PRISM: Multimodal Terrain Mapping for Rover Navigation in Unstructured Environments

The paper introduces PRISM, a multimodal terrain-perception system for rover navigation in unstructured environments. Its custom sensor suite aligns RGB, depth, and thermal imagery, while OmniUnet, a vision-transformer-based network, performs multimodal semantic terrain segmentation. The authors evaluate the system on two newly annotated datasets, BASEPROD and LAENTIEC, and demonstrate field deployment on a resource-constrained embedded computer. PRISM converts the perception output into traversability maps that can be consumed by a rover’s Guidance, Navigation, and Control subsystem for autonomous navigation. The abstract emphasizes thermal sensing as a complement to optical and depth data for terrain differentiation.