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

Learning-based navigation methods enable rapid perception and safe decision-making in dynamic obstacle environments but demand vast amounts of diverse scenario data to achieve robust generalization.

Unlike conventional methods that predict single-step action policies, diffusion-based trajectory generation can produce path-level action policies—enabling smoother, more effective driving by planning whole trajectories at once.

We propose the Hybrid A*-Diffusion Planner (HADP), which integrates A* search for global planning with conditional diffusion-based path planning for local obstacle avoidance.

In obstacle-free regions, HADP uses A* for optimal global path planning.

When a dynamic obstacle enters a predefined boundary, the system builds a semantic map and, conditioned on this map plus the robot’s pose and local goal, invokes a diffusion model to generate a local avoidance path then, after avoidance, re-plans globally with A*.

We conducted comparative experiments in both SEEN and UNSEEN environments against other learning-based navigation methods to demonstrate HADP’s superiority, and the results showed that HADP achieved the highest success rates in both settings.

Furthermore, the proposed method also showed reliable performance in real-world robot experiments.