Abstract: While there has been major investment in developing robot learning algorithms, achieving true autonomy remains a wide-open research question. A key limitation of current learning approaches is the assumption that learning is a one-time event on a pre-defined task-distribution and that an expert can manually specify how a robot should learn these tasks from multi-modal sensor data. This approach to setting up the learning problem fundamentally hinders robots from being able to learn new tasks or adapt to new situations autonomously. In this talk, I will discuss the challenges of the frontier of a lifelong learning robot and what algorithmic advancements are required to advance the state-of-the-art in autonomous robotics. In short, progress towards lifelong learning robots requires algorithms that enable a robot to learn new skills incrementally and continuously (without forgetting), autonomously (without expert intervention) and sample-efficiently. In this context, I will present our recent advances towards autonomous learning of robotic manipulation skills by enabling the robot to learn reward functions while only requiring minimal expert intervention. Specifically, I will present a unified framework for model-based and model-free reward learning algorithms, that can either learn rewards from demonstrations or under-specified (sparse) rewards. Finally, I will discuss the challenge of learning rewards that generalize to novel settings, and present initial insights from our empirical analysis of how well learned rewards generalize.

Bio: Franziska Meier is a research scientist at FAIR (Facebook AI Research). Previously she was a research scientist at the Max-Planck Institute for Intelligent Systems and a postdoctoral researcher with Dieter Fox at the University of Washington, Seattle. She received her PhD from the University of Southern California, where she defended her thesis on “Probabilistic Machine Learning for Robotics” in 2016, under the supervision of Prof. Stefan Schaal. Prior to her PhD studies, she received her Diploma in Computer Science from the Technical University of Munich. Her research focuses on machine learning for robotics, with a special emphasis on lifelong learning for robotics.