A Review of Terrain Perception for Ground Mobile Robots in Unstructured Environments

As ground mobile robots rapidly expand from structured urban roadways to unstructured outdoor environments, their demand across diverse domains including resource exploration, planetary exploration, agricultural cultivation, disaster rescue, and field research has become increasingly prominent, with terrain perception emerging as a core technology for ensuring high mobility and safe autonomous operation of wheeled and tracked robots. In off-road scenarios characterized by the absence of prior environmental models, highly heterogeneous terrain mechanical properties, and significant variations in physical attributes, constructing interpretable terrain perception models and achieving effective mapping from heterogeneous observation spaces to robust decision control represents the current critical bottleneck. Robots must not only identify terrain categories but also quantify the influence of terrain on their dynamic behavior, providing essential inputs for real-time decision-making and adaptive control. This paper systematically reviews the research progress in terrain perception for ground mobile robots, establishing a comprehensive technological chain from multi-source perception, semantic and physical property estimation, traversability assessment, to adaptive decision control, and proposes an integrated perception-planning-control framework oriented toward unstructured environments. Methodologically, three mainstream technological approaches can be summarized. Firstly, terrain perception based on exteroceptive information from visual cameras, LiDAR, and millimeter-wave radar, achieving prospective analysis of terrain geometric features and semantic information. Secondly, terrain perception based on interoceptive information from inertial measurement units and drive torque, realizing real-time estimation of ground mechanical properties through dynamic response. Thirdly, cross-modal terrain perception integrating heterogeneous sensor information, achieving unified representation of semantic and physical properties, thereby directly linking environmental understanding with motion planning and execution. The article first analyzes the complex characteristics of unstructured terrain, including irregular surface geometric morphology, non-uniform soil mechanical properties, and environmental uncertainty, elucidating how these factors affect robot motion stability, traction, and energy consumption. Unlike structured road environments, off-road terrain induces elastic and plastic deformation, subsidence, slippage, and complex wheel-ground interactions, all of which must be considered in high-level decision-making and low-level control. Regarding algorithmic evolution, this paper encompasses the complete spectrum from classical machine learning based on feature engineering to advanced deep learning models, discussing semantic segmentation methods utilizing convolutional neural networks, fully convolutional networks, and vision Transformers, as well as data-driven frameworks for estimating key physical properties such as friction coefficient, stiffness, cohesion, and sinkage from proprioceptive signals or fused multimodal data. Combined with typical datasets including RELLIS-3D, RUGD, GOOSE, YCOR, and TartanDrive 2.0, this paper analyzes the critical role of multimodal synchronized data in supporting joint learning of semantic and physical properties, thereby bridging the gap between perception and executable control. The research focus further concentrates on the mechanisms of terrain perception technology in decision planning and motion control, including traversability map construction, cost map generation, terrain-aware velocity planning and adaptive trajectory optimization, and methods for integrating semantic and physical property maps into decision planning and motion control. The integration of these methods significantly enhances system real-time performance, robustness, and energy-efficient mobility capabilities. Research indicates that addressing cognitive uncertainty brought by unstructured environments through constructing multimodal fusion perception frameworks and closed-loop coupled models of perception, planning, and control constitutes the core pathway for enhancing ground mobile robot adaptability to complex terrains. Recent innovations in end-to-end integration, such as deep reinforcement learning for terrain-adaptive strategies, Transformer architectures directly generating trajectories from raw multimodal inputs, and semi-supervised or self-supervised methods addressing sparse annotation, are driving development in this field. Finally, this paper identifies cross-domain model transfer, dynamic environment adaptation, real-time computational constraints, and heterogeneous sensor stream integration mechanisms as major ongoing challenges, and envisions future directions for achieving truly all-terrain autonomous mobility, emphasizing the need for tight coupling of robust terrain perception, continuous physical property estimation, and adaptive decision and motion control to ultimately achieve safe, efficient, and intelligent navigation in diverse off-road environments.