基于刚柔耦合臂及深度学习模型预测控制的柔性体柔顺抓取研究

The compliant grasp of objects by using manipulators is the key common basis for robots to perform tasks in many fields. This project aims at the compliant grasp of flexible objects. Based on the interdisciplinary research of information science, mechanics and control, the rigid-flexible coupled manipulator's rapid modeling, model reduction, and the grasp planning as well as real-time compliant grasp control are studied. Specifically, the rigid-flexible coupling of the manipulator and the deformation of the object are considered. Thereby, a fast modeling method performed on PC for the concerned multibody system based on symbolic operation and stereo vision is studied. Aiming at the large number of degrees of freedom, and the time-varying parameters dependency, this project focuses on the model reduction techniques of a time-varying parameters dependent model. In order to obtain the dynamic information of the target in real time, the hand-eye fusion is performed on the basis of the multi-mode information perception based on, e.g., vision and tactile. Combined with deep learning, those important contact points of grasp which meet the force-closure requirements are then generated and determined. Afterwards, the initial grasp force is quickly generated by the off-line training library, and the real-time optimization of the grasp force is subsequently completed by the force sensing array and the online machine learning. The model predictive controller receives inputs from grasp planning, and then dynamically adjusts the system model on runtime, so as to realize the compliant control of grasping of flexible objects, which is deemed as a highly nonlinear task. Finally, plenty of simulation and experimental verification are performed. This study is expected to provide theoretical basis and technical support for the compliant grasp of flexible objects by robots.

机械手臂类机器人对目标物体的柔顺抓取是众多领域里机器人完成作业任务的关键共性基础。本项目针对柔性物体的柔顺抓取，基于信息、力学、控制多学科交叉，深入研究刚柔耦合机械臂多体系统快速建模、模型降阶、抓取规划与柔顺实时抓取控制。具体地，考虑机械臂系统的刚柔耦合及柔性目标的变形，研究复杂多体系统基于符号化运算及立体视觉的计算机快速建模方法；针对全阶模型巨大的自由度数以及时变的参数依赖特性，重点研究时变-参数依赖的模型降阶技术；基于视触觉多模态信息感知等技术进行手眼融合以实时获取目标物动态信息，结合深度学习规划出满足力封闭性要求的抓取点。通过离线训练库快速生成初始抓取力，并用力传感阵列与在线学习完成抓取力的实时优化。模型预测控制器接收抓取规划信息后实时动态调整模型，实现对柔性体抓取这一高度非线性作业的柔顺控制；最后进行仿真和实验验证。本研究期望能为柔性物体的机器人柔顺抓取提供一定理论基础与技术支持。

本项目针对机器人抓取过程中迫切需要解决的机械臂抓取点和抓取力规划等关键共性基础问题展开深入研究。搭建了UR5机械臂、RG2和RobotiQ末端执行器等硬件组成的研究平台，建立了深度图像数据库，设计了基于深度图像的ALV视触觉规划算法，着重研究了机械臂对深度相机信息的感知，提高了信息利用率，实现了抓取点和抓取力规划。完成了基于模型预测控制的机械臂轨迹跟踪控制方法研究。针对机械手抓取过程中出现的模型失配问题，采用模型预测控制的滚动优化策略减少了模型失配对控制输出的影响，在有输入约束的条件下实现鲁棒控制。同时深入分析了抓取模型以及抓取过程中与环境之间的交互。研究了机械臂末端执行器抓取力跟踪方法，建立了基于灰色模型预测的导纳控制方法。实现了机械臂抓取过程力跟踪效果，避免了因接触产生过大的接触力而损坏抓取对象或者造成抓取失败等问题，能够实现精准力控柔顺抓取。最后，设计了大量的实验，完成了本项目中所设计的方法和算法的验证。通过本项目的研究，为机械臂的自主柔顺抓取柔性目标物提供了较好的理论基础和关键技术支持。