With the upcoming space missions for exploration in Mars and returning to the Moon, the need of autonomy regarding rover navigation is discernible. In this regard, Airbus Defence and Space has been developing Guidance Navigation and Control (GNC) algorithms for missions like ESA's ExoMars and the Sample Fetch Rover. While those classical rigid algorithms work for their target missions, some challenges arise regarding their flexibility for future applications. Aspects like run time and robustness to camera shifts and other miscalibration effects need to be considered, which are very common e.g. due to harsh launcher, lander and rover operations.

Thus, Airbus has taken a different approach with FARNAV (Fast And Robust NAVigation): an image based Deep Learning solution to create navigation maps (NavMaps). FARNAV’s final goal is to outperform the classical GNC stack computer vision algorithm. This paper is our first step into proving the feasibility of creating NavMaps with Convolutional Neural Networks. The tested architectures and challenges encountered when applying Deep Learning to the problem are analysed and the results discussed, concluding that the problem is solvable by the AI. The results open the path to further analysis on the integration and algorithm development towards its final goal.