Autonomous navigation plays an increasingly crucial role in rover-based planetary missions. End-to-end
navigation approaches developed upon deep reinforcement learning have been a rising research topic with
great adaptability in complex environments. However, most existing works focus on geometric obstacle
avoidance thus have limited capability to cope with ubiquitous non-geometric hazards in the wild, such as
slipping and sinking, caused by insufficient mechanical properties of the terrain. Autonomous navigation
in unstructured, harsh environments remains a great challenge requiring further in-depth study.
In this paper, a DRL-based navigation method is proposed to autonomously guide a planetary rover towards
goals via hazard-free paths with low wheel slip ratios. We introduce an end-to-end network architecture,
in which the visual perception and the wheel-terrain interaction are fused to learn the representation of
terrain mechanical properties implicitly and further facilitate policy learning for non-geometric hazard
avoidance. Our approach outperforms baseline methods in simulation evaluation with superior avoidance
capabilities against geometric obstacles and non-geometric hazards. Experiments conducted at a Mars
emulation site suggest the successful deployment of our approach on a planetary rover and the capacity of
dealing with locomotion risks in real-world navigation tasks.
