UTFacultiesETDepartmentsMS3Research ChairsApplied Mechanics and Data AnalysisProjectsOnline physics-based machine learning algorithms for the control of precision robotic systems

Online physics-based machine learning algorithms for the control of precision robotic systems


Many processes in the Dutch High-Tech industry require accurate positioning, e.g., printing, pick-and-place robots for electronic assembly, wafer-stages for lithography and assembly of optical components. In this project the interest lies in medical applications, like image guided needle placement, that require accurate positioning. As an example will be taken the innovative actuator for MRI compatible actuation that was recently developed by Precision Engineering group.

High-precision manipulators for mentioned tasks are often created by model-based mechanical design and control. The mechanical design aims at realising dynamical mechanical systems that can be accurately predicted given physics based models, typically deduced from rigid body dynamics theory. The design models and the predictable dynamics behaviour allow for feedforward compensation of the dynamics and high bandwidth for feedback control.

However, this design and control philosophy typically include few simplifications, some of which cannot be neglected. For example, dynamics effects like friction and the inherently limited compliance are usually not considered. These simplifications are partly aiming to reduce model complexity, and partly are introduced due to limited knowledge of many effects, e.g., production tolerances, temperature changes and load induced wear. On the other hand, machine and deep learning approaches are used to make inferences about the system given training data with no underlying model describing the physics. As the map between the features and quantity of interest is given in terms of some sort of the analytical function represented for example by the deep neural network, these methods cannot place guarantees on their predictive capability, or even stability. On the other hand, well studied classical system and parameter identification techniques, e.g. Kalman filters, that are used in a control based setting allow fast assimilation of data in physics based setting but cope with the inability to deal with nonlinear dynamics and non-Gaussian data sets. Therefore, the main goal of this project is to adopt optimal architectures from the deep learning, and to rephrase them in a more stable and computationally more efficient environment similar to those of Kalman type in which the hidden layers and latent variables have a physical meaning.

People Involved