Neural Network Modeling for the Meca500 Industrial Robot: Practical Implementation, Validation, and Testing

This paper presents a movement model for Meca500, an ultra-compact six-axis industrial robot, using a multilayer perceptron neural network (MLP). Accurate robot calibration is essential for advanced control techniques. However, the minimal difference between input and output velocity vectors, measured in millimeters, makes precise identification challenging. To address this, input-output data from various practical experiments were collected and used to generate multiple neural models, varying parameters such as training optimizers, neuron count, activation functions, and training algorithms. The best-performing model was selected through data analysis and validated. Results indicate that the optimized neural network accurately predicts the robot's output behavior with minimal mean absolute error (MAE): 0.15 mm at low velocity, 0.52 mm at intermediate velocity, and 1.5 mm at high velocity in both Y and Z directions. This study confirms the feasibility of using neural networks for precise robot system modeling, enhancing simulation accuracy before real-world deployment.