Cost-Ensemble-Guided Asymmetric Diffusion Model for Trajectory Planning of Manipulator

This paper proposes a diffusion model-based framework for manipulator trajectory planning, aimed at generating smooth, collision-free trajectories that satisfy both dynamic and kinematic constraints. Traditional trajectory planning methods often suffer from low sample efficiency, challenges in ensuring path smoothness, reliance on well-initialized paths, and a tendency to converge to local optima. While deep learning approaches show promise, they struggle to balance diversity and accuracy. To address these issues, this paper introduces an asymmetric diffusion mechanism that leverages the strengths of both latent and original task spaces. Forward diffusion ensures the kinematic feasibility of trajectories, while reverse diffusion enables high-resolution trajectory optimization. Additionally, a multi-cost function is employed to guide the diffusion process, ensuring that generated trajectories meet diverse task requirements. Experimental results demonstrate that the proposed framework outperforms existing methods in trajectory planning tasks, achieving higher success rates and real-time performance, making it suitable for real-time task demands in industrial scenarios.