Learning and Control with Inaccurate Models
Speaker: J. Zico Kolter , Stanford AI LabContact:
Date: July 27 2009
Time: 11:00AM to 12:00PM
Location: 32-D463 (Star)
Host: Russ Tedrake, CSAIL
Russ Tedrake, firstname.lastname@example.orgRelevant URL: http://ai.stanford.edu/~kolter/doku.php
We consider the problem of learning to control complex dynamical systems, specifically focusing on the examples of dynamic locomotion for quadruped robots and vehicle control at the so-called "limits of handling" (e.g. when the vehicle is sliding over the ground). A common thread in these and many other control problems is that often it is extremely challenging to build an accurate model of the system, making many existing control and planning algorithms perform poorly. In this talk I will present methods we have developed for dealing with such systems, which allow us to obtain state-of-the-art performance in many settings without the need to explicitly build accurate models. In particular, I will first present a policy search algorithm based on what we call the Signed Derivative, which allows us to learn controllers using a very inaccurate type of model that highlights accuracy only of certain model derivatives. Then, I will present a method that augments this and other control approaches by incorporating observed trajectories and multiple models into the control algorithm. We apply these algorithms to the aforementioned tasks of quadruped robot and vehicle control and demonstrate excellent performance: we present a quadruped robot controller that can robustly jump up stairs and a controller for a full-sized car that can robustly slide sideways into a narrow parking spot.
Zico is a Ph.D. candidate working with Andrew Ng.
See other events that are part of Robotics Seminar Series 2009/2010
See other events happening in July 2009