Diffusion-Based Behavioural Cloning for Aerial Informative Path Planning

Abstract

Informative path planning (IPP) tasks UAV with locating and accurately measuring high concentrations of specific phenomena in initially unexplored, budget constrained search space. Adaptive-IPP closes the loop between sensing and planning by replanning sensor trajectories as a probabilistic belief about the environment evolves but even state-of-the-art planners face two challenges while replanning their trajectory online: 1.) Replanning during deployment incurs substantial online computation and thus time costs during deployment, which is problematic in a time-budgeted mission and 2.) the resulting path-optimization landscape is non-convex, producing a fundamentally multimodal set of valid solutions. Behaviorally cloning such an expert with a deterministic, MSE-trained policy mode-averages these equally valid trajectories, producing plans that correspond to none of them and fail to learn the underlying monitoring objective in unseen environments.This thesis presents a path-planning framework that instead uses a conditional denoising diffusion model to generate valid 3D waypoint sequences for continuous environments. A novel CMA-ES-based data collection procedure exposes multiple deployable expert trajectories per belief state, giving the diffusion model a genuinely multimodal demonstration set to learn from. We tested our diffusion based model in simulated missions on synthetic Gaussian random fields and real-world agriculture maps based on the National Agricultural Imagery Programme (NAIP) and found that diffusion reduces time-integrated map uncertainty by roughly 34\% relative to the online replanning expert that it was trained on, while replanning approximately 11× faster by appropriately learning the goal of online replanning through the multimodal dataset, and consistently outperforms a deterministic learning baseline trained on the same demonstrations. These results suggest that diffusion-based behavioral cloning preserves valid planning modes and enables real-time, non-myopic informative path planning.