Final Report Summary - CON-HUMO (Control based on Human Models)
CON-HUMO developed novel concepts of automatic control, based on data-driven human models and machine learning for enabling innovative control applications. In particular, the scientific and technical developments focused on the area of human-robot interaction (HRI) as the derivation of a model with classical system identification techniques is difficult in this domain because of the complexity and stochastic nature of human behavior. To this end, our research made significant advancements towards deriving i) control properties of data-driven stochastic dynamical system models, ii) control design based on data-driven stochastic dynamical system models, and iii) data-driven stochastic non-parametric dynamic models of human behavior. In particular, we targeted advancements centered on Gaussian Process (GP) models as these models are a very powerful tool for model-based control approaches for nonlinear systems, and the fact that the prediction is based on the underlying data which offers highly flexible modeling techniques.
First, we systematically analyzed specifications of control-relevant characteristics such as equilibria, stability, controllability and observability in data-driven models. By mathematically deriving the stability properties, equilibrium distributions and learning behavior of data-driven models, we were able to enforce stochastic stability conditions for control applications.
Second, we developed stable robot control designs based on data-driven dynamics models. In particular, we investigated a class of stochastic optimal control approaches for robotics applications that explicitly incorporates the prediction uncertainty of GPs and other probabilistic representations of the model into the control design Furthermore, we developed a constrained control approach, which is crucial for guaranteeing safety in physical HRI. With these results, we demonstrated that this type of controls significantly improves the human-perceived quality of robotic assistance while ensuring the interaction is guaranteed to be safe for the humans. Third, to generate human model models for control designs, we developed methods for modelling dynamics of the human behavior. To this end, we employed data-driven identification of dynamical systems as they are advantageous when no analytical model is available and it requires only minimal prior knowledge. As a result, it enabled a more accurate prediction of human motion behavior with psychological theories of human-human interaction, which is an important precursor to generate intuitive robot behaviors. Altogether, the advancements of control theories and modelling techniques for stochastic non-linear systems in the CON-HUMO project demonstrated examples of how intuitive, effective and safe (physical) HRI can be realized in mathematical frameworks with formal guarantees for control stability.
First, we systematically analyzed specifications of control-relevant characteristics such as equilibria, stability, controllability and observability in data-driven models. By mathematically deriving the stability properties, equilibrium distributions and learning behavior of data-driven models, we were able to enforce stochastic stability conditions for control applications.
Second, we developed stable robot control designs based on data-driven dynamics models. In particular, we investigated a class of stochastic optimal control approaches for robotics applications that explicitly incorporates the prediction uncertainty of GPs and other probabilistic representations of the model into the control design Furthermore, we developed a constrained control approach, which is crucial for guaranteeing safety in physical HRI. With these results, we demonstrated that this type of controls significantly improves the human-perceived quality of robotic assistance while ensuring the interaction is guaranteed to be safe for the humans. Third, to generate human model models for control designs, we developed methods for modelling dynamics of the human behavior. To this end, we employed data-driven identification of dynamical systems as they are advantageous when no analytical model is available and it requires only minimal prior knowledge. As a result, it enabled a more accurate prediction of human motion behavior with psychological theories of human-human interaction, which is an important precursor to generate intuitive robot behaviors. Altogether, the advancements of control theories and modelling techniques for stochastic non-linear systems in the CON-HUMO project demonstrated examples of how intuitive, effective and safe (physical) HRI can be realized in mathematical frameworks with formal guarantees for control stability.