精密装配是高端装备制造过程中的关键环节，目前主要依赖人工操作，存在装配精度不足、批量装配质量稳定性差等问题，迫切需要采用自动化技术手段提高装配质量和效率。由于批量精密装配对象存在制造误差，同时装配对象复杂的几何特征导致装配过程力学特性呈现高度非线性，仅采用刚性位置伺服控制的工业机器人难以实现精密自动装配，因此，迫切需要研究机器人精密装配柔顺控制技术。本论文面向工业生产中典型的轴孔精密装配应用需求，围绕精密装配过程建模与机器人柔顺装配控制及其重构方法开展深入的理论与实验研究。本文首先建立轴孔精密装配过程模型，总结精密装配过程非线性与多自由度耦合的特点。进而面向不同配合性质提出边界分割方法，能够基于力学信号准确识别位姿误差方向，有利于提升控制精度与鲁棒性。进一步分析装配对象几何特征对力学信号分割边界和幅值的影响规律。基于精密装配过程模型，提出一种多自由度解耦自适应柔顺控制方法，实现截面正交对称装配对象的装配控制。其中，多自由解耦控制策略结合边界分割与位姿误差影响因素辨识方法，能够有效提升装配位姿精度以及对制造误差的鲁棒性；柔顺控制参数自适应策略能够同时提升控制过程的快速性与稳定性。结合强化学习方法对基于模型的柔顺控制方法进行优化，形成一种模型驱动的强化学习柔顺控制方法，以考虑复杂对象装配中未建模因素对控制性能的影响。为提高强化学习样本效率，提出动作空间维度拓展与局部连接策略两种方法。前者通过变维映射实现高维问题在低维空间的训练和低维网络向高维空间的拓展；后者通过分析状态与动作相关性避免无关状态分量对子策略的影响。为解决差异化几何特征对象的装配控制问题，提出一种重构已有柔顺控制器的方法，避免重新训练新对象控制策略导致的效率损失。该方法利用等效控制思想，建立几何特征参数与柔顺控制参数的解析关系，用于重构柔顺控制参数；提出一种维度相似性权重策略蒸馏方法，通过更细粒度的子策略级相似性分析，加速强化学习智能体的迁移过程。最后，本文以智能手机光学镜头组件的精密自动装配为应用背景，设计并搭建机器人精密自动装配实验系统，开展实验研究与验证。实验结果证明了本文所提出的模型和控制策略及算法的有效性。
Precision assembly is the key step in the manufacturing process of sophisticated equipment, which are mainly carried out manually, causing large amounts of problems such as insufficient accuracy and poor stability in mass assembly tasks. Hence, it is urgent to use automation technology to improve assembly quality and efficiency. Due to manufacturing errors and highly nonlinear mechanical characteristics caused by complex geometric features in assembly process, precision assembly tasks cannot be completed by traditional industrial robots controlled with position-servo. It is necessary to research on the robotic adaptive compliance control method. Aiming at sovling peg-in-hole precision assembly in typical industrial production, this paper carries out in-depth theoretical and experimental research on the modeling of precision assembly process, robotic adaptive compliance control and reconfiguration method. This paper first establishes the kinematic and mechanical models of peg-in-hole precision assembly process, which present nonlinearity and coupling. Boundary segmentation theory is proposed for assembly objects with symmetrical sections, which can accurately estimate the direction of pose errors based on mechanical signals, laying a foundation for improvements of control precision and robustness. Moreover, this paper reveals the influence of geometric features on the segmentation boundaries and amplitude of mechanical signals through theoretical analysis. Based on the referred models, this paper establishes the multi-degree-of-freedom decoupling adaptive compliance control method to accomplish assembly tasks of objects with symmetrical sections. In the control method, the decoupling control strategy is based on the boundary segmentation theory and identification of most influential factors on pose errors, which can effectively improve the assembly pose accuracy and ensure robustness in face of assembly objects with manufacturing errors. The adaptive strategy of compliance parameters is proposed to improve rapidity and stability of the assembly control method at the same time.The model-based compliance control method is optimized by reinforcement learning, forming a model-driven reinforcement learning compliance control method, which reveals the influence of the non-model factors caused by non-orthogonal symmetric objects. To improve the sample efficiency of reinforcement learning, action space dimension extension and local connection policy methods are proposed. The former realizes training of high dimensional problems in low dimensional space and the extension of low dimensional networks to high dimensional space through variable dimensional mapping. The latter avoids the influence of irrelevant state components on sub policies by analyzing the correlation between states and actions.To accomplish assembly tasks of objects with different geometric features, this paper proposes a reconfiguration method of existing compliance controllers to avoid the efficiency loss caused by the retraining process on the new objects. In the reconfiguration method, the equivalent control method is raised to establish the analytical relationship between geometric feature parameters and compliance control parameters, which is used to reconfigure compliance control parameters. The dimensional similarity weighted policy distillation method is given to accelerate the transfer process of reinforcement learning agents by using more granular sub policy level similarity analysis.Finally, this paper designs experiments based on precision assembly tasks of optical lens components in smartphones, which verify the effectiveness of the compliance control and reconfiguration methods proposed above.
