The paper introduces LDP (Local Diffusion Planner), a novel approach for efficient robot navigation and collision avoidance using diffusion models. The authors address the challenges of data urgency and myopic observation in local navigation by collecting diverse expert policy data from multiple scenarios and preferences. They train a diffusion agent using Denoising Diffusion Probabilistic Models (DDPM) to generate collision-free action sequences. LDP incorporates global paths as additional conditions to enhance the planner's ability to navigate through complex environments. Experimental results demonstrate that LDP outperforms baseline algorithms in terms of success rate, runtime, and success weighted by path length (SPL), showing superior robustness and generalization capabilities in various scenarios. The method is validated through real-world experiments on an Ackermann steering robot, highlighting its practical value. Future work includes improving data quality and real-time performance.The paper introduces LDP (Local Diffusion Planner), a novel approach for efficient robot navigation and collision avoidance using diffusion models. The authors address the challenges of data urgency and myopic observation in local navigation by collecting diverse expert policy data from multiple scenarios and preferences. They train a diffusion agent using Denoising Diffusion Probabilistic Models (DDPM) to generate collision-free action sequences. LDP incorporates global paths as additional conditions to enhance the planner's ability to navigate through complex environments. Experimental results demonstrate that LDP outperforms baseline algorithms in terms of success rate, runtime, and success weighted by path length (SPL), showing superior robustness and generalization capabilities in various scenarios. The method is validated through real-world experiments on an Ackermann steering robot, highlighting its practical value. Future work includes improving data quality and real-time performance.