2.5D Object Mapping using Gaussian Processes for Robot Navigation

Abstract

Mapping and planning are fundamental to robotic navigation in unknown environments. This work introduces a probabilistic framework that combines Gaussian processes (GPs) for 2.5D object modeling with an informative motion planner, using LiDAR-based measurements. The mapping approach employs a flexible, nonparametric representation to process 3D point cloud data to create compact volumetric representations through contours and heights, enabling robust shape estimation even from sparse data. Building on this, the GP representation-based informative motion planner incorporates information gain into the dynamic window approach (DWA) to enhance navigation performance. Simulations validate the framework by comparing its mapping accuracy with OctoMap and elevation map, and its planning efficiency with a baseline DWA.