Abstract
Many robotic applications require control of the applied forces or moments. Model predictive control allows for the direct or indirect control of forces, while taking constraints into account. However, challenges arise when the robot environment that affects the force is highly variable, uncertain and difficult to model. Learning supported model predictive control makes it possible to combine the advantages of optimal control, such as the explicit consideration of constraints, with the advantages of machine learning, such as adaptive data-based modeling. In this paper Gaussian processes are used to model the contact forces that are applied in model predictive force control. The Gaussian process learns the static output mapping describing the interaction of the robot with the environment. It is shown that stability guarantees can be derived in a similar way as in classical predictive control. A proof-of-concept experimental implementation of a direct hybrid position force controller for a lightweight robot shows real-time feasibility.
Published Version
Talk to us
Join us for a 30 min session where you can share your feedback and ask us any queries you have
Schedule a call