绳驱机械臂是一个驱动箱和机械臂本体分离的机器人系统，它具备优秀的灵活性，适合在太空设备维修等各种狭小复杂环境下作业。在机械臂的运动控制规划中，通常需要对机械臂的可操作性进行优化，避免机械臂陷入奇异状态。不同于传统的工业机械臂，绳驱机械臂有其自身的特点，因此在可操作性优化中，有两个问题亟待研究：一是根据绳驱机械臂结构特点，定义出合理的可操作性描述指标；二是绳驱机械臂运动学建模复杂，需要设计一个实时性好通用性强的可操作性优化算法，来控制其在运动过程中自动完成可操作性优化。具体研究内容如下：首先，对绳驱机械臂进行了结构分析和运动学建模，抓住绳驱机械臂的驱动源为绳索这一特点，从关节-末端映射和关节-绳索映射的运动学模型出发，以最小化绳索速度为优化目标，提出了基于最小绳速法的关节-绳索映射的运动学模型。仿真实验表明，相比于经典的关节伪逆法，最小绳速法的规划方法能够使绳索长度变化量减少了46.16%。基于最小绳速法的映射关系，首次提出了绳索速度可操作性椭球和绳驱代价度两个概念，并通过实验验证了绳驱代价度能够正确衡量绳驱机械臂的可操作性，绳驱代价度越小，机械臂可操作性越高。然后，针对绳驱机械臂可操作性优化问题，提出使用无模型强化学习算法来控制机械臂达到实时优化的效果。多组消融实验确定最优物理量的设定和超参数的取值，并在训练中使用了广义值函数估计器、U形分布采样、特征归一化三个技巧来帮助策略网络收敛，使得算法能够控制机械臂以 5mm 的末端定位精度完成轨迹跟踪任务。接着设计了奖励融合法和多评价网络法两种算法实现轨迹跟踪过程中的可操作性优化，通过仿真和实物实验表明，多评价网络法的优化效果最佳，在不损失末端定位精度的前提下显著减少绳驱代价度，相比于非优化算法，在直线轨迹跟踪中减少绳驱代价度75.00%，而在圆轨迹跟踪中减少绳驱代价度59.93%。最后，针对两个绳驱机械臂协同作业的应用需求，在单臂可操作性优化的基础上，开展双臂协同可操作性优化算法的设计。根据双臂系统的特点，通过对两个过渡绳索速度可操作性椭球取并集的方式，首次提出了双臂绳驱代价度来描述双臂系统整体的可操作性。此外还设计了多评价网络多智能体TD3算法来控制双臂系统，使得双臂系统在完成末端目标接近任务时，双臂绳驱代价度指标相比于非优化算法而言下降了49.14%。

The cable-driven manipulator is a robot system that separates the driving box and the manipulator. It has excellent operability and flexibility, so it is suitable for working in various narrow and complex environments. It is widely used in space equipment maintenance, underwater exploration, nuclear waste treatment, and other fields. In the motion control planning of the manipulator, it is usually necessary to optimize the manipulability of the manipulator to avoid the singular state of the manipulator, which leads to a series of problems such as the loss of joint freedom, the overload of driving velocity and the violent oscillation of joints. Different from the traditional industrial manipulator, the cable-driven manipulator has its own characteristics. Therefore, in the manipulability optimization, there are two problems to be studied First, according to the structural characteristics of the cable-driven manipulator, a reasonable manipulability description index is defined. Second, the kinematics modeling of the cable-driven manipulator is complex, and the traditional optimization algorithm has high computational cost and poor real-time performance. Therefore, it is necessary to design an operable optimization algorithm with good real-time performance and strong versatility to control the cable-driven manipulator to automatically complete the manipulability optimization during the movement. The specific research contents are as follows :Firstly, the structural analysis and kinematic modeling of the cable-driven manipulator are carried out. The driving source of the cable-driven manipulator is cable. Based on the kinematic model of joint-end mapping and joint-cable mapping, the kinematic model of joint-cable mapping based on the minimum cable length method is proposed to minimize the cable velocity. Simulation experiments show that compared with the classical joint pseudo-inverse method, the planning method of the minimum cable length method can reduce the change of the cable length by 46.16% , while the change of the joint angle does not differ greatly. Combining the mapping relationship of the minimum cable length method with the definition of traditional manipulability, two concepts of cable velocity manipulability ellipsoid and cable drive cost index are proposed for the first time. It is verified by experiments that the cable drive cost index can correctly measure the manipulability of the cable drive manipulator. The smaller the cable drive cost index, the higher the manipulability of the manipulator, which provides a theoretical basis for the following control optimization algorithm. Then, for the manipulability optimization of the cable-driven manipulator, a model-free reinforcement learning algorithm is proposed to control the manipulator to achieve real-time optimization. After analyzing the basic physical quantities of the reinforcement learning algorithm in the field of cable-driven manipulator control, the setting of the optimal physical quantity and the value of the hyper-parameters are determined by the ablation experiment. In the training, three techniques of generalized value function estimator, U-shaped distribution sampling, and feature normalization are used to help the strategy network converge, so that the reinforcement learning algorithm can control the manipulator to complete the trajectory tracking task with the end positioning accuracy of 5mm. Then, two algorithms, the reward fusion method and the multi-critic network method, are designed to realize the manipulability optimization in the trajectory tracking process. The simulation and physical experiments show that the multi-critic network method has the best optimization effect, and the cable drive cost index is significantly reduced without losing the end positioning accuracy. Compared with the non-optimization algorithm, the cable drive cost index is reduced by 75.00 % in the linear trajectory tracking and 59.93 % in the circular trajectory tracking.Finally, aiming at the application requirements of the cooperative operation of two cable-driven manipulators, the design of dual-arm cooperative manipulability optimization algorithm is carried out, based on the manipulability optimization of single arm. According to the characteristics of the dual-arm system, the cost of the dual-arm cable drive is proposed for the first time to describe the overall manipulability of the dual-arm system by taking and combining the two transition cable velocity manipulability ellipsoids. Then, a multi-critic network multi-agent TD3 algorithm is designed to control the dual-arm system, so that the dual-arm system can reduce the cost of the dual-arm cable drive and optimize the manipulability of the whole dual-arm system in the process of completing the end target approaching task. The simulation results show that the cost index of the dual-arm cable drive of the dual-arm cooperative manipulability optimization algorithm is 49.14 % lower than that of the non-optimization algorithm.