Cinemática inversa para un robot planar 2R basada en una red neuronal con datos sintéticos a través de un enfoque de ensamble de modelos

This article presents the development of a neural network-based model, trained with synthetic data, to replace the geometric inverse kinematics of a 2R planar robot. This approach aims to simplify the implementation of kinematics, reducing development time and computational resource usage. The model was created in Google Colab (Python) using TensorFlow/Keras, which facilitated its creation and training. Furthermore, the system integrates real-time image processing to recognize and follow contours, which the robot subsequently traces. A linear and vertical motion was implemented using a rack-and-pinion mechanism, enabling discontinuous tracing between contours of an image. The results, averaging three neural network models, show high accuracy in predicting the angles of the robot's first two joints, with an RMSE of 0.2293 and 0.0739 compared to the geometric inverse kinematics.