Ivan Ang

This paper proposes a mechanism to automatically assess the safety of autonomous robots, and in particular autonomous cars. Most methods to assess the safety of autonomous cars generate adver-sary strategies, which will then be used to test whether the car beingassessed can avoid accidents with the adversary. To generate such adversarial strategies, many have proposed learning techniques that requirea large amount of accident data. But, such data are difficult to obtain because accidents are rare. To alleviate this issue, we leverage the observation that when safe and colliding adversary trajectories are closer together, the vehicle is less safe because there is generally less buffer to avoid accidents. Specifically, we generate/utilise data on adversaries’ safe trajectories, which is more abundant than accidents data, and compute colliding adversarial trajectories that are as close as possible to the safe trajectories. The average distance between safe and colliding adversarial trajectories provides an indicator of the vehicles’ safety. To compute colliding adversarial trajectories, we take into account that the driving strategy of the vehicle being assessed is not fully known, and therefore propose a multi-objecive POMDP framing of the problem and an on-line planning method, called Constraint-Aware Tree (CAT), to compute approximate solution to the multi-objective POMDP. Evaluations of four learning-based autonomous driving software on pedestrian crossing and lane merging scenarios, derived from the National Highway Traffic Safety Administration (NHTSA), indicate the viability of the proposed testing mechanism in assessing a variety of autonomy software. Moreover, evaluations of CAT on the nuScenes dataset indicate that CAT generates more colliding adversarial trajectories in less time compared to state-of-the-artlearning-based method, STRIVE.