Toward End-to-End Autonomy in Agricultural Weeding Robots: A Review of Cross-Module Integration, Validation Gaps, and Sustained Operation

Agricultural weeding robots are progressing from single-function automation toward end-to-end autonomy that integrates task planning, field navigation, weed perception, precision execution, and operational support. This review synthesises 217 sources identified through structured searches of major multidisciplinary, engineering, and Chinese literature databases, supplemented by citation tracking, followed by documented screening, evidence-use coding, technical-theme mapping, and a review-defined L1–L5 validation-coverage assessment. Three main findings emerge. First, reported technical evidence is concentrated at physical closed-loop feasibility in real agricultural environments rather than at complete production service. Among 157 studies with classifiable technical validation, 78 (49.7%) attained L3 as their highest level, whereas only one met the predefined criteria for complete production service and sustained operation with quantified supervision; that study concerned area-based mowing rather than target-level selective weeding. A separate analysis of 49 direct robotic-weeding technical studies showed the same concentration at L3. In a scenario-based boundary-robustness analysis, the number of direct studies attaining L4 or above increased from 1/49 under the strict criteria to 2/49 and 3/49 under two deliberately relaxed scenarios, while L3 remained the dominant highest level. Second, validation depth is uneven: navigation, trajectory control, and several execution modalities have stronger real-field evidence than global task planning, weed perception, fault diagnosis, and autonomous replenishment. Third, the principal system bottleneck is increasingly cross-module rather than purely algorithmic: spatial, temporal, and task-state inconsistencies can propagate through interfaces into missed or repeated treatment, crop damage, unfinished tasks, additional downtime, and human intervention. The review therefore argues for validation centred on recoverable complete production tasks, with explicit reporting of task completion, agronomic outcomes, faults and recovery, replenishment, downtime, resource use, and supervision.