Year: 2026
Type: conference-paper
Abstract: Diffusion models have recently emerged as a powerful deep learning approach for robot motion planning, capable of generating multimodal, high-quality trajectories in complex environments.However, their slow sampling speed poses a significant challenge for real-time applications, particularly in dynamic settings that require replanning at high frequencies.This challenge is addressed by implementing and comparing several acceleration strategies.Leveraging the resulting speedups, a dynamic planning benchmark is introduced, where obstacles move randomly through the environment.In this setting, replanning can be executed up to 100 Hz, utilizing a cost-based trajectory selection mechanism that balances goal proximity, predicted collision risk, and trajectory smoothness.Additionally, a straight-line seed injection heuristic is introduced in order to improve near-goal performance.Comparisons with warm-start-based replanning using diffusion models and MPC show that accelerated diffusion models outperform both methods in dynamic settings, highlighting their strength in generating multimodal solutions.
Source: Lirias
Institutions: KU Leuven, Flanders Make (Belgium)
Language: English
Subfield: Control and Systems Engineering
Field: Engineering
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7165284781