A Review of Learning Off-Road Terrain Traversability for Autonomous Ground Vehicles

Abstract Autonomous ground vehicles (AGVs) and mobile robots have significantly impacted off-road operations such as search and rescue, planetary exploration, and agriculture. One of the most difficult challenges for these scenarios is understanding the extremely complex traversable terrain and how to avoid getting stuck/immobilized in such terrain. Terrain traversability is an interdisciplinary field that generates traversable maps by extracting features from multimodal sensors to plan and control the AGV over unstructured and nondeformable or deformable surfaces. In this article, we investigate state-of-the-art techniques that utilize learning-based models to predict terrain traversability using exteroceptive and proprioceptive sensing. The literature is enriched with probabilistic and risk-aware terrain traversability to handle aleatoric and epistemic uncertainties and challenges. In addition, we discuss various planning and control techniques for off-road terrain traversability and control over deformable terrain, considering model-based, inverse dynamics-informed thresholds-based, and learning-based terramechanics. The key aspects, advantages, limitations, and open research challenges for off-road terrain traversability for AGVs are critically examined.