Collision filtering for autonomous vehicle simulations

Abstract

Collision filtering for autonomous vehicle simulations is disclosed herein. A simulation system can be configured to identify a simulated collision involving a simulated autonomous vehicle and a simulated agent representing an obstacle with which the simulated autonomous vehicle may collide. For example, the simulated collision can be simulated based on logged data from a real-world autonomous vehicle. The simulation system can classify the simulated collision as unrealistic or realistic. For example, classification can involve considering a nature of the simulated agent (including, e.g., whether the simulated agent is a drivable obstacle), a nature of the collision, whether the collision is avoidable through an evasive maneuver, whether the collision is similar to a known false positive, and/or input from a human operator. The simulation system can filter the simulated collision from participating in a collision triage analysis in response to classifying the simulated collision as an unrealistic simulated collision.