controlled by the tested algorithm that result in safety-critical events, such as collisions. We demonstrate the effectiveness of the presented method on an example of a critical scenario generation for a vehicle control policy based on Reinforcement Learning methods.