Feedforward Control for a Manipulator with Flexure Joints Using a Lagrangian Neural Network

Feedforward control of a manipulator can be generated with a sufficiently accurate stable inverse model of the manipulator. A Feedforward Neural Network (FNN) can be trained with experimental data to generate feedforward control without knowledge about the system at hand. However, the FNN output can show unphysical behaviour especially in operational regimes where the training data is sparse. Instead, the output of a Lagrangian Neural Network (LNN) is limited by physical constraints and hence is expected to predict the (inverse) multibody system behaviour more robustly. We propose to generate the feedforward control by first training a LNN that captures already most features in experimental data and next add a FFN to account for a relatively small residual. Experimental results from a fully actuated 2-DOF manipulator with flexure joints show that the accuracy of the controlled motion using this approach is comparable to using an identified inverse plant model built from the system’s equations of motion.