For my master's thesis we investigated the fusion of geometric and a visual data that estimate traversability using custom trained neural nets. 

We focus on unstructured, outdoor environments for robot hiking purposes with the ANYmal quadruped. A traversability semantic segmentation model using image data was developed. This was projected onto a multi-modal elevation map, and fused with score from a geometric traversability estimation network to create a cost map. A exploration module selects the best goal to follow in the absence of human input, and a path is computed by an MPPI local planner.

We optimize this for onboard deployment, such that the entire pipeline can be run at a rate of over 6Hz. We show that, using this traversability cost map, exploration based goals can lead to comparable performance of in hiking trail path planning as human selected goals, as well as the feasibility of this pipeline for autonomous hiking.