基于示教学习的混合柔顺控制方法在未来聚变堆部件自动化装配中的应用研究

Efficient assembly in complex environments is one of the key issues that must be addressed in the future commercialization of fusion reactors. However, in the future, the fusion reactor exhibits a local unstructured state and an unknown maintenance state of the assembly contact parameters under the action of high temperature, magnetic field, radiation, etc., which greatly increases the assembly difficulty. The master-slave operation puts people into the control loop, and can smoothly complete the assembly trajectory planning, but the maintenance efficiency is low. Impedance control, although suitable for complex assemblies, is susceptible to end effector reference trajectories and environmental parameters. This project intends to adopt a master-slave control + impedance control hybrid method based on teaching learning technology to establish a hybrid compliance controller based on dynamic primitive equations and impedance control equations: firstly, using the master-slave control to complete the teaching assembly and record the trajectory data (position Information and force information), and based on dynamic equations and teaching trajectories to establish dynamic primitive equations. When performing automated assembly tasks, due to environmental errors, the robot can't track the pose and force/torque of the generalized trajectory at the same time. Based on the impedance control, the generalized trajectory is fine-tuned so that the contact force/torque matches the teaching force/torque to achieve automated and compliant assembly. The method can provide a theoretical basis for the fusion teleoperation smooth assembly technology, and also provide reference for the compliant technology research in other extreme structural environments.

复杂环境下的高效装配是未来聚变堆实现商用化所必须解决的关键问题之一。然而，未来聚变堆在高温、磁场、辐射等作用下表现出局部非结构化、装配接触参数未知的维护状态，大大增加装配难度。主从操作将人纳入控制环中，能平顺的完成装配轨迹规划，但维护效率低。阻抗控制虽适用于复杂装配，但易受末端执行器参考轨迹和环境参数影响。本项目拟采用基于示教学习技术的主从控制+阻抗控制混合方法，建立基于动态基元方程和阻抗控制方程的混合柔顺控制器：首先利用主从控制完成示教装配并记录轨迹数据（位置信息和力信息），并基于动力学方程和示教轨迹建立动态基元方程。在执行自动化装配任务时，由于环境误差存在导致机器人无法同时跟踪泛化轨迹的位姿和力/力矩，基于阻抗控制对泛化轨迹微调使得接触力/力矩与示教力/力矩匹配，从而实现自动化柔顺装配。该方法可为聚变遥操作柔顺装配技术提供理论依据，同时也为其它极端结构化环境下的柔顺技术研究提供参考。

复杂环境下的高效装配是未来聚变堆实现商用化所必须解决的关键问题之一。然而，未来聚变堆在高温、磁场、辐射等作用下表现出局部非结构化、装配接触参数未知的维护状态，大大增加装配难度。. 本项目以重载机械臂在聚变环境下执行轴孔装配为研究目标，重点首先开展基于示教学习技术的主从控制+阻抗控制混合方法研究。首先在示教试验过程中解决了双边主从力反馈柔顺控制，然后在DMPs运动方程研究过程中攻克了重载机械臂的刚柔耦合动力学建模问题。最终通过DMPs技术将主从示教和阻抗控制结合成功实现了实验室环境下的轴孔装配。. 另外，基于项目组在神经网络技术方面的积累，项目组成员开发了一种基于N-step Dueling-DQN的深度强化学习方法，该方法通过神经网络融合来自RGBD相机和力传感器数据，再结合混合力/位控制策略，同样在实验室环境下使机械臂能够以类似于人类的手眼协作的方式在不确定和不稳定的环境下完成轴孔装配任务，装配精度达到0.1mm。