Hu, Dong (The Hong Kong Polytechnic University), Zhou, Yangyang (The Hong Kong Polytechnic University), Hu, Fengqing (The Hong Kong Polytechnic University), Huang, Chao (The Hong Kong Polytechnic University), Wu, Jingda (Nanyang Technological University), Savkin, Andrey (University of New South Wales)

Generalizing Autonomous Navigation Via Human-Guided Diffusion Policies with Adversarial Reinforcement Learning

Scheduled for presentation during the Regular Session "S27a-Safety and Risk Assessment for Autonomous Driving Systems" (TH-LM-T27), Thursday, November 20, 2025, 10:30−10:50, Broadbeach 3

2025 IEEE 28th International Conference on Intelligent Transportation Systems (ITSC), November 18-21, 2025, Gold Coast, Australia

This information is tentative and subject to change. Compiled on October 18, 2025

Abstract

To address the challenges of autonomous vehicle navigation in dynamic and complex environments—particularly in LiDAR-free and low-cost settings—this study proposes an online adversarial reinforcement learning (RL) framework that integrates real-time human guidance with diffusion-based policies. Unlike prior diffusion RL works, our method enables stable online training of diffusion policies by leveraging human guidance to stabilize data distribution and improve early-stage exploration. Furthermore, adversarial training is introduced to enhance robustness and generalization in dynamic multi-task scenarios. The framework uses only onboard cameras without reliance on prior maps or expensive sensors, allowing for cost-effective deployment. Experimental results demonstrate that our approach substantially enhances adaptability and multi-task generalization, outperforming state-of-the-art baselines in training efficiency, success rate, and safety across various dynamic and complex environments.