Artificial Intelligence Seminar - Armin Lederer, ETH Zurich

Title: Gaussian Process-Based Online Learning Control for Safe Autonomy

Abstract:

In the past decade, technological advancements have laid the foundations for robotic systems that have the potential to fundamentally change our lives. This concerns for example medical problems, where robotic devices can be attached to humans to facilitate rehabilitation or to assist them after losing sensorimotor capabilities. Other examples can be found in search and surveillance problems, where swarms of underwater and aerial vehicles can be used to explore dangerous environments. The complexity of such application scenarios and the challenging environments the robots often must execute them in poses unprecedented requirements on autonomous systems. On the one hand, they need the capability to adapt their behavior to uncertain and dynamically changing environments. On the other hand, the autonomous operation of robotic systems generally requires safety certificates, in particular when they operate in the proximity of humans or in remote locations.

In this talk, we will address these challenges with a particular focus on high frequency, low level control by presenting a safe online learning control approach based on Gaussian process regression. The probabilistic foundation of Gaussian processes provides us with an explicit representation of the uncertainty of a learned model. This allows us to adapt the robustness of control algorithms, such that safety can be effectively ensured through continual replanning. By establishing a direct connection between data and local model uncertainty, we show that simple sampling and online learning strategies can already provide strong performance guarantees for learning control systems. To realize this beneficial behavior on resource-constrained systems, we propose computationally efficient yet guarantee-preserving approximations. The efficacy of the developed methods is illustrated using realistic simulations and real-word experiments throughout the presentation.