A team of researchers has built AI agents that invent their own worst-case driving scenarios to find holes in self-driving software.
The system, called EvoDrive, uses large language model agents to repeatedly rewrite scenario generators that simulate crashes and near-misses. A critic filters out scenarios that are physically implausible, while a separate evaluator decides which proposals are worth running through a limited simulation budget. The framework keeps every tested scenario in what the researchers call a Pareto archive, preserving a spread of options that trade off realism against how hard they are to survive. Tests on the MetaDrive and CARLA simulators showed EvoDrive found a wider range of failure cases than existing generators, and the scenarios it produced were useful for training driving policies, not just breaking them.
Safety testing for autonomous vehicles has long relied on handcrafted scenario libraries built from known accident patterns. That approach only tests for crashes engineers already anticipated. By letting agents evolve scenarios on their own, EvoDrive is designed to surface edge cases nobody wrote a rule for, which is exactly the kind of long-tail failure that has embarrassed self-driving programs before.
This is still a simulation-only benchmark paper, not a deployed safety tool. Whether an evolved scenario in MetaDrive maps to anything that happens on a real street is the open question every synthetic-scenario method eventually has to answer.