农业对国计民生有着重要意义，粮食安全是社会稳定、经济增长和人民生活水平提高的重要保障。随着农业从业者减少和农业机械化、信息化、智能化水平提高，充分利用农业机器人完成作业成为了实现农业生产降本增效的重要手段。本文着眼于农场中的农业机器人的智能任务规划问题，设计人工智能算法，优化农业机器人调度。本文研究的农业机器人任务规划问题是具有多参数的多农机在不规则多地块内细分到作业行的任务规划，这是一种多目标的与节点尺度和入口点相关的车辆路径问题。本文改进和设计了有序多变异遗传算法、基于深度强化学习的单目标任务规划算法和基于约束强化学习的多目标任务规划算法，并设计了含运动学约束的路径规划和路径跟随算法，搭建了仿真环境，对任务规划算法进行了验证。本文的主要贡献和创新点如下： （1）针对细分到作业行的任务需要确定作业行入口点的问题，提出了应用于遗传算法的入口点取反算子、组内排序算子和贪婪路径算子，以及基于节点-入口点二维概率分布采样的动作网络，解决了作业行入口的选择问题，实现了农机细分到作业行的任务规划，提升了农业机器人作业精细化程度。 （2）针对作业行和农业机器人具有多参数的问题，建立了基于图神经网络和交叉注意力机制的神经网络编码器，并在解码器的已生成序列编码部分考虑了农机特征，提升了网络对作业行节点和农机特征的提取能力和对农机特征的利用效率，解决了农业机器人和作业行之间的匹配以及单个农机的任务的排序问题，实现了多地块多参数农机端到端的快速任务规划，使得算法能应用于多样的生产场景。 （3）针对农业机器人多目标任务规划问题，将难以设计的时间和油耗多目标优化问题建模为可定量描述的在时间约束条件下优化油耗的问题，设计了具有额外安全预算的约束策略优化算法，解决了农机任务规划中多个目标间的组合没有理论和经验依据的问题，实现了针对不同时间要求给出不同规划方案的目标，有利于在农业生产中在满足时间要求的同时降低油耗，提升经济效益。 在本文设计的仿真环境中，本文应用迪杰斯特拉算法和杜宾斯曲线实现了路径规划，设计了比例微分控制器实现农机控制，通过含运动学的仿真验证了智能任务规划算法的应用前景。

Agriculture holds vital significance for economies and livelihoods, with food security serving as a cornerstone for social stability, economic growth, and living standards. As the agricultural workforce decreases and agricultural mechanization, informationization, and intelligence increase, leveraging agricultural robots to perform tasks has become a vital means to reduce costs and enhance efficiency in agricultural production. The thesis addresses task planning challenges of agricultural robots in smart farms, utilizing artificial intelligence algorithms for optimizing robot planning. The agricultural robot task planning problem addressed in this thesis involves planning for multiple multi-parameter robots subdividing tasks into working lines within irregular fields. This problem represents a multi-objective vehicle routing problem related to node scale and entry points. The thesis introduces improvements and designs of the ordered multi-mutation genetic algorithm, single-objective and multi-objective task planning algorithms based on deep reinforcement learning and constrained policy optimization, respectively. Additionally, it devises path planning and path-following algorithms with kinematic constraints, establishes a simulation environment, and validates the task planning algorithms. The main contributions of the thesis can be summarized as: (1) Addressing the issue of determining entry points for working lines, the thesis proposes 3 specific genetic algorithm operators, including entry points inversion, intra-group sorting, and greedy path operators, as well as an action network based on two-dimensional probability distribution of nodes-entry pairs, achieving the subdivision of tasks into working lines, and improving the precision of agricultural robot operations; (2) Addressing the multi-parameter issues of working lines and robots, a neural network encoder based on graph neural networks and cross-attention mechanisms is established and the robots features are considered in the generated sequence encoding part of the decoder, enhancing the network's ability to extract and utilize features of working lines and robots, solving the assignment and arrangement of working lines, achieving rapid end-to-end task planning and enabling the algorithm to be applied in diverse scenarios; (3) Addressing the multi-objective task planning problem, multi-objective optimization problem is modeled as quantitatively describable fuel consumption optimization under time constraints, and optimized by a novel algorithm with extra safety budgets proposed in the thesis, realizing low fuel costs under different time constraints, enhancing economic benefits while meeting time requirements. In the simulation, the thesis uses Dijkstra's algorithm and Dubins curves for path planning and designs a proportional-derivative controller to control the agricultural robots. Through kinematic-constrained simulations, the application prospects of the intelligent task planning algorithms were verified.