Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Accelerating db-A* for Kinodynamic Motion Planning Using Diffusion

Julius Franke, Akmaral Moldagalieva, Pia Hanfeld, Wolfgang Hönig

TL;DR

This work addresses kinodynamic motion planning by learning diffusion-model priors that generate problem-instance conditioned motion primitives to accelerate db-A* planning. The diffusion models are trained on ground-truth primitives derived from randomly generated problem instances solved with a kinodynamic planner, enabling generation of problem-tailored trajectories. Empirical results show up to about 30% improvements in planning time and solution quality across multiple robot dynamics (notably the $2^{\text{nd}}$-order unicycle and car with trailer), demonstrating the practical impact of instance-conditioned diffusion priors. The approach includes dataset construction, conditioning on both solution-specific and environment-specific features, and a deployment strategy with model caching and precomputation, with future work extending to more dynamics and richer workspace representations.

Abstract

We present a novel approach for generating motion primitives for kinodynamic motion planning using diffusion models. The motions generated by our approach are adapted to each problem instance by utilizing problem-specific parameters, allowing for finding solutions faster and of better quality. The diffusion models used in our approach are trained on randomly cut solution trajectories. These trajectories are created by solving randomly generated problem instances with a kinodynamic motion planner. Experimental results show significant improvements up to 30 percent in both computation time and solution quality across varying robot dynamics such as second-order unicycle or car with trailer.

Accelerating db-A* for Kinodynamic Motion Planning Using Diffusion

TL;DR

This work addresses kinodynamic motion planning by learning diffusion-model priors that generate problem-instance conditioned motion primitives to accelerate db-A* planning. The diffusion models are trained on ground-truth primitives derived from randomly generated problem instances solved with a kinodynamic planner, enabling generation of problem-tailored trajectories. Empirical results show up to about 30% improvements in planning time and solution quality across multiple robot dynamics (notably the -order unicycle and car with trailer), demonstrating the practical impact of instance-conditioned diffusion priors. The approach includes dataset construction, conditioning on both solution-specific and environment-specific features, and a deployment strategy with model caching and precomputation, with future work extending to more dynamics and richer workspace representations.

Abstract

We present a novel approach for generating motion primitives for kinodynamic motion planning using diffusion models. The motions generated by our approach are adapted to each problem instance by utilizing problem-specific parameters, allowing for finding solutions faster and of better quality. The diffusion models used in our approach are trained on randomly cut solution trajectories. These trajectories are created by solving randomly generated problem instances with a kinodynamic motion planner. Experimental results show significant improvements up to 30 percent in both computation time and solution quality across varying robot dynamics such as second-order unicycle or car with trailer.

Paper Structure

This paper contains 16 sections, 8 equations, 2 figures, 3 tables, 1 algorithm.

Table of Contents

  1. Introduction
  2. Related Work
  3. Problem Definition
  4. Background
  5. Motion Primitives
  6. Kinodynamic Motion Planning with db-A*
  7. Diffusion Models
  8. Approach
  9. Dataset Creation
  10. Conditioning
  11. Training
  12. Deployment
  13. Results
  14. Benchmarking
  15. Ablation studies
  16. ...and 1 more sections

Figures (2)

  • Figure 1: Examples for the $2^{\text{nd}}$ order unicycle - Top left: Bugtrap with solution (red: start, green: goal), Top right: Ten sampled motion primitives (starting from origin) Bottom row: Two random instances with a solution found by our method (red: start, green: goal)
  • Figure 2: Violin plots of regret for duration and costs for three dynamics. The introduced diffusion model outperforms the baseline across different robot dynamics and in all metrics.