Realistic Extreme Behavior Generation for Improved AV Testing

Abstract

This work introduces a framework to diagnose the strengths and shortcomings of Autonomous Vehicle (AV) collision avoidance technology with synthetic yet realistic potential collision scenarios adapted from real-world, collision-free data. Our framework generates counterfactual collisions with diverse crash properties, e.g., crash angle and velocity, between an adversary and a target vehicle by adding perturbations to the adversary's predicted trajectory from a learned AV behavior model. Our main contribution is to ground these adversarial perturbations in realistic behavior as defined through the lens of data-alignment in the behavior model's parameter space. Then, we cluster these synthetic counterfactuals to identify plausible and representative collision scenarios to form the basis of a test suite for downstream AV system evaluation. We demonstrate our framework using two state-of-the-art behavior prediction models as sources of realistic adversarial perturbations, and show that our scenario clustering evokes interpretable failure modes from a baseline AV policy under evaluation.

Figures (11)

• Figure 1: An example counterfactual collision from our framework generated by modifying a reference scenario from the Waymo Open Dataset waymo-open-dataset with a perturbation derived from the MTR behavior model shi2022motion. Our framework modifies the trajectory of an adversarial agent, using realistic behavior perturbations, to encourage a collision with a target agent along a reference trajectory. In this example, all reference trajectories are highlighted in blue, while the adversary's trajectory is colored red. The non-adversarial agents' initial and final positions are highlighted by a green and blue star, respectively; the adversary's initial and final position are highlighted by a yellow and red star, respectively. We present a counterfactual in which the adversary performs an aggressive, over-wide right turn and collides with traffic stopped in the oncoming lane. Our approach offers valuable counterfactual scenarios -- grounded in the notion of realism -- on which to evaluate the maturity of existing AV collision avoidance technology.
• Figure 2: The risk contour $c$ offers a segmentation between realistic and unrealistic behavior with respect to a generic feed-forward behavior model.
• Figure 3: The number of clusters balances a representative set of scenarios, measured through the minimum cluster size, and the effectiveness of the clustering algorithm, measured by the Silhouette score. $k$, the number of clusters, achieves a low minimum cluster size and high Silhouette score for $k=8$.
• Figure 4: Clusters are numbered in descending order with respect to the severity of the collision by virtue of the mean adversarial vehicle speed.
• Figure 5: Singular values of the sketching decomposition of the Hessian matrix for the Trajectron++ and MTR models salzmann2020trajectronshi2022motion. Rapid decay of singular values indicates that a low-rank approximation captures most of the singular-value energy within the matrix.