Structured Prediction Techniques for Imitation Learning


I completed my Ph.D. work at Carnegie Mellon University’s Robotics Institute under Professor J. Andrew Bagnell in 2009. My graduate work focused primarily on imitation learning, structured prediction, and functional gradient approaches to optimization. More recently I have been developing trajectory optimization and control methods for robotic manipulation using tools from Riemannian geometry.

Early on, with Professor Bagnell and Dr. Martin Zinkevich, I formalized a reduction from an inverse optimal control (IOC) form of imitation learning to maximum margin structured classification that prompted a widespread sharing of ideas between the two fields. Since then, we have developed online, batch, and functional subgradient methods to efficiently solve the optimization problems associated with both structured prediction and imitation learning problems in robotics. Collectively, our framework is known as maximum margin planning (MMP), and our specific class of linear and nonlinear gradient-based approaches to solving these problems is known as LEArning to seaRCH (LEARCH). Applications of this framework are diverse and include footstep prediction, grasp prediction, heuristic learning, overhead navigation, LADAR classificationoptical character recognition, as well as parsing and other problems areas that I have not directly worked on myself. A portion of my thesis work is dedicated to generalizing these ideas to high-dimensional configuration spaces where optimal planning itself is intractable and leveraging ideas from Riemannian geometry to improve the computational efficiency of the corresponding control problems. See the pages on IOHC and CHOMP for details.

For the past two years I have been at Google developing large scale learning systems to assess the quality of ad landing pages, but at this juncture I'm moving back into academia and I'll be at the Max Planck Institute of Intelligent Systems (Tübingen) and the University of Stuttgart for the next two-three years. I'm always happy to talk about my research or to discuss ideas. Send me a line.