The movement of the human is recorded so that it is reproducible for the robot. (Graphic: DFKI GmbH)
The movement of the human is recorded so that it is reproducible for the robot. (Graphic: DFKI GmbH)
Thanks to human negative feedback, the robot learns from its own misconduct. (Graphic: DFKI GmbH)
Thanks to human negative feedback, the robot learns from its own misconduct. (Graphic: DFKI GmbH)

Advanced AI - Interactive Machine Learning

The team "Interactive Machine Learning" aims to develop methods and approaches that enable machines to learn from human-machine or machine-machine interactions. Machines in this context can be either robotic systems or synthetic agents that interact with their pendants in simulated or real environments. Using these approaches, autonomous robots and synthetic agents, which operate in complex environments along humans or other systems over long time periods, will be able to continuously learn. As a result of these interactions that provide the basis for learning, robots will be able to not only improve their own behavior but also quickly adjust to different challenges within their team of other machines and/or humans. This will allow for a sustainable cooperation within a team that optimally utilizes each member’s expertise while facilitating the exchange of skills and knowledge. 

The team develops new machine learning approaches that enable autonomous robots based on simple, generalizable behavioral primitives to learn complex behavior in order to learn from the interactions themselves. For human-machine interactions, the behavior of the robot can be adapted in ways that increase its predictability for humans, thus making it a more easily acceptable interaction partner. The robot’s predictability for humans is in turn directly linked to the human’s predictability for the system, which is an important aspect of human-robot-interaction safety. For machine-machine interactions, there is broad potential for optimization within the machine modeling and the cooperative and competing behavior for complex, structured and unstructured problems in robot interactions.

Team lead: Dr. Elsa Kirchner, Dr. Patrick Draheim

Intrinsic interactive reinforcement learning: Using error-related potentials

Thanks to human negative feedback, the robot learns from its own misconduct.

© DFKI GmbH
last updated 16.07.2019
to top