Machine Learning Based Robotics
par/by Greg Grudic
Department of Computer Science
University of Colorado
Autonomous robot navigation in unstructured outdoor environments remains a critical challenge for tasks such as reconnaissance, search and rescue and automated driving. Completion of the DARPA Grand Challenge (www.darpa.mil/grandchallenge ), was an exciting step toward this goal, but competitors still required extensive use of well chosen GPS way points, sometimes only a few metres apart. Indeed, successful navigation (between way points a few hundred metres apart) in unfamiliar dynamic outdoor environments is a key open research problem. Such Robotic tasks are characterized by a high dimensional input space that represents the world mediated by robot sensors (vision, sonar data, etc). The robot experiences millions of sensor readings at many frames per second, which must be processed and acted upon in real time. The key open questions are: 1) What information must be extracted from sensors? and, 2) How can the robot use this information to act appropriately in the world? Machine Learning techniques offer powerful tools to model complex real world situations and produce coherent behavior. Indeed many of the fundamental goals of Machine Learning are also those of Robotics, which establishes a synergy between the two fields that can serve as a catalyst for advancing theory and practice in both.
This talk will describe our research efforts in reformulating Robotics within the framework of Machine Learning. I will describe the Machine Learning algorithms we are currently applying to navigation in unstructured outdoor environments, which represents work done under the DARPA Learning Applied to Ground Robots (LAGR) program, and NSF funded Human-to-Robot Skill transfer research. These are density based algorithms that use near field (within 5m of the robot) stereo data to build image-based models of traversable and non-traversable terrain, which are then applied to the far field (beyond 5m) for long range path planning. The models are built in real-time as the robot navigates, and are intended to be used over the robot's lifetime. This long term learning poses many interesting questions regarding how to maintain large model sets, and when and how to apply, refine or discard known models. Experimental evidence on actual robot trials shows that this approach significantly outperforms purely stereo based navigation.
Greg Grudic received his Ph.D. in Electrical and Computer Engineering at the University of British Columbia in 1997. He was a Post Doctoral Fellow at the GRASP Lab at the University of Pennsylvania between 1998 and 2001. In 2001 he joined the Department of Computer Science at the University of Colorado, where he is currently an Assistant Professor.