Generative Trajectory Planning in Dynamic Environments: A Joint Diffusion and Reinforcement Learning Framework

ICLR 2026 Conference Submission18881 Authors

Keywords: Diffusion model, Reinforcement learning, trajectory optimization

Abstract: Real-time trajectory optimization requires planners that can simultaneously ensure safety and energy efficiency in environments containing both static and dynamic obstacles. This paper introduces a generalized framework that combines diffusion-based trajectory generation with deep reinforcement learning (DRL). The diffusion component generates diverse candidate trajectories by modeling feasible sub-paths, where a sub-path denotes a short-horizon segment aligned with receding-horizon execution. In this formulation, the entire trajectory is decomposed into consecutive sub-paths, enabling the diffusion model to learn local collision avoidance and smoothness while maintaining consistency across the fully identified path (e.g., global path and whole trajectory). The DRL component then evaluates these candidates online, selecting actions that improve safety while adapting to dynamic obstacles and maintaining energy-efficient behavior. The joint design leverages the generative diversity of diffusion and the adaptive decision-making of DRL, producing a planner that is both responsive and reliable. To assess effectiveness, the method is evaluated in unmanned aerial vehicle (UAV) path optimization scenarios with dynamic obstacles. The results demonstrate that sub-path training enhances the generalization of diffusion-based planners by linking local feasibility to global performance, and that the approach offers a practical solution for real-time UAV trajectory optimization with improved safety and efficiency.

Primary Area: learning on time series and dynamical systems

Submission Number: 18881