Despite supercomputer level control systems and pin-point GPS data, the lead vehicle in last month's Grand Challenge race was only able to complete 7.4 miles of a 140 mile course.

Recognizing the difficulty of fully autonomous navigation, a scientist at the University of Missouri-Columbia is designing a semiautonomous approach that might have a better chance of producing useful machines in the near term.

Marjorie Skubic, a computer scientist at the University of Missouri's College of Engineering, has demonstrated a prototype robot that can read sketches drawn on a PDA and then execute a proposed path through a room. Skubic is conducting the research with Missouri colleague James Keller, an expert in fuzzy-logic-based pattern recognition, and with Pascal Matsakis of the computing and information science department at Ontario's University of Guelph.

Robots can be easily equipped with precise sensor and directional-movement systems. The rub for artificial-intelligence researchers is deriving the functionality for constructing a global picture of an environment and then finding a path through it.

The solution for Skubic and her partners is to pursue machines that understand basic spatial relationships in a real-time interactive mode and then use that information to navigate complex obstacle courses or to carry out commands. The human director would supply the high-level cognitive understanding of the space, and the robot would execute low-level distance and motion calculations.

"It turns out that both maps and everyday conversations share a simple set of spatial elements and relationships that are used to navigate around obstacles," said Skubic. Skubic's group is creating a robotic AI system that can understand those basic terms so that human operators would be able to direct robots through a room in an intuitive conversational mode. The goal is to create a more practical and flexible means of directing robots and autonomous vehicles.

Humans use relative spatial terms all the time without specifying concrete data. One might tell a colleague, for example, that the conference room is down the hall and two doors on the right, leaving the task of navigating the hall and finding the door to the colleague's cognitive and motor skills. Similarly, someone directing a robot in a hazardous environment might say: "There is a pillar in front of a doorway," followed by the instruction "Go around the pillar and through the doorway." The robot would be able to parse the sentences for spatial relationships, then use its sensors to detect the designated objects and navigate around them. The scheme thus allows the robot to accomplish simple tasks without needing a complex cognitive processor to understand the entire environment.

The underlying spatial language is based on constructing histograms representing the distance relationships between objects. The parameters used to construct the graphs would include the angles between points in two objects, the distance between points and a "gravitational" factor that falls off with the inverse square of the distance. The histograms are then processed with a fuzzy logic system that can identify objects and their relationships and map them onto conversational phrases such as, "The desk is to your right."

Skubic expects this type of conversational system to be immediately useful in structured environments such as offices or homes. The conversational interface could also be employed in video surveillance systems or for object tracking in various applications. For example, video images of vehicles in motion could be analyzed and reported verbally by computer.

Another direction for this type of conversational control system is being applied to groups of robots. Using a collection of robots designed by her students, Skubic is developing a system that will allow them to adopt configurations and operate as a team.

"Football couches use a sketching method to discuss different play strategies with their team. That type of sketch would be easy to capture with our interface," she said. The individual robots would have built-in routines for common configurations such as forming into a line. The more demanding applications would be the kind of autonomous vehicles the military hopes to create through programs such as the Grand Challenge and the Mars rovers.

"NASA has done a wonderful job with the Mars rover, but because there is no way to rescue the vehicle if it gets caught on an obstacle, they have had to be very conservative with movement," she said.

An informal and intuitive interface would not be appropriate in that situation. However, as intelligent sensors, data fusion and high-level cognitive processing continue to develop, planetary rovers and remote vehicles in hazardous situations may eventually have the skill to respond to conversational commands without committing low-level errors.