Error compensation of collaborative robot dynamics based on deep recurrent neural network

Establishing the dynamics model of robot and its parameters is significant for simulation analysis, control algorithm verification, and implementation of human–machine interaction. Especially under various working conditions, the errors of the calculated predicted torque of each axis have the most direct negative effect. The general robot dynamics model rarely takes the minor and complex characteristics into consideration, such as the reducer flexibility, inertia force of motor rotors, and friction. However, as the structure of collaborative robots is lighter and smaller than the ordinary industrial robots, the characteristics neglected by general dynamics models account for a relatively large amount. The above facts result in a large error in the calculation and prediction of collaborative robots analysis. To address the short comings of general robot dynamics model, a network based on long short-term memory (LSTM) in deep recurrent neural network was proposed. The network compensates the general dynamics model of a self-developed six-degree-of-freedom collaborative robot based on the consideration of gravity, Coriolis force, inertial force, and friction force. In the experiment, the nondisassembly experimental measurement combined with least-squares method was used to identify the parameters. The motor current was used to evaluate the joint torque instead of mounting an expensive and inconvenient torque sensor. The excitation trajectory based on the Fourier series was optimized. The raw experimental data were used to train the proposed LSTM network. About the accuracy of the dynamic model and the compensation method for the collaborative robot, the root-mean-square error of the calculated torque relative to the actual measured torque was used to train the network and evaluate the proposed method. The analysis and the results of the experiment show that the compensated collaborative robot dynamics model based on LSTM network displays a good prediction on the actual torque, and the root-mean-square error between predicted and actual torques is reduced from 61.8% to 78.9% compared to the traditional model, the effectiveness of the proposed error compensation policy is verified.