With the acceleration of agricultural intelligent transformation, deep reinforcement learning (DRL), leveraging its adaptive perception and decision-making capabilities in complex environments, has emerged as a pivotal technology in advancing the intelligent upgrade of agricultural machinery and equipment. For example, in UAV path optimization, DRL can help UAVs plan more efficient flight paths to cover more areas in less time. To enhance the systematicity and credibility of this review, this paper systematically examines the application status, key issues, and development trends of DRL in agricultural scenarios, based on the research literature from mainstream Chinese and English databases spanning from 2018 to 2024. From the perspective of algorithm–hardware synergy, the article provides an in-depth analysis of DRL’s specific applications in agricultural ground platform navigation, path planning for intelligent agricultural end-effectors, and autonomous operations of low-altitude unmanned aerial vehicles. It highlights the technical advantages of DRL by integrating typical experimental outcomes, such as improved path-tracking accuracy and optimized spraying coverage. Meanwhile, this paper identifies three major challenges facing DRL in agricultural contexts: the difficulty of dynamic path planning in unstructured environments, constraints imposed by edge computing resources on algorithmic real-time performance, and risks to policy reliability and safety under human–machine collaboration conditions. Looking forward, the DRL-driven smart transformation of agricultural machinery will focus on three key aspects: (1) The first aspect is developing a hybrid decision-making architecture based on model predictive control (MPC). This aims to enhance the strategic stability and decision-making interpretability of agricultural machinery (like unmanned tractors, harvesters, and drones) in complex and dynamic field environments. This is essential for ensuring the safe and reliable autonomous operation of machinery. (2) The second aspect is designing lightweight models that support edge-cloud collaborative deployment. This can meet the requirements of low-latency responses and low-power operation in edge computing scenarios during field operations, providing computational power for the real-time intelligent decision-making of machinery. (3) The third aspect is integrating meta-learning with self-supervised mechanisms. This helps improve the algorithm’s fast generalization ability across different crop types, climates, and geographical regions, ensuring the smart agricultural machinery system has broad adaptability and robustness and accelerating its application in various agricultural settings. This paper proposes research directions from three key dimensions-“algorithm capability enhancement, deployment architecture optimization, and generalization ability improvement”-offering theoretical references and practical pathways for the continuous evolution of intelligent agricultural equipment.

The emergence of ResNet provides a powerful tool for training extremely deep networks. The core idea behind it is to change the learning goals of the network. It no longer learns new features from scratch but learns the difference between the target and existing features. However, the difference between the two kinds of features does not have an independent and clear meaning, and the amount of learning is based on the absolute rather than the relative difference, which is sensitive to the size of existing features. We propose a new network that perfectly solves these two problems while still having the advantages of ResNet. Specifically, it chooses to learn the quotient of the target features with the existing features, so we call it the quotient network. In order to enable this network to learn successfully and achieve higher performance, we propose some design rules for this network so that it can be trained efficiently and achieve better performance than ResNet. Experiments on the CIFAR10, CIFAR100, and SVHN datasets prove that this network can stably achieve considerable improvements over ResNet by simply making tiny corresponding changes to the original ResNet network without adding new parameters.

The emergence of ResNet provides a powerful tool for training extremely deep networks. The core idea behind it is to change the learning goals of the network. It no longer learns new features from scratch but learns the difference between the target and existing features. However, the difference between the two kinds of features does not have an independent and clear meaning, and the amount of learning is based on the absolute rather than the relative difference, which is sensitive to the size of existing features. We propose a new network that perfectly solves these two problems while still having the advantages of ResNet. Specifically, it chooses to learn the quotient of the target features with the existing features, so we call it the quotient network. In order to enable this network to learn successfully and achieve higher performance, we propose some design rules for this network so that it can be trained efficiently and achieve better performance than ResNet. Experiments on the CIFAR10, CIFAR100, and SVHN datasets prove that this network can stably achieve considerable improvements over ResNet by simply making tiny corresponding changes to the original ResNet network without adding new parameters.