Advancement: EvoScenario: Generating long tail scenarios for testing autonomous vehicles using evolutionary search

Ishaan Paranjape
Computational Media PhD Student
Virtual Event
Jim Whitehead

Join us on Zoom: / Passcode: 708703

Description:  Autonomous vehicles need to be rigorously tested, preferably before deployment at scale across the world. Simulation testing helps address this need. For robust AVs, test scenarios must widely cover the situations they would encounter in the real world, prioritizing the more dangerous ones. In the distribution of all test scenarios, the rare, absurd and dangerous ones generally belong to the long tail. Unfortunately, we lack the ability to generate such scenarios. The proposed project, EvoScenario, addresses this shortcoming by using evolutionary search to procedurally generate scenarios. We explore multi-objective and diversity search algorithms to evaluate which ones are more effective in generating long tail scenarios. The outcome of running EvoScenario is a catalog of automatically generated test scenarios which are rare, diverse and unsafe. We evaluate the rareness of these scenarios by comparing with outlier scenarios from the HighD dataset. AV developers can consequently use these scenario catalogs to test their AV technology stack to increase its reliability before deploying it to the real world.