Realistic and Interactive Safety-Critical Scenario Generation through Guided Diffusion

Abstract

This thesis introduces a novel framework for generating safety-critical traffic scenarios for autonomous vehicles (AVs), aiming to create realistic simulations that reflect actual dangers faced by motion planners on the road. Despite advancements in AV technology, many existing motion planning systems are not exposed to a wide variety of challenging traffic situations, potentially leading to unforeseen safety risks in scenarios they have not encountered before. To address this issue, this framework strategically generates synthetic data that captures a broad spectrum of traffic scenarios, enriching the diversity and realism of simulations for autonomous vehicles. Central to this framework is a conditional guided diffusion policy which is used to enhance scenario generation, offering improved controllability, multi-modality, and scene-consistency as compared to existing approaches. The model serves as a scenario optimizer, embedding an understanding of typical traffic behaviors to ensure that the generated scenarios are not only varied and challenging but also realistic in terms of driving dynamics and adherence to traffic laws. Additionally, the framework incorporates a game-guided strategy that combines iterative best response (IBR) with tau-Gradient Descent Ascent. This method enables dynamic interactions among agents, with IBR allowing these agents to adapt through simulated decision-making processes that replicate human driving behavior in an effort to maintain safety. By generating highly realistic and interactive synthetic traffic scenarios, this framework aims to significantly improve the training datasets available for motion planners, enhancing their capability to handle complex driving scenarios and thereby contributing to the increased resilience and decision-making skills of deployed AVs.