CoIRL-AD: Collaborative-Competitive Imitation-Reinforcement Learning in Latent World Models for Autonomous Driving

A new framework called CoIRL-AD combines imitation and reinforcement learning to improve the robustness of end-to-end autonomous driving systems, according to a paper posted on the arXiv preprint server. The approach targets persistent failures in long-tail and cross-city driving scenarios. The paper, submitted by Xiaoji Zheng and revised in June 2026, introduces a competitive dual-policy architecture that decouples imitation learning and reward optimization into separate actors [1][2]. The framework, detailed on the arXiv open-access repository, uses imagined future rollouts for long-horizon reward estimation and includes a competition mechanism that selectively transfers beneficial behaviors while keeping reinforcement learning anchored to expert-like driving [2]. Experiments on the nuScenes benchmark showed that CoIRL-AD consistently improved robustness over strong imitation-learning baselines, with especially large gains in cross-city generalization and long-tail scenarios [2]. The initial submission was made on October 14, 2025, with a file size of 12,000 KB; the revised version, uploaded on June 15, 2026, measured 12,147 KB [1]. Code for the project is available on GitHub [2]. The paper appears under the Computer Vision and Pattern Recognition category on arXiv, a repository that hosts over two million e-prints and receives roughly 24,000 submissions per month as of late 2024 [6]. Unlike peer-reviewed journals, arXiv provides moderated but not peer-reviewed distribution of preprints across fields including computer science, physics, and mathematics [6]. The platform also supports community-developed tools through arXivLabs, a framework launched in 2020 that allows third-party collaborators to build features such as citation explorers and code finders directly on article pages [4][5].