Real-Time Adaptive Motion Planning via Point Cloud-Guided, Energy-Based Diffusion and Potential Fields

TL;DR

The paper tackles real-time motion planning under partial observability in environments with static obstacles and dynamic pursuers. It presents a hierarchical approach that fuses energy-based diffusion models with artificial potential fields, conditioned on point-cloud obstacle encodings and implemented via a temporal U-Net with cross-attention; classifier-free guidance enables flexible conditioning, while local APF refinement provides reactive obstacle avoidance. The framework supports compositional sampling to handle unseen obstacle configurations and employs dynamic refinement to adapt trajectories as pursuers move, storing executed history to condition future planning. Experimental results across static mazes, pursuit-evasion tasks, 3D environments, and physical RC car demonstrations show high success rates, reduced collisions, and fast iteration times, outperforming several baselines in both efficiency and robustness.

Abstract

Motivated by the problem of pursuit-evasion, we present a motion planning framework that combines energy-based diffusion models with artificial potential fields for robust real time trajectory generation in complex environments. Our approach processes obstacle information directly from point clouds, enabling efficient planning without requiring complete geometric representations. The framework employs classifier-free guidance training and integrates local potential fields during sampling to enhance obstacle avoidance. In dynamic scenarios, the system generates initial trajectories using the diffusion model and continuously refines them through potential field-based adaptation, demonstrating effective performance in pursuit-evasion scenarios with partial pursuer observability.

Figures (11)

• Figure 1: Overview of our approach: (1) Initial energy-based diffusion trajectory planning, where light-blue regions represent high-energy potential fields (obstacles), lightblue dots trace the evader's (green star) state history, red sphere indicates a dynamic pursuer, and dark purple areas show safe, low-energy navigation spaces. (2) When the pursuer approaches the evader, local path exploration is performed via small noise perturbations. (3) Final trajectory refinement using denoising and APF, demonstrating adaptive path planning that avoids obstacles.
• Figure 2: Trajectory planning pipeline for a Pursuit-Evasion case: In the high-level planner, obstacle point clouds are processed by a pre-trained encoder to obtain a low-dimensional encoding of the point cloud features. The diffusion model then generates multiple candidate trajectories conditioned on start/goal states and these obstacle features. The best trajectory is passed to the low-level planner, which performs refinement when a pursuer approaches the evader. This refinement combines denoising with APF while conditioning on the evader's state history. The refined trajectory is then used as input for the next planning iteration, creating an adaptive system that continuously responds to the pursuer's movements while navigating through the environment.
• Figure 3: Trajectory generation comparison on Maze2D in the presence of unseen objects. left: MDP fails to avoid the obstacles. right: Our method successfully avoids them.
• Figure 4: Trajectory generation on Maze3D environment with box and sphere obstacles.
• Figure 5: Pursuit-evasion simulation: Our method (top) successfully adapts to unseen obstacle configurations while the DRL-based SAC method (bottom) struggles. Orange lines show evader path, red circles represent pursuers with detection zones (light yellow circular regions), green areas indicate safe zones, and green/purple dots mark start/goal states.