毕业生知识库

机器人智能控制技术是机器人领域的重要研究课题之一，也是我国《“十四五”机器人产业发展规划》列出的机器人共性技术之一。机器人的持续发展，以及机器人在智能制造、智慧服务、智能医疗、国防安全等领域应用范畴的不断扩大，对机器人智能控制理论和技术不断提出新的要求。例如，舱外货物搬运等在轨任务对空间机械臂提出高可靠、高精度跟踪控制的要求，切割焊接等作业任务对工业机器人提出高精度、高速度跟踪控制的要求。目前机器人智能控制理论和技术的研究仍存在系统模型呈现复杂不确定性、实现最优控制性能与闭环系统稳定性的平衡等难点与挑战。开展机器人智能优化控制理论和技术的研究，对于克服上述挑战，实现机器人的高性能跟踪控制具有重要意义，也有助于推进我国机器人核心技术攻关行动的开展。

模型预测控制是一类被广泛研究和应用的智能优化控制方法。因具有鲁棒性好，能够显式、动态地处理系统约束，可通过目标函数配置实现最优控制性能等优点，模型预测控制成为解决机器人智能优化控制问题的重要手段。尽管关于模型预测控制的研究已有丰富的理论和应用成果，但利用其实现机器人的优化控制，仍有许多关键问题亟需解决，例如，如何在线处理机器人模型的不确定性，在系统动态未知情况下满足模型预测控制方法对于系统模型的需求；如何设计模型预测控制的滚动优化策略，满足机器人对控制实时性的要求；如何保证闭环系统的稳定性等。鉴于此，本论文融合自适应神经网络，提出了针对机器人系统的模型预测控制方法，主要研究内容和贡献点总结如下：

1、面向带有输入约束的模型不确定机器人系统，提出了一种基于预测网络和动作-评价网络的模型预测控制方法。该方法建立了包含两组神经网络的模型预测控制框架，其中：第一组神经网络为预测网络，其构建了机器人系统的自适应更新预测模型，用于实时估计系统动态；第二组神经网络采用激活函数相同但权值不同的动作-评价网络，用于求解模型预测控制的优化问题，建立了神经网络权值的在线学习策略，以保持最优跟踪性能和预测系统稳定间的平衡。同时本方法利用李雅普诺夫定理和数学归纳法，证明了闭环系统的稳定性，闭环系统所有变量在所设计的控制方法作用下保持一致最终有界，并通过仿真实验验证了本方法的有效性，其能够实现模型不确定机器人的优化跟踪控制。

2、面向带有输入约束的模型不确定机器人系统，提出了一种引入学习终端成本的事件触发模型预测控制方法。该方法在自适应神经网络预测模型基础上，首先设计了包含预测跟踪误差和评价网络的模型预测控制的学习终端成本，并构造了其全局学习机制。然后，引入了事件触发机制，根据预测网络的权值和预测跟踪误差，设计了优化求解的触发条件，仅在满足触发条件时非周期地求解模型预测控制中的优化问题，从而提高求解结果的利用效率。同时本方法验证了模型预测控制的循环可行性，并利用李雅普诺夫定理证明了闭环系统所有变量满足一致最终有界。最后通过硬件实验验证了在此方法作用下，机器人在全局稳态优化和瞬态快速收敛方面都具有良好的性能，且具有对有限外界扰动的鲁棒性。

3、面向带有状态约束的模型不确定机器人系统，提出了一种基于Tube和滑动模态的自适应模型预测控制方法。首先利用高阶全驱系统方法，构建了机器人系统的标称模型，并作为预测模型。在此基础上，设计了引入滑动模态的标称模型预测控制器，放松了对于终端状态的约束，实现了标称系统对期望轨迹的精准稳定跟踪。然后，构建了基于节点自适应神经网络的辅助控制器，对机器人的非线性不确定系统模型进行动态补偿，并实现了将标称系统与实际系统间的估计偏差限制在Tube不变集内。同时本方法验证了标称模型预测控制循环可行，并利用李雅普诺夫定理证明了闭环系统所有变量满足一致最终有界。最后通过硬件实验验证了在此方法作用下，机器人能够实现高效、快速的轨迹跟踪。

The robotic intelligent control technology is one of the important research topics in the field of robots, and also one of the robotic generic technologies listed in the “14th Five-Year Plan for the Robot Industry Development” of China. With the constant development of robots and the continuous expansion of the application scope of robots in the fields of intelligent manufacturing, intelligence services, smart medical, national defense security, etc., new requirements have been put forward constantly for the robotic intelligent control theory and technology. For example, high reliability and high precision tracking control of the space manipulator is required for on-orbit tasks such as extravehicular cargo handling, as well as high precision and high speed tracking control of the industrial robot is required for cutting and welding tasks. At present, there are still some difficulties and challenges in the research of robotic intelligent control theory and technology, such as the complex and uncertain system model, and the balance between the optimal control performance and the stability of closed-loop systems. Research on the theories and technologies of robotic intelligent optimization control is of great significance to overcome the above challenges and realize high-performance tracking control of robots, which is also helpful to promote the core technology research of robots in China.

Model predictive control (MPC) is a kind of intelligent optimization control method which has been widely studied and applied. It has many advantages, such as good robustness, the ability to deal with system constraints explicitly and dynamically, and can achieve the optimal control performance of the system through the configuration of the objective function, etc. Therefore, MPC becomes an important method to solve the problem of the intelligent optimization control of robots. Although there have been abundant theoretical and applied achievements in the research on MPC, there are still many key problems that need to be solved in utilizing MPC to realize the optimal control of robots. For example, how to deal with the uncertainty of the robotic model online to meet the requirements of MPC for system model when the system dynamics are unknown, how to design the rolling optimization strategy of MPC to meet the real-time requirements of robot control, how to ensure the stability of the closed-loop system, and so on. In view of this, this thesis proposes MPC methods for robotic systems combined with adaptive neural networks. The main research contents and contributions are summarized as follows:

1. A model predictive control method based on the predictive network and actor-critic network is proposed for robotic systems with input constraints and uncertain dynamics. This method establishes a model predictive control framework consisting of two groups of neural networks. The first group of neural networks, which is the predictive network, is constructed as an adaptively updated predictive model of the robotic system, for estimating the system dynamics in real-time. The second one, which takes into account the actor-critic network with different weights and the same activation function, is developed for solving the optimization problem of MPC. Online learning strategies of neural networks' weights are established for balancing between optimal tracking performance and predictive system stability. At the same time, the stability of the closed-loop system is proved according to the Lyapunov theorem and mathematical induction, and the ultimately uniformly boundedness (UUB) of all variables is verified. Simulation studies are conducted to explain the effectiveness of the proposed method, it can realize the optimal tracking control of the robot with an uncertain model. 

2. An event-triggered model predictive control method with learning terminal cost is proposed for robotic systems with input constraints and uncertain dynamics. Firstly, on the basis of the predictive model of adaptive neural networks, a learning terminal cost for MPC is designed with predictive tracking error and the critic network, and a global learning mechanism for the terminal cost is constructed. After that, the event-triggered mechanism is introduced, and a triggering condition of the MPC solving is developed based on the predictive network's weights and the predictive tracking error. The optimization problem of MPC is solved aperiodically only when the triggering condition is satisfied, thereby improving the utilization efficiency of the solution. Meanwhile, this method verifies the recursive feasibility of MPC and proves that all variables are ultimate uniform bounded (UUB) utilizing the Lyapunov theorem. Finally, experiments are conducted to demonstrate that the robot has good performance on global steady-state optimization, transient fast convergence as well as robustness against limited external disturbances under the action of the proposed method.

3. A sliding mode-based adaptive tube model predictive control method is proposed for robotic systems with state constraints and uncertain dynamics. Firstly, utilizing the high-order fully actuated system approaches, the nominal model of the robotic system is constructed as the predictive model. Based on the nominal model, a nominal model predictive controller with the sliding mode is designed, which relaxes the terminal constraints, and realizes the accurate and stable tracking of the desired trajectory by the nominal system. Then an auxiliary controller based on the node-adaptive neural networks is constructed to dynamically compensate nonlinear uncertain dynamics of the robotic system, furthermore, the estimation deviation between the nominal system and the actual system is limited to the Tube invariant sets. At the same time, the recursive feasibility of MPC is verified, and the ultimately uniformly boundedness (UUB) of all variables is proved according to the Lyapunov theorem. Finally, experiments show that the robot can achieve fast and efficient trajectory tracking under the action of the proposed method.