Prior-Constrained Explorative Guidance for Generalization in Diffusion Motion Planning
Sunhwi Kim, Junsu Kim, Seungjae Baek, Jaechan Shin, Jungeun Lee, Seongjae Lee, Kyungdon Joo, Jeong hwan Jeon
AI summary
Problem
Diffusion-based motion planners struggle to generalize to unseen environments and face an unresolved trade-off between sample diversity and conditioning consistency during inference-time guidance.
Approach
The method estimates environmental gradients via local trajectory perturbations and blends Gaussian Process noise with Gaussian noise during reverse diffusion to constrain exploration and preserve solution diversity.
Key results
- Up to 30 percentage point success rate improvement over baselines
- Preserves trajectory smoothness and collision avoidance in unseen environments
- Successful real-world deployment in novel environments without retraining
- Computational efficiency maintained through parameter scheduling
Why it matters
Enables reliable, generalizable motion planning for robotics applications in dynamic or unstructured environments without costly retraining.
Abstract
Diffusion-based planners have achieved gener- alization comparable to classical planners by leveraging inference-time optimization through guidance. However, their limited ability to capture environmental variations often con- strains their responsiveness in unseen settings. In addition, the diversity-consistency trade-off inherent in guidance has re- mained unresolved. In this work, we propose Prior-Constrained Explorative Guidance (PCEG), a novel approach that gathers environmental information through local exploration and pre- vents guided samples from converging prematurely to similar solutions by leveraging a trajectory prior. The collected in- formation is included in the guidance via stochastic gradient estimation, while a succinct parameter scheduling strategy enables latent optimization driven by environmental signals without significant computational overhead. Furthermore, dur- ing the modal-seeking stages of the reverse diffusion process, we employ a Gaussian Process (GP) to enforce dynamics-informed priors, effectively constraining the exploration region of each sample and thereby enhancing solution diversity. Across diverse benchmarks including 7-degree-of-freedom (7-DoF) robot-arm manipulation, PCEG substantially improves success rate by up to 30 percentage points compared to competitive diffusion planners without compromising trajectory quality, even in scenarios involving unseen obstacles. Real-world experiments further validate these findings, showcasing the generation of smooth, collision-free trajectories in novel environments. The project page is available at https://rml-unist.github. io/PCEG/.