Font Size: 

The joint stiffness of the robot arms plays an important role in posture and movement control. This article proposes a real-time and quantitative method to inverse the stiffness of the robot arms. This method uses the TubeNet that was proposed by Liu recently through special architecture design of nets. Firstly, establish a finite element simulation model based on the actual physical model. According to the comparison of experimental modal analysis and computational modal analysis, verify and modify the simulation model. Next, the simulation model was used to verify the sensitivity between the stiffness of each joint and the natural frequencies of each order, and the first few natural frequencies with the highest sensitivity were selected for inverse problem analysis. Finally, the joint stiffness is sampled, and then the simulation model is used to establish a joint stiffness and natural frequency data set. Establish a forward TubeNet neural network. By training the network of forward problems, the relationship between joint stiffness and natural frequency is obtained. The direct-weights-inversion (DWI) formulae based on TubeBet can compute the joint stiffnesses at the joints explicitly in real-time using the parameters trained on the positive problem. Moreover, the article discusses the activation function suitable for inverse problem analysis, and studies the influence of the number of neurons and the number of hidden layers on the accuracy of inverse problem solving. Using DWI formula can make the maximum error of the joint stiffness solved within 0.72%.

Robot; Joint stiffness; TubeNet neural network; Inverse problem