Computer Vision‐Based Navigation for Agricultural Machinery: Technologies, Challenges, and Trends

Visual navigation offers a low‐cost, information‐rich solution for agricultural robots, but its deployment is constrained by the diversity and uncertainty of agricultural environments. This review systematically evaluates key technologies—path extraction, tracking control, and multisensor integration—through the S 2 TAP framework, which examines five interconnected dimensions: Scenario structure, sensing modality, task abstraction, agent mobility, and performance trilemma (the inherent difficulty of simultaneously maximizing accuracy, efficiency, and robustness). The analysis reveals a fundamental dichotomy between wheeled agents (precision tracking) and tracked agents (disturbance rejection), and identifies deep learning’s accuracy gains at the cost of efficiency. Based on these trade‐offs, the paper proposes actionable pathways including geometry‐first perception, hardware‐aware neural architecture search (HW‐NAS), uncertainty‐aware control, and community‐driven standardization (AgriROS and digital twins). Emerging frontiers such as neuromorphic vision and federated learning are also discussed, providing a road map toward robust, scalable visual navigation for autonomous agriculture. This paper presents a structured narrative review that synthesizes the literature through a novel analytical framework rather than an exhaustive systematic review. The paper provides a good reference for the development of visual navigation technology for agricultural robots and vehicles.