Kinematics Prediction of a Robotic Arm Using Traditional Machine Learning and CNN-LSTM Techniques

Abstract

Learning inverse kinematics in 3 degrees of freedom (3-DOF) robots usually requires manual intervention, which limits their autonomy. This research proposes to optimize this process using Machine Learning techniques, evaluating the performance of models trained with tabular (organized in tables) and visual (videos and images) data. A robotic arm that could trace a given letter in space was simulated. Cartesian coordinates and angular positions of the robot's joints were extracted to structure a dataset organized in tables (tabular data) and the videos corresponding to each writing task. With these datasets, the performance was compared on the one hand using tabular data with traditional Machine Learning algorithms such as Support Vector Machines (SVMs), Artificial Neural Networks (ANNs), Decision Tree (DTs), and Random Forest (RFs); on the other hand, with the videos with contemporary Machine Learning algorithms with convolutional neural networks of short and long temporal memory CNN-LSTM. The results show good predictability for most of the algorithms under study, with coefficients of determination that are very close to one. However, the CNN-LSTM trained with videos has less error than traditional Machine Learning algorithms and performs automatic feature extraction.