Robust Robot Learning
Period: 2021 – 2024
Learning methods offer an attractive alternative to traditional robot programming. A flexible use of learning robots in modern production, however, requires the consideration of uncertainties, such as environmental changes and disturbances, during the learning process. Such uncertainties are taken into account through a reformulation of the learning process as a two-player game, where the adversary represents the uncertainty.
Solving the two-player game leads to robust solutions against the adversary and thus, depending on the formulation of the adversary, also against the uncertainties. While there is a lot of research on the fundamental formulation of robust learning algorithms, few is concentrated around industrial robot applications. In this context, algorithms with efficient computation, as well as their deployment in industrial robotic applications are being explored.
Contact: Janosch Moos