ABSTRACT: With recent advances in machine learning for intelligent agent design, the robotics community has made immense progress towards autonomous skill learning. However, the day of deploying robots to assist people in their homes still seems far away. One of the key challenges towards this goal is to enable a robot to robustly and safely deal with unknown situations and unexpected events.
Current robot learning approaches assume that one can prepare a robot for every possible task and environment variation. With this assumption learning becomes a large data collection effort, followed by a one-time training phase. Once training has converged, learning of the system is terminated and the robot is left to execute it’s skills in the real world, without the ability to further adapt and improve them. Yet, to be truly autonomous, robots need to be able to react to unexpected events and then update their models to include the just encountered data points. Furthermore, a robot’s ability to adapt to new situations should improve over time. In short, true autonomy requires continuous learning. However, continuously updating models without forgetting previously learned mappings remains a fundamental open research problem. In this talk, I will present learning algorithms, based on localized inference schemes, that alleviate the problem of forgetting when learning continuously. Furthermore, I will introduce our recent advances on learning-to-learn for robotics, which accelerates learning of novel task variations. I will demonstrate the effectiveness of our learning algorithms on the challenging task of learning the dynamics of a 7-DOF torque-controlled manipulator, which operates at a 1000Hz. Finally, I will conclude this talk by presenting my vision on how to enable life-long learning for robotics.
BIOSKETCH: Franziska Meier is a research scientist at the Max-Planck Institute for Intelligent Systems since August 2016. She is also a postdoctoral researcher with Dieter Fox at the University of Washington, Seattle, since February 2017. Before that she was a PhD student at the University of Southern California. 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 and attended the Georgia Institute of Technology as graduate student in Computer Science. Her research focuses on machine learning for robotics, with a special emphasis on continuous learning for robotics.
Date(s) - Feb 22, 2018
3:00 pm - 4:00 pm