Accelerating Diffusion Models for Adaptive Motion Planning in Dynamic Environments

Matthias Van Eysendeyk, Edward Sandra, Dries Dirckx, Wilm Decré

Keywords: Motion Planning, Diffusion Models, Dynamic Environments

Abstract:

Diffusion models have recently emerged as a powerful approach for robot motion planning, capable of generating multimodal, high-quality trajectories. Their main limitation is slow sampling, which hinders real-time replanning in dynamic environments. This work studies several acceleration strategies for diffusion-based motion planning and leverages the resulting speedups to enable replanning at up to 100 Hz in a dynamic benchmark with randomly moving obstacles. A cost-based trajectory selection mechanism is used to exploit multimodality during
replanning by balancing motion time, predicted collision risk, and smoothness. Comparisons against warm-start diffusion and MPC show that accelerated diffusion models achieve the best performance in dynamic settings.

Submission Number: 43