SPHINX: An explain-before-explore framework for adversarial scenario generation for autonomous driving

The researchers propose SPHINX, a closed-loop framework for adversarial driving scenario generation for autonomous driving decision systems, with the core principle of explaining before exploring. Addressing the problem that existing methods mainly rely on the prior knowledge of large language models and vision-language models to generate scenarios, the authors argue that adversarial scenarios should be generated based on failure diagnoses of the driving policy (such as hesitation and multi-frame inconsistency). SPHINX uses explainable AI methods to analyze the policy, identifying key visual concepts and decision uncertainty that affect the policy's output, and then has a vision-language model reason about and critique the current policy's failure modes to generate targeted adversarial scenarios for policy retraining. The authors state that, on the benchmarks and test suites evaluated, the framework applies to multiple state-of-the-art autonomous driving architectures, brings consistent robustness improvements compared with existing scenario generation methods, and can provide interpretable explanations of policy failures.