Detailed Research Interests

I am interested in building robots and related systems that are cost-effective in today's marketplace. It is clear that sensing and cognition have a long way to go before an autonomous system can match the ability of even a small child. Yet, it is also clear that autonomous systems have a place in our world now if they can compete with humans because they are better, faster, cheaper, safer or even more entertaining.

I am also interested in the study of the mobile robot as a wholistic system. Aspects of that study include the interrelationships of subsystem and system level requirements (constraints), the development of performance models (objectives), and highly principled and well-argued approaches to the construction of the best achievable solution. I want to understand what works well and why so that I can profitably manipulate the governing relationships in order to do even better.

Some recent area os research include:

Off-Road Mobility

The goal of this work is to improve the performance and reliability of vehicles that drive themselves outdoors, off the road. I am interested in the systems aspects of constructing high-performance autonomous vehicles.

Vision-Based Localization

There are basically two methods of position estimation: deduced reckoning and triangulation.. Triangulation methods are those which associate measurements gathered on the vehicle with some prestored map of the area to produce a "fix" on the position of the vehicle. Systems based only on deduced reckoning get more and more lost over time, so some form of triangulation is almost always necessary in practice - if accurate absolute positioning is desired.
I have developed a visual guidance system for factory robots which uses only cameras (vision) and odometry sensing (wheel rotations). It uses no infrastructure such as laser retroreflectors or wires or magnets embedded in the floor. It relies only on the naturally occuring texture on the floor. This system is an existence proof that competent and reliable robot localization can be accomplished with nothing but vision and the odometry sensing available on any mobile robot. Coincidentally, the image processing involved is a 2D analog of the Global Positioning Satellite navigation system (GPS).
There are many ways in which this principle can be used in different ways and in different environments including the outdoors. I would like to extend this work to having ground vehicles navigate from overhead imagery and bulding autonomous vehicle convoys based on nothing but vision.

High Speed Motion Planning & Control

While much is known about how to handle the problem of motion planning in general, relatively little work has been done on the problem of planning robot motions using highly complex models in difficult environments or planning at such high speeds that a good decision must be made in a very small fraction of a second. I am interested in addressing these challenges in both practice and in theory. Of late, I have been working on effective path sampling strategies for high speed and complex environments.

Alonzo Kelly <alonzo@cmu.edu>

Back to Robotics Institute Homepage