Learning semantic traversability priors using diffusion models for uncertainty-aware global path planning

Ethan Fahnestock, Erick Fuentes, Philip R Osteen, Siddharth Ancha, Nicholas Roy

Keywords: Global navigation, a-priori unknown environments, diffusion models, uncertainty-aware

TL;DR: Explores learning environment priors and exploiting them to improve global navigation through unknown terrain

Abstract: Robots have limited sensor ranges, restricting what they can observe, complicating navigation through a-priori unknown environments. If environment structure is present, priors over this structure can extend the utility of local observations and improve navigation performance. In this work, we propose learning priors over the semantic structure of navigation environments using state-of-the-art generative diffusion models. We show that diffusion models can capture complex spatial dependencies in overhead semantic maps, and are able to infer the semantics of far-away unobserved regions conditioned on local semantics already observed by the robot. By sampling a diverse, multi-modal set of high-fidelity semantic maps that are consistent with observed regions, we are able to estimate far-field navigation costs in an uncertainty-aware manner. Our preliminary investigations suggest that diffusion-based uncertainty-aware navigation costs can enable a downstream global planner to find more efficient paths and improve navigation performance.

Submission Number: 22