ScenePilot: Controllable Boundary-Driven Critical Scenario Generation for Autonomous Driving

Researchers introduced ScenePilot, a reinforcement-learning framework for generating critical autonomous-driving scenarios near safety boundaries.

It combines constrained multi-objective reinforcement learning with feasibility-aware shielding to keep scenarios physically valid while still challenging the autonomy stack.

The method is designed to trigger realistic failures rather than impossible crashes, improving boundary-driven stress testing for self-driving systems.

On SafeBench, ScenePilot reports higher collision rates, and adversarial fine-tuning further improves robustness.