Enabling a mobile service robot to move in a human populated environment is not only a question of safety but also of predictability, consistency, and the general feeling of comfort of the human. There are currently many approaches of solving this issue but most of them are built on constraints and static learning methods and not on long-term learning through interaction. The main focus of this talk and the underlying thesis, therefore, is the creation of novel approaches to shape a robots spatial behaviour ”on-the-fly” using long-term experiences from engagements in joint movements with lay users in an elder care home and trying to find and understand adaptation needs and thereby create a predictable, readable and consistent robot behaviour.
The techniques used build on the popular Robot Operating System (ROS) navigation stack and use a state-of-the-art local planner to transfer qualitative representations of the interactions between human and robot into discrete movement commands. The Qualitative Trajectory Calculus (QTC) is the underlying qualitative representation for the robots behaviour and the means of identifying the current situation based on a particle filter approach. The initial Markov Models used for the classification and the probability tables to determine the robots behaviour are learned from demonstration and are altered (shaped) towards the users preferences using explicit and implicit feedback.
Christian Dondrup is currently a PhD candidate at university of Lincoln. His topic is: "Shaping human-aware navigation and human-robot joint motion using long-term adaptation." He is working in the STRANDS, 4-year EU FP7 Integrated Project. Previous he was a MSc student in Bielefeld University, with the topic: "Keyword Tagging and Learning using a Detection Based System in a Human-Robot Tutoring Scenario".