Approach to Inverse Dynamics Model for a Serial 3 DoF Robot Using Neural Networks

The dynamic model of a robot is complex as the system is generally nonlinear, multivariable, and time-invariant. Methods have been developed to model the dynamics of robots such as the Newton-Euler and Lagrange-Euler algorithms. For those methods, it is important to have accurate data such as localization of the link’s center of mass about its base, the inertia matrix, lengths, and viscosity. Uncertainties of the data make the model different from the real system. For those reasons, we applied a Neural Network model to identify the inverse dynamic model of a 3-degree-of-freedom robotic with joint torques and joint values. The robot model was developed using Matlab Simscape Multibody in Simulink Toolbox, and sample input torque and output joint time evolution data were obtained to train the model. The data obtained was divided into training data and validation data. The model was validated using the mean square error (MSE), $\mathrm{R}^{2}$, and Cosine Similarity metrics comparing the measures with those of the ground truth.