[Colloquium] REMINDER: Talk by Nathan Ratliff Today

[Katie Casey]

DEPARTMENT OF COMPUTER SCIENCE

UNIVERSITY OF CHICAGO

Date: Monday, February 15, 2010
Time: 2:30 p.m.
Place: Ryerson 251, 1100 E. 58th Street

----------------------------------------------------------

Speaker:	Nathan Ratliff

From:		TTI-C

Web page:	http://www.tti-c.org/ratliff.php

Title: Learning to Search: Structured Prediction Techniques for Imitation Learning

Abstract: Modern robots successfully manipulate objects, navigate rugged terrain, drive in urban settings, and play world-class chess.  Unfortunately, programming these robots is challenging and time-consuming; even when the desired behavior is clear and easily demonstrated, the parameters governing the behavior of a given system are unintuitive making it difficult to write a planner or controller that embodies this behavior.  Inspired by successful end-to-end learning systems such as classical neural network controlled driving platforms (Pomerleau, 1989), learning-based "programming by demonstration" paradigms have gained currency as a method to achieve intelligent robot behavior.  But modern state-of-the-art systems in robotics are governed by highly structured planning, control, and sensor processing algorithms.  It is less clear now how to effectively and efficiently train modern robotic systems using classical learning techniques.  My research approaches the problem from a new angle.  Rather than redefining robot architectures to accommodate existing learning algorithms, I develop new learning techniques that leverage high performing tools from modern robotics.

My presentation begins with a discussion of a novel imitation learning framework we developed called Maximum Margin Planning (MMP).  Given demonstrated trajectories, these algorithms generalize the demonstrations by learning a cost function under which optimal planning or control algorithms such as A* perform as desired.  In the linear setting, this framework has firm theoretical backing in the form of strong generalization and regret bounds.  Further, I have developed practical nonlinear generalizations that are effective and efficient for real-world problems.  This framework reduces imitation learning to a modern form of machine learning known as Maximum Margin Structured Classification (Taskar et al. 2005); these algorithms, therefore, apply both specifically to training planners within an imitation learning paradigm, as well as broadly to solving a range of structured prediction problems of importance across machine learning and robotics.

In difficult high-dimensional planning domains, such as those found in many manipulation problems, high-performance planning technology remains the focus of significant  research in modern robotics.  I will present some recent work which moves toward simultaneously advancing this technology while retaining the key characterizations of learnability we leveraged under the MMP framework.

I'll demonstrate these algorithms on a range of applications including overhead navigation, quadrupedal locomotion, heuristic learning, manipulation planning, grasp prediction, driver prediction, pedestrian prediction, optical character recognition, and LADAR classification.


Please note that refreshments will be served after the talk at 3:30 in RY 255.
-------------- next part --------------
An HTML attachment was scrubbed...
URL: http://mailman.cs.uchicago.edu/pipermail/colloquium/attachments/20100215/b930423e/attachment.htm