Model-Based Robot Control with Gaussian Process Online Learning: An Experimental Demonstration

Autonomous robotic systems are increasingly deployed in operating environments with partially known and time-varying dynamics, which requires control algorithms to adapt their behavior accordingly. While it has been shown that this adaptivity can be safely realized by employing model-based control laws in combination with Gaussian process online model learning, the practical implementation usually requires approximations which do not admit theoretical guarantees needed for safe deployment. In this work, we address this discrepancy between theory and practice by demonstrating the practical applicability of Gaussian process-based online learning with theoretical guarantees in a real-world robotic experiment. For this purpose, we propose an online learning control architecture, which employs locally growing random trees of Gaussian processes with prediction error bounds. Moreover, we improve the practical applicability of this method by introducing an approach for online hyperparameter optimization. This allows us to demonstrate the effectiveness of Gaussian process online learning with prediction error bounds for control in hardware experiments.