Robot joint type recognition using machine learning

Reconfigurable robots and mountable measurement systems face attraction due to significant changes in development of wireless networks and internet of things. This research benefits cross disciplinary viewpoints to build a self-aware robotic module that utilizes data analysis and machine learning for a network of sensors mounted on an arbitrarily chosen serial manipulator. Serial robotic arms consist of series of links and joints. Joints connect links to each other with possibility of generating rotational or translational motion. The goal of the study is to detect the mechanical joint type of robotic manipulators which can be rigid (body), revolute joints, prismatic joints, corrupted and random data via machine learning techniques. The dataset of robotics manipulators is collected via Aaria platform with diverse structures. Measurements and classifications are based on raw data collected from real and simulated sensors. These sensors are capable of providing absolute values of acceleration and angular velocity measured by Inertial Measurement Units (IMUs). Raw input data is transformed into useful features by feature extraction. The machine learning models which are explored in this study consists of logistic regression, 5-nearest neighbors, linear discriminant analysis, random forest, XGBoost and LSTM models. The most accurate methods are XGBoost and LSTM with accuracy of approximately 69% and 70% respectively