仿人机器人实时动作模仿的最优控制策略  

Optimal control strategy for real-time motion imitation of humanoid robots

作　　者：韩可 李世其[1] 周雨玫 熊友军 HAN Ke;LI Shiqi;ZHOU Yumei;XIONG Youjun(School of Mechanical Science and Engineering,Huazhong University of Science and Technology,Wuhan 430074,China;UBTECH Robotics Co.Ltd.,Shenzhen 518000,Guangdong China)

机构地区：[1]华中科技大学机械科学与工程学院,湖北武汉430074 [2]深圳市优必选科技股份有限公司,广东深圳518000

出　　处：《华中科技大学学报（自然科学版）》2023年第11期1-8,共8页Journal of Huazhong University of Science and Technology(Natural Science Edition)

基　　金：广东省重点领域研发计划资助项目(2019B010154003).

摘　　要：为实现仿人机器人对人类上肢动作的高相似模仿,提出了基于递归神经网络(RNN)的拟人机械臂实时控制算法.首先,将动作模仿转化为对腕部和肘部双演示轨迹的同步跟踪问题.为了协调双轨迹跟踪过程中的潜在冲突及关节极限,采用最优控制技术将该问题构造为一个时变凸二次规划(QP)问题.机械臂的关节位置极限和速度极限以不等式约束形式加入到优化问题中.然后,设计了递归神经网络SWET-RNN,对该QP问题进行实时化求解.理论证明了SWET-RNN的全局收敛性和稳定性,其能够在有限时间内收敛到所述QP问题的最优解.最后,真机实验表明:部署了SWET-RNN算法的动作模仿控制器对腕部目标轨迹的平均跟踪误差为6.5 mm,肘部平均跟踪误差为7.3 mm,单次调用平均耗时仅为0.126 ms;仿人机器人能够在不超越自身物理能力限制的前提下,对演示动作进行高精度、低延迟的模仿.To achieve high-fidelity imitation of human upper limb motion by humanoid robots,a real-time control algorithm for anthropomorphic manipulators based on recurrent neural networks(RNN)was proposed.First,the motion imitation was transformed into a simultaneous tracking problem for the wrist and elbow demonstration trajectories.To coordinate the potential conflicts and joint limits during dual trajectory tracking,the problem was constructed as a time-varying convex quadratic programming(QP)problem using optimal control techniques.The limits of joint positions and velocities of the manipulator arm could be added to the optimization problem in the form of inequality constraints.Then,the SWET-RNN was designed to solve this QP problem in real time.The global convergence and stability of the model was theoretically demonstrated,and it was able to converge to the optimal solution of the stated QP problem in finite time.Finally,real-world experiments show that the motion imitation controller with the SWET-RNN algorithm has an average tracking error of 6.5 mm for the wrist and 7.3 mm for the elbow of the target trajectory,and the average time consumption of a single call is only 0.126 ms.The humanoid robot is able to imitate the demonstrated motion with high accuracy and low latency without exceeding the limits of its physical capabilities.

关 键 词：仿人机器人 人类动作模仿 最优控制 时变凸二次规划 递归神经网络(RNN) 

分 类 号：TP242.6[自动化与计算机技术—检测技术与自动化装置]