Experimental Learning for Traversability Estimation and Stochastic Motion Planning on a Planetary Rover

To safely navigate autonomously on unstructured terrain, a planetary rover requires the ability to: a) accurately and reliably estimate terrain traversability, and b) plan and execute safe paths to reach the desired goal. However, real-world considerations such as sensing limitations and uncertainties make these tasks particularly challenging in practice. This work proposes novel approaches to address some of these challenges, with experimental validation on a realistic platform and a Mars-analogue environment (Fig. 1).