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

2026-03-16

AI Summary

This work addresses the insufficient robustness of autonomous driving systems in rare yet safety-critical long-tail scenarios by proposing the first closed-loop minimax optimization framework that formulates adversarial training as a zero-sum Markov game. The approach jointly optimizes driving policies and adversarial scenarios through iteratively evolved adversaries driven by preference learning, with theoretical guarantees of convergence to a Nash equilibrium and maximization of a certified lower bound on real-world performance. Experimental results demonstrate that the framework effectively uncovers diverse safety-critical failure modes and significantly enhances the generalization and robustness of both policy and motion planners under previously unseen long-tail risks.

Technology Category

• Multiagent Systems: Adversarial Agents
• Search and Optimization: Adversarial Search
• Machine Learning: Adversarial Learning & Robustness

Application Category

• Responsible Web: Machine-in-the-loop, human agency and autonomy
• Search and Retrieval-Augmented AI: Web learning to rank, online learning, and counterfactual learning for ranking
• Economics, Online Markets and Human Computation: Uses of LLMs and GenAI for marketplace design, bidding, and strategic interactions

Abstract

Deploying autonomous driving systems requires robustness against long-tail scenarios that are rare but safety-critical. While adversarial training offers a promising solution, existing methods typically decouple scenario generation from policy optimization and rely on heuristic surrogates. This leads to objective misalignment and fails to capture the shifting failure modes of evolving policies. This paper presents ADV-0, a closed-loop min-max optimization framework that treats the interaction between driving policy (defender) and adversarial agent (attacker) as a zero-sum Markov game. By aligning the attacker's utility directly with the defender's objective, we reveal the optimal adversary distribution. To make this tractable, we cast dynamic adversary evolution as iterative preference learning, efficiently approximating this optimum and offering an algorithm-agnostic solution to the game. Theoretically, ADV-0 converges to a Nash Equilibrium and maximizes a certified lower bound on real-world performance. Experiments indicate that it effectively exposes diverse safety-critical failures and greatly enhances the generalizability of both learned policies and motion planners against unseen long-tail risks.

Problem

• long-tail robustness
• autonomous driving
• adversarial training
• safety-critical scenarios
• policy optimization

Innovation

• closed-loop adversarial training
• min-max optimization
• zero-sum Markov game
• long-tail robustness
• iterative preference learning