Lecture Title: "Enabling Robot Videographers to Record the Visual Footage that Human Experts Want"
Presented By: Florian Shkurti, University of McGill
Abstract: The adoption of robotics is becoming widespread in many sectors ofsociety, most notably in the contexts of automated transportation, warehousing, and advancedmanufacturing. Yet, for robots that operate in more challenging and unstructured natural domains(e.g. underwater, air, deserts, forests, lakes), where the promise of automated environmentalmonitoring presents exciting possibilities for societal progress, open research problems stillabound. In this talk I will focus on the problem of enabling robot videographers/documentariansthat autonomously navigate in unstructured 3D environments, alongside scientists, to help themrecord visual footage that they deem valuable for their work. I will present a method to inferthe expert's reward function over images, using a small number of labeled and a large number ofunlabeled examples. This reward function is used to guide the robot's exploration and datacollection in unknown environments. I will also present vision-based algorithms for tracking andnavigation that are robust to long-term loss of visual contact with the subject, by making useof the subject's learned behavior, estimated via inverse reinforcement learning. Finally, I willdescribe a visual and inertial localization and mapping method that enables robust navigation ina wide range of challenging environments. Experimental validation of these methods onunderwater, aerial and ground robots will be shown.
Biography: Florian Shkurti is a Ph.D. candidate in computer science and robotics at McGill University, working with Gregory Dudek. His research is at the intersection of mobile robotics, computer vision, and machine learning. His favorite research problems revolve around increasing the autonomy of mobile robots, and include: inverse reinforcement learning, imitation learning, the control of dynamical systems under uncertainty and partial observability, visibility-aware multi-robot path planning, as well as robust visual mapping and localization in 3D. He is a recipient of the Lorne Trottier Fellowship, the AAAI-15 Robotics Fellowship, and the NSERC Alexander Graham Bell CGS Doctoral Award. He is also a member of the Center for Intelligent Machines and the NSERC Canadian Field Robotics Network. He did his M.Sc. in computer science at McGill, and his Hon. B.Sc. in computer science and mathematics at the University of Toronto.
For additional information, please contact Eyal deLara at email@example.com