ADV-0: Closed-Loop Min-Max Adversarial Training for Long-Tail Robustness in Autonomous Driving

ADV-0 is a new closed-loop adversarial training framework for autonomous driving that uses min-max optimization to improve robustness against rare but safety-critical scenarios. The system treats the interaction between driving policy and adversarial agents as a zero-sum game, converging to Nash Equilibrium while maximizing real-world performance bounds.

• →ADV-0 addresses the critical problem of autonomous driving robustness in rare but dangerous long-tail scenarios through adversarial training.
• →The framework uses a closed-loop min-max optimization approach that treats driving policy and adversarial agents as a zero-sum Markov game.
• →Unlike existing methods, ADV-0 aligns attacker utility directly with defender objectives to avoid misalignment issues.
• →The system theoretically converges to Nash Equilibrium and provides certified lower bounds on real-world performance.
• →Experimental results show improved generalizability for both learned policies and motion planners against unseen risks.