Diffusion Model and RRT*-Based Methods for Reflected Mass Optimization in Motion Planning
David Gutierrez-Moreno, Simon Armleder, Yuhwan Kwon, Takumi Hachimine, Yoshihisa Tsurumine, Takamitsu Matsubara, Gordon Cheng · 2025
Optimizing a robot’s posture can be advantageous for managing interaction forces with the environment. By optimizing the Reflected Mass (RM) along entire trajectories, the robot’s posture can be adjusted to minimize impact forces for safety or maximize them for tasks that require high force, such as pushing or striking. However, the integration of RM optimization within motion planning remains underexplored. To address this, we introduce two new approaches for optimizing RM in motion planning: a probabilistic generative model based on diffusion techniques and a sampling-based method using Rapidly-Exploring Random Trees Star (RRT*). Both methods optimize the RM within the motion planning framework, enabling new strategies for enhancing robot interactions in diverse and dynamic environments. Experimental validation in simulation and on a UR5 robot demonstrates the effectiveness of these approaches in controlling RM, offering promising directions for future research and applications.