This is a page that curates AI-related papers published worldwide. All content here is summarized using Google Gemini and operated on a non-profit basis. Copyright for each paper belongs to the authors and their institutions; please make sure to credit the source when sharing.
EvaDrive: Evolutionary Adversarial Policy Optimization for End-to-End Autonomous Driving
Created by
Haebom
Author
Siwen Jiao, Kangan Qian, Hao Ye, Yang Zhong, Ziang Luo, Sicong Jiang, Zilin Huang, Yangyi Fang, Jinyu Miao, Zheng Fu, Yunlong Wang, Kun Jiang, Diange Yang, Rui Fan, Baoyun Peng
Outline
EvaDrive is a novel multi-objective reinforcement learning framework for achieving human-like, iterative decision-making in autonomous driving. To overcome the limitations of existing generation-evaluation frameworks, it establishes a closed-loop co-evolution between trajectory generation and evaluation through adversarial optimization. A hierarchical generator combines autoregressive intention modeling and diffusion-based refinement to propose candidate paths, while a trainable multi-objective evaluator explicitly preserves diverse preference structures without reducing them to a single scalar. Guided by a Pareto frontier selection mechanism, this adversarial interaction enables iterative, multi-round refinement, maintaining trajectory diversity while avoiding local optima. It achieves state-of-the-art performance on the NAVSIM and Bench2Drive benchmarks.
Takeaways, Limitations
•
Takeaways:
◦
Solving the scalarization bias problem caused by scalar rewards in existing reinforcement learning methods.
◦
Explicitly maintain diverse preference structures through multi-objective evaluators.
◦
Implementation of closed-loop coevolution of trajectory generation and evaluation via adversarial optimization.
◦
Generate different driving styles (dynamically weighted without external preference data).
◦
Achieving SOTA performance on NAVSIM and Bench2Drive benchmarks.
•
Limitations:
◦
The paper does not explicitly mention the specific Limitations. Further research is expected to address potential issues, such as computational costs and generalization performance degradation, that may arise during actual implementation and application.
