Abstract: Demanding performance requirements combined with increasing complexity, uncertainty and human interaction in many emerging application problems, e.g. in robotic, transportation, or power systems, are pushing traditional control methods to their limits. A new opportunity to address these challenges is offered by sensor technologies with the ability to collect large amounts of data online. While machine learning provides powerful techniques to analyze and utilize such large-scale data, safety concerns when integrating them in a closed-loop, automated decision-making process represent a key limitation for leveraging their potential.

In this talk, we will discuss some of our recent work towards an automatic controller synthesis that utilizes online data to enhance system performance, while ensuring satisfaction of safety conditions at all times. We show how a predictive controller can be systematically tailored to the particular system at hand by improving predictions, quantifying uncertainties and/or tailoring the objective function online based on data, providing a high performance controller with reduced development times. Then, a safety wrapper is introduced that exploits reachability analysis to ensure satisfaction of constraints for any online control scheme. The key novelty is the learning capability of the wrapper itself, utilizing data to find the largest region of safe operation where a performance-maximizing controller can be employed. Finally, experimental results are shown for a quad-rotor safely learning to fly.

Abstract: With the success of machine learning, we increasingly see learning algorithms make decisions in the real world. Often, however, this is in stark contrast to the classical train-test paradigm, since the learning algorithm affects the very data it must operate on. I will explain how statistical confidence bounds can guide data acquisition in a principled way to make effective decisions in a variety of complex settings. I will discuss several applications, ranging from autonomously guiding wetlab experiments in protein structure optimization, to safe automatic parameter tuning on a robotic platform.

Abstract: Programming robots remains notoriously difficult. Equipping robots with the ability to learn would by-pass the need for what often ends up being time-consuming task specific programming. In this talk I will describe the ideas behind two promising types of robot learning: First I will discuss apprenticeship learning, in which robots learn from human demonstrations, and which has enabled autonomous helicopter aerobatics, knot tying, basic suturing, and cloth manipulation. Then I will discuss deep reinforcement learning, in which robots learn through their own trial and error, and which has enabled learning locomotion as well as a range of assembly and manipulation tasks.

Bio: Pieter Abbeel (Associate Professor, UC Berkeley EECS) works in machine learning and robotics, in particular his research is on making robots learn from people (apprenticeship learning) and how to make robots learn through their own trial and error (reinforcement learning). His robots have learned: advanced helicopter aerobatics, knot-tying, basic assembly, and organizing laundry. He has won various awards, including best paper awards at ICML and ICRA, the Sloan Fellowship, the Air Force Office of Scientific Research Young Investigator Program (AFOSR-YIP) award, the Office of Naval Research Young Investigator Program (ONR-YIP) award, the DARPA Young Faculty Award (DARPA-YFA), the National Science Foundation Faculty Early Career Development Program Award (NSF-CAREER), the MIT TR35, the IEEE Robotics and Automation Society (RAS) Early Career Award, and the Dick Volz Best U.S. Ph.D. Thesis in Robotics and Automation Award.