Colloquia

Samenvatting

The end-effector positioning accuracy of a robotic system can be greatly improved by the addition of a feedforward which compensates for the system's known dynamics, significantly reducing the effort required from the feedback controller. The performance of the feedforward controller is highly dependent on the quality of the inverse dynamic model of the system which is conventionally derived using a model-based approach i.e. using laws of physics such as the Euler-Lagrange equations. This approach takes quite some engineering effort and usually simplifications are made, leading to unmoddelled dynamics. 

In this thesis research is done on a data-driven approach, where the inverse dynamic model is identified purely from data using machine learning techniques. A physics informed neural network (PINN) is derived, where the Euler-Lagrange equations are leveraged to constrain the solution of a feedforward neural network (FNN) to be physically feasible. The result is a neural network which still acts as a universal approximator, and is therefore applicable to a wide range of robotic systems, with the guarantee of having physically plausible solutions and consequently being a more robust and accurate approximator than a non-constraint neural network. Additionally the PINN is combined with a FNN working in parallel, where the PINN approximates the conservative forces and the FNN the non-conservative forces. 

The neural networks are trained on measured data from a 2 degree of freedom manipulator with flexure joints, and its approximations are implemented as a feedforward signal. Experiments show that the machine learning based feedforward can successfully represent the inverse dynamic model, and the feedforward significantly improves the end-effector positioning accuracy.