Machine Learning Seminar

Machine Learning Seminar
November, 01, 2018
10:30
in Room 1061 Electrical Eng. Building Technion City

Mathias Bürger

Bosch Center for Artificial Intelligence
On:

Multi-Robot Coordination under Uncertainty – Integrating Formal Methods, Auctioning and Learning

Planning efficient and coordinated policies for large teams of robots is still a challenging problem. We will particularly point out two additional challenges, observed in real-world applications: First, uncertainties or a lack of knowledge about the environment, as it is typical in environments shared between robots and humans; and second, temporal dependencies within the tasks, as they are typical in service or industrial manufacturing environments. We will discuss in this talk a scalable approach to coordinate a team of heterogeneous robots from a single goal specification, formulated in Linear Temporal Logic, while explicitly considering uncertainty and incorporating observations during execution. Building upon the concept of MDP options, we propose a decentralized auction-based procedure that allows robot to coordinate their concurrent actions during execution and adjust their policies when additional observations are made online. For persistent tasks, including repetitive elements, the approach naturally extends into a learning setting, where online reinforcement learning concepts are used to continuously improve the team performance. We will show in several case studies that the proposed approach enables a tight and flexible coordination of robotic teams, where robots automatically prepare jobs that are only required in the future and adapt to changes in the environment. Bio: Mathias Bürger is heading the research group on Reinforcement Learning and Planning at the Bosch Center for Artificial Intelligence. He joined Bosch in 2014 and since then worked in several research activities at the intersection of artificial intelligence and robotics. Currently, he is also principal investigator for the H2020 project Co4Robots. He holds a PhD degree in control theory from the University of Stuttgart. For his PhD work he received the 2014 EECI PhD Award in recognition of the best PhD thesis in Europe in the field of Control for Complex and Heterogeneous Systems.