Safe Autonomous Driving Scenarios
adversarial generation, preference alignment, and LLM agents for closed-loop autonomous driving evaluation.
I study how to generate traffic scenarios that are both realistic and safety-critical, so that autonomous driving systems can be tested and improved before deployment.
This line includes preference-aligned adversarial scenario generation, retrieval-augmented LLM attack agents, and Pareto evolution for balancing attack strength with scenario realism.
Representative work: SAGE, EvoDrive, and LLM-attacker.