HADP: Hybrid A*-Diffusion Planner for Robust Navigation in Dynamic Obstacle Environments

Autonomous driving in dynamic environments requires the ability to rapidly perceive changes and make safe, efficient decisions. While numerous studies have explored learning-based planning methods to address this challenge, diffusion-based planning methods have recently emerged as a promising approach in robotic path planning by modeling action sequences as probabilistic trajectories and capturing diverse planning behaviors under uncertainty. However, they still face significant challenges in generalizing to unseen situations and demand extensive data collection to achieve robust performance. In particular, efforts to ensure general robustness by learning from all possible scenarios can introduce inefficiencies, as many cases may be irrelevant during actual deployment. To address these challenges, this study proposes the Hybrid A*-Diffusion Planner (HADP), designed to achieve robust navigation performance even in unseen dynamic environments. In obstacle-free environments, HADP employs the A* algorithm for global path planning, whereas when a dynamic obstacle is detected within a predefined boundary, it records observations and leverages a conditional diffusion model to generate a local avoidance path. After avoidance, the path is re-planned using A*. To enable the diffusion model to generate effective avoidance paths, a semantic map is constructed by incorporating key environmental information within the predefined boundary. This map, along with the robot’s position and local goal information, serves as a conditional input to the diffusion model. HADP is trained on an expert trajectory dataset created by recording expert-driven navigation across various dynamic obstacle avoidance scenarios. Experiments in various simulations in dynamic obstacle environments demonstrated HADP’s superior success rate, trajectory smoothness, and robust generalization performance in unseen environments compared to existing approaches. Real-world tests further confirmed its practical applicability and highlighted its potential as a robust and scalable solution for dynamic obstacle environments, and the simulation code used in this study is publicly available at https://github.com/DSLab-MJU/HADP