Parameter Identification of a Robot Arm Manipulator Based on a Convolutional Neural Network

Dynamic parameters are crucial in designing robotics systems because they reflect an actual robot. Conventional identification methods require that the robot execute the optimal motion; however, they spend time trying all possible trajectories in the robot. This article shows the identification of a robot arm of 2 degree-of-freedom with an algorithm based on a convolutional neural network (CNN) and their dynamic model. The proposal consists of a CNN that uses an image construct with a proposed conversion technique and the robot signals. The algorithm gets the parametric residuals from this signal image to find the parameters without trajectory optimization. An embedded system on a Field Programmable Gate Array (FPGA) has the classical controller Proportional-Derivative to execute a predefined trajectory for the identification. The identified parameters and the predefined motion rebuild the torque with the dynamic model. A proposed evaluation metric based on the discrete cosine transform evaluates the similarity of the actual and reconstructed torques. Four numeric tests verify the algorithm by torques created with the dynamic model, the predefined trajectory, and a parameter set. The similarity of numeric torques and their rebuilding overcomes 97.38%, and the experimental and rebuilt torque with the identified parameters is over 93.55%. The proposed algorithm is compared with least-squares, and the results show that the proposed method provides better identification of the experimental robot.