Deep-learning damped least squares method for inverse kinematics of redundant robots

• Published in: Measurement : journal of the International Measurement Confederation 2021-02, Vol.171, p.108821, Article 108821
• Main Authors: Wang, Xiaoqi, Liu, Xing, Chen, Lerui, Hu, Heyu
• Format: Article
• Language: English
• Subjects: Algorithms; Artificial neural networks; Configurations; Convergence; Damped least square method; Data points; Deep learning; Deep neural network; Inverse kinematics; Iterative methods; Kinematics; Least squares method; Neural networks; Optimal enhanced coefficient; Optimization; Redundancy; Redundant robot; Robots; Simulation; Stability
• ISSN: 0263-2241; 1873-412X
• DOI: 10.1016/j.measurement.2020.108821

•Optimal enhanced coefficient is proposed to optimize damped least squares method.•The deep neural network we designed effectively predicts the coefficient.•The proposed method is applied to solve inverse kinematics of redundant robots.•The method is better than existing methods in solving speed and convergence. In the robot field, it has always been a hard issue of solving inverse kinematics (IK) problems of redundant robot. Although many researchers have come up with solutions for redundant robots with different configurations, there is still an issue of computational efficiency for redundant robot with complex configuration. In this paper, we proposed a novel optimization method which solves redundant robot IK with ultra-high speed and accuracy. On the basis of damped least square (DLS) method with a good stability, we introduced an optimal enhanced coefficient for the first time to achieve faster iteration and more accurate convergence. This is also due to the good performance of the deep neural network we designed in coefficient prediction. 106 data points were selected from the robot working space as the original data training network. The simulation result showed that, unlike other algorithms which take hundreds of iterations to achieve convergence, through the optimized method proposed in this paper, the average number of iteration was just less than 10 and the solving speed was greatly improved on the basis of ensuring stability. In addition, the convergence of the algorithm was greatly improved and when the error threshold was 0.01 mm, the convergence reached 95.47%, which is better than most existing algorithms. In this paper, we evaluated the common IK methods, and the results showed that this method performed better in convergence, accuracy and speed.