Identification of Robot Forward Dynamics via Neural Network

In recent years machine learning techniques have received an increasing interest, since they can be successfully applied to several application domains, among which robotics. In this field, neural networks can be used to approximate the dynamic model of the manipulator, which is highly non-linear and often affected by some uncertainties regarding the dynamic parameters. These parameters are typically not provided by the manufacturer or partially unknown. In this work, we propose the use of neural networks to perform the black-box identification of the robot forward dynamics. This consists in performing a one step ahead prediction of the robot joint positions, velocities and accelerations based on the knowledge of the joint positions, velocities at the previous time step and the actuation torques. To this purpose, we analyzed both static and dynamic neural networks structures with different combinations of hyper-parameters. We tested the performance by comparing the predicted output with both simulated data and those acquired on a real industrial manipulator, ABB IRB 140.