A group of faculty and graduate students has been enhancing Xavier's cognitive skills so it understands its actions, realizes if something has gone wrong and learns from its mistakes.
The project's coordinators, Research Computer Scientist Reid G. Simmons and Computer Science Professor Tom M. Mitchell, envision making Xavier smart enough to travel the hallways of a campus building, ride the elevators, deliver food and drinks, retrieve recyclable items and then, when its battery runs low, come back to the lab and plug itself into the wall.
Their research goal is to create a robot that can plan and account for uncertainty, recognize features in its environment and generalize to learn relevant things about them. It's hoped that by using experiences, Xavier can learn from just a few examples.
Xavier is capable today of generating speech and understanding it via Carnegie Mellon's Sphinx speech recognition system. It can maneuver in unstructured environments without running into things. It contains a topological map of the rooms and corridors on the fifth floor of the university's Wean Hall and uses the map to navigate through the area.
"Xavier knows where it is. We tell it where to go," says Simmons. "It plans a route using the map. It has a fuzzy idea of where it is at all times and, as it recognizes landmarks, it uses them to reduce its uncertainty. It takes into account that some things are more reliable to recognize than others."
The robot is unique in design. Its custom-designed upper body sits on a commercially built base. The robot is equipped with a color laptop computer, two PCs, 24 sonar sensors, a laser light striper, a color camera on a pan/tilt head, eight bump panels and a speech generation board. It also contains a radio ethernet that allows it to connect to the Internet.
At this point, some of Xavier's problems are more social than technical, Simmons says.
"People tend to be malicious and do tricks to trap the robot," says Simmons. "Others are frightened by it. Others try to ignore it. We want the robot to live in the environment like everyone else, but there must be some social adjustment. The main problem is the best-laid plans can go awry. The robot has to expect the unexpected and have contingency plans ready to achieve its goals. It's the 10 percent of the time when things go wrong that takes 90 percent of our research time to fix. Right now we're spoon feeding it about how to recognize and recover from mistakes. We'd like to have it learn from its mistakes automatically.'
Xavier's address on the World Wide Web is http://www.cs.cmu.edu/~Xavier