PRESTO: Fast Motion Planning Using Diffusion Models Based on Key-Configuration Environment Representation

TL;DR

PRESTO addresses fast motion planning in unseen C-spaces by representing obstacles with a sparse set of key configurations and guiding a conditional diffusion model with TrajOpt-inspired costs to generate high-quality trajectory seeds. The seeds are refined through trajectory optimization, enabling efficient collision-free plans even in narrow passages and under limited compute. The approach outperforms both pure learning and pure planning baselines, while ablations demonstrate the advantages of C-space key-config representations and cost-informed training. The work enables practical, generalizable planning for high-dimensional robots and offers configurable trade-offs between planning quality and computation time.

Abstract

We introduce a learning-guided motion planning framework that generates seed trajectories using a diffusion model for trajectory optimization. Given a workspace, our method approximates the configuration space (C-space) obstacles through an environment representation consisting of a sparse set of task-related key configurations, which is then used as a conditioning input to the diffusion model. The diffusion model integrates regularization terms that encourage smooth, collision-free trajectories during training, and trajectory optimization refines the generated seed trajectories to correct any colliding segments. Our experimental results demonstrate that high-quality trajectory priors, learned through our C-space-grounded diffusion model, enable the efficient generation of collision-free trajectories in narrow-passage environments, outperforming previous learning- and planning-based baselines. Videos and additional materials can be found on the project page: https://kiwi-sherbet.github.io/PRESTO.

Figures (7)

• Figure 1: Overview of PRESTO. PRESTO aims to generate collision-free trajectories in complex, unseen C-spaces. First, we approximate these C-spaces using key configurations from prior data and generate trajectories based on this representation. A conditional diffusion model, trained with a motion planning loss, provides initial solutions that are subsequently refined through trajectory optimization.
• Figure 2: Trajectory generation pipeline of PRESTO. We obtain the environment representation for an unseen problem by checking the collision states at the key configurations used during training. Using the trained conditional diffusion model, we generate multiple trajectories conditioned on this representation and then select the least-colliding trajectory after post-processing.
• Figure 3: Main results. We report the success rate (%), the collision rate (%), and the penetration depth (m) across 180 problems. (Top) The evaluation environments feature consistent 3-tier shelf fixtures with randomized object positions that vary across levels. (Bottom) We show PRESTO's performance changes across domains and computational budgets compared to the baselines.
• Figure 4: Ablation study results. We report the success rate (%), the collision rate (%), and the penetration depth (m) averaged across 180 problems for PRESTO and the self-variant baselines. (Left) We show performance changes with varying post-processing iterations. (Right) We present the performance of trajectories directly generated by the diffusion models without post-processing.
• Figure 5: Ablation studies with guided-sampling. We report the success rate (%), the collision rate (%), and the penetration depth (m) averaged across 180 problems for PRESTO and the self-variant baselines with guided-sampling. (Left) We show performance changes with varying post-processing iterations. (Right) We present the performance of trajectories directly generated by the diffusion model without post-processing.