abstract
- This paper introduces a novel regularization method to enhance trajectory generation for 6-Degrees of Freedom (DoF) robotic motion planning. The proposed approach integrates diffusion policies with conditional diffusion processes and a Mutual Information Neural Estimator (MINE) to improve trajectory-following accuracy. It generates 6-DoF action sequences from an initial noise distribution via a diffusion policy, refining them with a MINE. This enables learning of complex, nonlinear features in high-dimensional configuration spaces while optimizing the gradient field of an energy-based model for smooth and continuous motion trajectories. Experimental results demonstrate that this framework significantly enhances accuracy, allowing the use of smaller models without performance loss. These advancements create new opportunities for efficient and adaptive robotic control in real-world applications.