| dc.description.abstract | In this thesis, the feasibility of a deep-learning-based terrain characterization method was assessed in comparison to a traditional analytical approach. Both solutions were implemented on a wheeled mobile robot equipped with a standard stereo depth camera, an IMU, and a GPS unit. Classical technique was derived from existing literature, while a deep-learning based implementation was developed alongside the navigation system and data processing utilities. The study conducted extensive experiments in a real-world setting located in an unstructured forest environment and gathered results in the form of energy consumption and roughness. While both methods proved effective at navigation from point A to point B, inherent limitations highlighted avenues for future advancements. Key findings include the need for an extensive, high-resolution dataset to optimize machine learning performance and a more robust navigational with global perception. The outcomes of this research pave the way for future exploration into refining terrain characterization techniques for more diverse environments and applications. | en_US |
