毕业生知识库

我国产业转型升级和劳动力成本上升的双重影响下，由机器人引发劳动力
置换将成为提升我国制造业竞争力的重要举措，为我国经济持续增长提供内生
动力。因此，如何提升机器人面对复杂问题的智能性是亟待解决的重要问题。目
前，数据驱动与人工智能展现出巨大潜力，其离不开大数据的支撑。然而在机器
人领域，大数据的获取绝非易事，这要求我们必须将目光转向虚拟空间。
近年来，随着工业 4.0、数字孪生、元宇宙等概念的兴起，在人们对利用虚
拟空间展现出巨大热情的同时，虚拟空间也不断释放其在工业场景中的应用潜
力。在未来工厂里，传统控制方法势必由实时数据驱动的虚实互动方法所取代。
然而这类方法尚缺乏理论支撑与方法探索。
本论文基于平行系统和 ACP 方法 (Artificial Societies, Computational Experiments, and Parallel Execution)，提出平行机器人框架，研究平行机器人的实现方
法，研发平行书法机器人系统，并围绕平行机器人框架研究上层、下层任务的数
据驱动方法，分别解决基于机器人书法笔触规划问题和永磁同步电机最优跟踪
控制问题。本文的主要工作如下：
1. 针对机器人书法创作的灵活性与精确性不足的问题，本文结合基于 ACP
的平行机器人框架设计并实现了平行书法机器人系统。系统由执行系统、软件定
义的系统和采集系统组成。其中，执行系统由机械臂和辅助设备组成，用以执行
执行笔触书写任务；软件定义的系统由机器人中间件和仿真算法组成，用以调度
系统资源并实现对书写过程和书法写作结果的仿真模拟；采集系统由动作捕捉
系统、相机和辅助装置组成，用于采集人类书法创作过程并自动处理采集数据。
平行书法机器人系统的运行逻辑围绕描述、预测、模仿、引导 4 个过程展开，以
数据驱动方式实现机器人书法学习和创作的闭环。实验表明平行书法机器人系
统可以有效采集、仿真并执行书法创作过程。
2. 针对机器人书法笔触规划方法中存在交互耗费大、收敛困难、精度受限
的问题，本文结合平行书法机器人系统，提出以数据驱动的方式改进机器人笔触
规划的学习过程。本文通过描述过程优化仿真参数，从而以计算实验代替实际交
互实验进行强化学习，减轻系统的交互耗费；通过采集和处理人类书写过程，使之作为专家监督数据以模仿学习的方式学习专家规划策略，提升规划效果和学
习速度；同时基于平行系统附加的可观性扩展原笔触规划问题，从而实现更精细
灵活的书法笔触规划。实验表明，本文方法可以有效优化仿真参数和采集专家数
据，并基于少量交互数据有效提升笔触规划的学习速率和规划效果。
3. 针对连续时间系统的状态跟踪问题，本文提出了无模型近似最优跟踪
(MfNOT) 方法。本方法基于平行控制，以控制量作为增广状态对原系统和原损
失函数进行增广，然后从理论上证明了增广系统与原系统的稳定性和最优性具
有一致性，并由此得到零稳态误差跟踪控制律。通过积分强化学习，本方法从在
线执行数据中学习并调节网络参数以优化性能指标。在永磁同步电机上的实验
表明， MfNOT 方法可以从电机的在线执行数据中学习，优化后的控制器在超调
量、调整时间等动态性能指标超过手工调参 PI 控制器。此外，本文所提出方法
使控制量平滑变化，提升了电机起动阶段的能量效率。
本文提出的平行机器人系统是一个融合多领域、跨学科的综合系统，包含了
机器人软硬件、传感、控制、机器学习等内容，以数据驱动为核心，使机器人获
得持续的再优化能力。本文初步探索了平行机器人系统的实现方式，及其框架内
的数据驱动规划与控制方法。
 

Under the dual influence of China's industrial transformation and upgrading and rising labor costs, labor replacement induced by robots will become an important measure to improve the competitiveness of China's manufacturing industry and provide endogenous power for China's sustained economic growth. Therefore, how to improve the intelligence of robots in the face of complex problems is an important problem to be solved urgently. At present, data-driven technologies and artificial intelligence have shown great potential, which cannot be separated from the support of big data. However, in the field of robotics, the acquisition of big data is not easy, which requires us to turn our attention to virtual space.

In recent years, with the rise of the concepts of Industry 4.0, Digital Twins, meta-universe and so on, while people show great enthusiasm for using virtual space, virtual space is also constantly releasing its application potential in the industrial scene. In the factory of the future, traditional control methods are bound to be replaced by real time data driven virtual interaction. However, these methods lack theoretical support and methodological exploration.

Based on Parallel systems and ACP (Artificial Societies, Computational Experiments, and Parallel Execution), this paper proposes the framework of Parallel robotics and studies its implementation methods. The parallel calligraphy robot system was developed, and the data-driven methods of the upper and lower levels of tasks were studied around the framework of the parallel robot. The problems of robot calligraphy stroke planning and the optimal tracking control of permanent magnet synchronous motor were solved respectively. The main work of this paper is as follows: 

1. Aiming at the problem of insufficient flexibility and accuracy of robot calligraphy, this thesis designs and implements a parallel calligraphy robot system based on the parallel robotics framework based on ACP. The system consists of an execution system, a software-defined system, and an acquisition system. Among them, the execution system consists of a robotic arm and auxiliary equipment to perform the task of writing brushstrokes; the software-defined system consists of robot middleware and simulation algorithms to schedule system resources and simulate the writing process and calligraphy writing results; the acquisition system is composed of a motion capture system, cameras, and auxiliary devices, which are used to collect the process of human calligraphy creation and automatically process the collected data. The operation logic of the parallel calligraphy robot system revolves around the four processes of description, prediction, imitation, and guidance, which realize the closed-loop of robot calligraphy learning and creation in a data-driven manner. Experiments show that the parallel calligraphy robot system can effectively collect, simulate, and execute the process of calligraphy creation.
2. Aiming at the problems of high interaction cost, difficult convergence, and limited accuracy in the robot calligraphy stroke planning method, this thesis combines the parallel calligraphy robot system and proposes a data-driven way to improve the learning process of robot brush stroke planning. This thesis optimizes the simulation parameters by the describing process, to replace the actual interactive experiment with computational experiments for reinforcement learning and reduce the interactive cost of the system; collecting and processing the human writing process is used as expert supervision data to learn expert planning strategies in the way of imitation learning. In this way, our method improves the planning effect and learning speed. Furthermore, based on the additional observability of the parallel system, the original brushstroke planning problem is extended, so as to achieve more refined and flexible calligraphy stroke planning. Experiments show that our method can effectively optimize the simulation parameters and collect expert data, and effectively improve the learning rate and planning effect of brushstrokes based on a small amount of interactive data.
3. In this thesis, a model-free approximate optimal tracking (MfNOT) method is proposed for the state tracking problem of continuous-time systems. Based on parallel control, this method augments the original system and the original loss function with the control amount as the augmented state, and then theoretically proves that the stability and optimality of the augmented system and the original system are consistent, and thus the zero-steady state error tracking control law is obtained. Through integral reinforcement learning, MfNOT learns from online execution data and tunes network parameters to optimize performance metrics. Experiments on permanent magnet synchronous motors show that the MfNOT method can be learned from the online execution data of the motor, and the optimized controller surpasses the manually adjusted PI controller in dynamic performance indicators such as overshoot and adjustment time. In addition, the control variables of the method proposed in this thesis change smoothly, which improves the energy efficiency of the motor starting stage.

The parallel robotics system proposed in this paper is a multi-domain and interdisciplinary integrated system, including robot hardware and software, sensing, control, machine learning and other contents, with data driven as the core, so that the robot can obtain continuous re-optimization ability. In this paper, the implementation of parallel robot system and the data-driven planning and control method within the framework are preliminarily explored.