Abstract: Agricultural robots have been increasingly recognized as key intelligent equipment for improving agricultural productivity, operation quality, and sustainable development in smart farming. Compared with conventional agricultural machinery, agricultural robots are required to integrate environmental perception, autonomous decision-making, and precise execution in complex field operations. However, agricultural scenarios are highly open, dynamic, and unstructured. Significant spatiotemporal variations in soil conditions, crop morphology, terrain characteristics, operating loads, and meteorological factors introduce substantial uncertainty into robotic perception, decision-making, and control. These characteristics make agricultural robots representative crop-robot-environment coupled systems. As a result, their intelligent operation is still restricted by incomplete sensing information, insufficient model accuracy, limited adaptability of control strategies, low repeatability of field validation, and weak coordination among robots, tasks, and environments. Digital twin technology provides a promising approach to addressing these problems by establishing a dynamic connection between physical entities and virtual models through real-time data interaction, model updating, and predictive simulation. By constructing virtual representations of agricultural robots, operating environments, and task processes, digital twins can support state monitoring, process prediction, strategy evaluation, and control optimization, while reducing the dependence on costly and time-consuming field trials. In this study, a systematic review was conducted on the technical pathways and future prospects of digital twin-empowered agricultural robots. The evolution of agricultural machinery from traditional equipment to intelligent robotic systems was first reviewed, and the core characteristics of agricultural robots were summarized from the perspectives of perception, decision-making, and control. Furthermore, the concept, system elements, and functional value of digital twins were analyzed to clarify their applicability to agricultural robotic systems under complex field conditions. Based on this analysis, the empowerment pathways of digital twins were summarized within a perception-decision-control framework. For perception enhancement, digital twins can integrate robot state data, environmental information, historical operation data, and mechanism models to construct a dynamic virtual mapping of agricultural operations. Multi-source data fusion, virtual sensing, state reconstruction, and anomaly correction can be used to improve the completeness, reliability, and interpretability of sensing information. For decision optimization, digital twins can provide a controllable and evaluable virtual environment for operation-process simulation, multi-strategy comparison, intelligent algorithm training, and policy iteration. This enables path planning, task allocation, operation-parameter matching, and risk avoidance to shift from experience-based decision-making to model-supported and prediction-based optimization. For control improvement, digital twins can connect decision outputs, execution states, external disturbances, and control constraints through online model updating, disturbance prediction, and adaptive parameter tuning, thereby supporting control compensation, parameter adaptation, and closed-loop optimization. Representative application scenarios were then analyzed, including fault prediction and health management, autonomous operation, and operation scheduling and coordination. In fault prediction and health management, digital twins integrated real sensing data with virtual simulation data to enable key component monitoring, anomaly detection, performance degradation prediction, and predictive maintenance. In autonomous operation, digital twins supported virtual testing and optimization of motion-control and operation-control algorithms, enabling agricultural robots to adjust trajectories, operating parameters, and execution strategies according to field conditions. In operation scheduling and coordination, digital twins can model field environments, robot fleets, task processes, and environmental constraints to evaluate scheduling schemes and optimize task allocation, path coordination, and resource deployment. The research results indicate that digital twins not only enhance the role of agricultural robots as information hubs within the agricultural production chain, but also accelerate the transition of agricultural operations from mechanization toward data-driven and model-enabled intelligent systems. Finally, this review summarizes the challenges faced by the current application of digital twin technology and prospects their future development directions in agricultural robotics. This review contributes to a comprehensive understanding of how digital twin empowers agricultural robots and provides theoretical and practical guidance for future research.