Visual Perception and Autonomous Navigation Control for Orchard Robots: From Scene Understanding to Closed-Loop Motion Execution

Autonomous navigation is essential for orchard robots performing spraying, weeding, inspection, harvesting, and transport, yet semi-structured orchards combine repeatable tree-row geometry with canopy occlusion, uneven illumination, weak road boundaries, terrain variation, and intermittent global positioning. This review examines orchard visual navigation through a control-oriented chain of scene information, perception and localization, navigation-reference construction, motion execution, and closed-loop evaluation. Particular emphasis is placed on orchard-specific interfaces: repeated row geometry can stabilize reference extraction but can also create perceptual ambiguity; canopy occlusion can simultaneously degrade visual features and GNSS availability; and headland turning requires explicit switching among row following, turning, relocalization, and task resumption. Passable areas, tree-row/canopy structure, and obstacle/end-of-row semantics are compared according to the control variables they generate, including drivable corridors, lateral and heading errors, curvature, safety clearance, event states, and tracking targets. Traditional vision, deep-learning detection and segmentation, depth vision, visual SLAM, multi-sensor fusion, and path-tracking control are synthesized in terms of operating conditions, failure modes, and downstream control effects. The review further analyzes perception-to-control error propagation, end-to-end latency, cross-season generalization, and the mismatch between image metrics and field navigation performance. A unified perception–representation–localization–planning–control–evaluation framework and full-chain benchmark protocol are presented.