Title: EvoDrive: Leveraging Pareto Evolution and Self-Improving LLM Agents for Safety-Critical Autonomous Driving

Abstract:

Creating safety-critical scenarios is a vital step in validating and enhancing autonomous driving systems. However, this process presents a fundamental challenge: it must maximize adversarial potential to reveal system failures while simultaneously maintaining realistic conditions. Current approaches typically rely on manual heuristics to balance these competing demands, which restricts generation to established priors and ignores less explored patterns. Although recent advancements in open-ended agentic evolution have the potential to surpass these limitations, unrestricted general agents often lack rigorous grounding in simulators. Consequently, they tend to simplify complex multi-objective tensions into single-scalar maximization problems.

To address these issues, we introduce EvoDrive, the first automated framework for multi-objective scenario generation that utilizes LLM-based agentic evolution. EvoDrive utilizes a simulator-grounded actor-critic architecture. In this setup, a memory-driven actor iteratively suggests enhancements to the generators, while critics eliminate implausible candidates. Additionally, a self-evolving world evaluator directs promising proposals toward simulation budget optimization. By maintaining a Pareto archive of evaluated candidates, EvoDrive preserves a diverse range of trade-offs between attack effectiveness and realism, using simulation feedback to steer future evolutionary steps. Evaluations on the MetaDrive and CARLA benchmarks demonstrate that EvoDrive substantially extends the Pareto frontier across various generators and yields high-value scenarios suitable for policy training.

Source: arXiv Generated at: 2026-06-03 00:00:00 UTC