Abstract: Traditional autonomous driving systems for agricultural machinery encounter a multitude of significant limitations when operating within complex and unstructured environments, prominently including prohibitive costs associated with high-precision sensors, signal susceptibility to interference from canopy cover or terrain features, and poor environmental adaptability when faced with dynamic or unforeseen conditions. As the intelligent transformation of agriculture accelerates in tandem with rapid advancements in artificial intelligence, the end-to-end (E2E) autonomous driving paradigm offers a powerful and innovative solution for intelligent navigation and precise operations in these challenging settings. This paper provides a systematic review of the theoretical adaptability of E2E autonomous driving technology and conducts a full-path analysis of its application scenarios. By employing a combined approach of literature survey and technical analysis, systematically examining relevant research from multiple dimensions including technological evolution, methodological principles, typical application scenarios, and key technical challenges. The review commences with an in-depth analysis of the unique characteristics and challenges inherent to agricultural environments, such as their high degree of dynamic variability including encompassing fluctuations in lighting, weather, crop growth stages, and soil conditions; a frequent lack of stable navigation references like permanent lane markings, and the stringent precision requirements mandated by agronomic tasks to maximize yield and minimize resource waste. To address these issues, this paper provides a detailed introduction to the principles of E2E autonomous driving technology, including its intelligent strategies and development paths, as well as an E2E autonomous driving framework that integrates multi task learning. And deeply dissects how multimodal perception fusion schemes and deep learning network architectures can be specifically designed for agricultural contexts. The end-to-end method constructs a unified neural network model to directly map sensor data to control commands, eliminating information loss between modules and demonstrating excellent environmental adaptability. By integrating low-cost sensors like cameras and Inertial Measurement Units (IMUs), these systems can achieve a rich, redundant environmental perception, while advanced network architectures, such as Convolutional and Recurrent Neural Networks, are optimized to process this data, thereby satisfying the critical dual demands for both low cost and high precision in field operations. Subsequently, the analysis extends to a series of typical agricultural scenarios, including open fields, orchards and forestlands, special crop environments, and multi-machine collaboration, detailing the concrete implementation pathways and distinct technical advantages of the E2E approach. For instance, in open fields, E2E systems demonstrate superior capabilities in complex behaviors like headland turning and non-linear path following. For orchards and forestlands, which present complex 3D navigation challenges, the paper explores how Deep Reinforcement Learning-based frameworks enable robust 3D spatial path planning and maneuverability in GPS-denied areas. In specialized high-value crop environments, E2E models facilitate centimeter-level precision and complex environment adaptation for delicate tasks. For multi-machine collaboration, Multi-Agent Reinforcement Learning architectures are analyzed as a means to achieve effective swarm intelligence coordination in communication-constrained settings. Recognizing the practical hurdles to implementation, this paper also identifies key challenges related to agricultural environment data scarcity, computational resource limitations, complex system integration requirements, and cost-sensitivity issues in agricultural scenarios. At the same time, key solutions have been summarized based on existing research results and technological development trends, such as simulation-to-real transfer, domain adaptation techniques, and model optimization for edge deployment. Finally, this paper presents a forward-looking perspective on the deep integration trend of E2E technology with emerging technologies like multimodal large models and agronomy knowledge graphs. This convergence promises to imbue intelligent machines with unprecedented reasoning capabilities, allowing them to understand complex commands and agronomic contexts. In doing so, this review aims to provide a definitive technical roadmap and theoretical reference for the research and development of the next generation of cost-effective, highly robust intelligent agricultural machinery, ultimately serving as a key driver for the advancement of high-quality smart agriculture.