Motion Planning Diffusion: Learning and Planning of Robot Motions with Diffusion Models

TL;DR

This work addresses accelerating robot motion planning by learning priors over trajectories with diffusion models. It casts planning as inference and trains a diffusion prior on expert trajectories, enabling posterior sampling via guided reverse diffusion to produce low-cost, collision-free plans. The key contributions include introducing Motion Planning Diffusion (MPD), a planning‑as‑inference framework that combines diffusion priors with cost guidance, and extensive evaluation showing improved success rates and multimodality over CVAE priors, including generalization to unseen obstacles and real-world demonstration on a Panda arm. The approach demonstrates that diffusion priors can speed up planning and provide diverse, high-quality trajectories for high-dimensional robotics applications, with potential impact on real-time and robust autonomous systems.

Abstract

Learning priors on trajectory distributions can help accelerate robot motion planning optimization. Given previously successful plans, learning trajectory generative models as priors for a new planning problem is highly desirable. Prior works propose several ways on utilizing this prior to bootstrapping the motion planning problem. Either sampling the prior for initializations or using the prior distribution in a maximum-a-posterior formulation for trajectory optimization. In this work, we propose learning diffusion models as priors. We then can sample directly from the posterior trajectory distribution conditioned on task goals, by leveraging the inverse denoising process of diffusion models. Furthermore, diffusion has been recently shown to effectively encode data multimodality in high-dimensional settings, which is particularly well-suited for large trajectory dataset. To demonstrate our method efficacy, we compare our proposed method - Motion Planning Diffusion - against several baselines in simulated planar robot and 7-dof robot arm manipulator environments. To assess the generalization capabilities of our method, we test it in environments with previously unseen obstacles. Our experiments show that diffusion models are strong priors to encode high-dimensional trajectory distributions of robot motions.

Figures (4)

• Figure 1: Execution of motions in the real-world Panda Shelf environment. The motion in blue and red start and end at the same configurations, but it is possible to see two modes resulting from sampling trajectories with MPD. This environment includes obstacles (represented as boxes in the digital twin), which are not present in the environment used for training (Fig. \ref{['fig:environments:pandapickandplace']}).
• Figure 2: The environments considered in our experiments include robot navigation tasks of a point mass in $2$D and $3$D, and a $7$-dof Franka Emika Panda manipulator. In (a) and (b), the green and red dots are the initial and goal states, respectively, and the blue line is the result of RRT Connect.
• Figure 3: (a)-(d) Diffusion steps of MPD on a batch of $100$ trajectories in the PointMass2D Dense - Extra Obstacles environment. Notice how noise transforms into multimodal, smooth and collision-free trajectories. (e) Trajectories generated by CVAEPosterior. Obstacles in red were not present in the training environment. Trajectories in orange are collision-free, black in collision. The start and goal configurations are in green and blue.
• Figure : Motion Planning Diffusion