Task Parametrization through Multi-modal Analysis of Robot Experiences

With quickly progressing and increasingly complex robot control and reasoning systems, a large gap of practical real-world knowledge for robots needs to be filled. While two prominent directions exist, namely designing all knowledge manually, or completely bootstrapping it, we emphasize the combination of both: Starting with simple heuristics, we let robots explore a task, record memories, interpret their findings, and improve their own multi-modal understanding to better their own performance. In this work, we present a software system for autonomous robots that allows them to learn task nuances, and make informed decisions based on experience. They store these comprehensive probabilistic models of any task they perform in a robot knowledge service, benefiting from a shared knowledge base and centralized, well-maintained reasoning algorithms.