This research paper by Dominik Baumann and Thomas B. Schon discusses safe reinforcement learning in the context of uncertain variables. The authors argue that most safe learning algorithms only consider the internal dynamics of a robot and possibly external constraints. However, in real-world applications, robotic systems are subject to changes in their environment that influence their dynamics. These changes are represented by discrete context variables that are often assumed to be known. The authors propose a method for identifying these variables through experiments and provide frequentist guarantees for multi-class classification, enabling the estimation of the current context from measurements.
Publication date: 12 Jan 2024
Project Page: Not specified
Paper: https://arxiv.org/pdf/2401.05876
