For the feedforward control of a manipulator, the required driving forces can be computed with an accurate stable inverse system model. Training of a Lagrangian Neural Network (LNN), or Deep Lagrangian Networks (DeLaN), has been proposed to estimate the conservative part of the driving forces for a specified trajectory. Such network is bound to physical constraints and hence can predict the (inverse) multibody system behaviour quite accurately and robustly from a relatively small dataset. In a previous paper we extended this framework into a so-called DeLaN+D by adding additional damping and friction terms in the underlying equations. The performance of this network has been evaluated with simulated and experimental data from a fully actuated 2-DOF manipulator with flexure joints. In the present paper we consider a more complex system of a redundantly actuated 3-DOF manipulator. This system has four brushless direct current (BLDC) motors that show considerable cogging behaviour. In a first principles model the system’s multibody dynamics is expressed in three independent coordinates, e.g. both translations and the rotation of the end-effector (EE). The cogging contributions depend heavily on accurate knowledge of all rotation angles of the four motors. The kinematic relations between the EE coordinates and motor angles are known in a model. However, experimental data is most likely harder to interpret due to model inaccuracies. For this reason two different DeLaN+D implementations are considered. A 3-DOF network estimates the effective EE forces and torque from the EE coordinates. Alternatively, a 4-DOF network estimates motor torques directly from the motor angles. Using simulated data, the accuracy of the feedforward estimates is similar for both networks and clearly improves the closed-loop tracking accuracy. In experiments the 4-DOF approach shows better results compared to the 3-DOF network. However, the current DeLaN+D implementations do not offer better estimates than a first principles model.