SEATTLE — Robots stole the show at the recent International Joint Conference on Artificial Intelligence hosted by the American Association for Artificial Intelligence.

Besides the fifth annual RoboCup, which featured robots playing soccer, other competitions included robots rescuing people trapped in a collapsed building, interactive hors d'oeuvre-serving robots trying to be the most conversational "butler," plus a national Botball tournament among high-school teams of robot builders from all over the country.

"We look to the natural world for inspiration; every stance must express an intention," said Damian Isla, researcher at the Massachusetts Institute of Technology (MIT) Media Lab, about his "dog," Duncan, in the paper "A Layered Brain Architecture for Synthetic Characters," co-authored by graduate student Robert Burke and professor Bruce Blumberg.

Like Sony's cute dog robots, MIT's Duncan endears itself by design. The animated canine can perform all the standard dog commands, plus he "herds" sheep. More important, Duncan builds an internal model of his world from his senses to perform unstructured commands like "find" a sheep that has strayed. "We chose a dog because they are very empathetic around people . . . and canine psychology helped formulate our brain model," said Isla.

The brain model permits Duncan to extract "perceptions" of objects from his stream of raw sensory inputs and to formulate actions that respond appropriately to those perceptions. To do all that, his brain model needed a short-term memory in order to meld the senses of sight and sound and coordinate the two in terms of locations so that an object's visual appearance must come from the same location as its sound. From there an action can be chosen and the navigation subsystem invoked to plan a route that enables the action to be carried out.

At the Seattle conference, more than 2,600 attendees listened to over 200 presentations, which included technical papers that fleshed out all the aspects of creating a robotic intelligence in the likeness of the living, with sessions like "Cognitive Robotics." "Spatial vs. Temporal Reasoning," "Causality," and "Belief Revision."

Whole sessions were devoted to such subtopics of cognitive modeling as modeling with diagrams, modeling with categorization and modeling with "perceptual grounding." The newest approach, grounded models, differs from past approaches in that it encourages the robot to synthesize its concepts directly from its sensors. In contrast, previous robotic "reasoning" was based on axiomatic systems — diagrams or categories — that were preprogrammed into it ahead of time by their "creator."

Grounded models instead start with an empty set of categories that self-organize depending on their sensory inputs. As explained by professor Josefina Sierra-Santib of the Escuela Tcnica Superior de Informtica at the Universidad Autnoma (Madrid, Spain), "grounded models are based on conceptualization of information gathered by sensors and support a form of intuitive reasoning, which can become the basis of [self-organized] axiomatizations." Intuitive reasoning, for Sierra-Santib, ensues when a grounded robot tries to fit new sensory inputs into the categories it has self-organized from previous sensory inputs.

Other sessions covered specific types of cognitive functions, such as planning. The planning sessions focused on improvements in traditional forward chaining (deducing possible future events from current events) as well as high-profile problems, such as dealing with incomplete knowledge and uncertainty, or applying rule-of-thumb heuristics to simplify problems.

New approaches

Several papers addressed combinatorial "explosions." One such paper showed a new approach invented by Emmanuel Guere and Rachid Alami, researchers at the French National Center for Scientific Research's Laboratory for Analysis and Architecture of Systems (Toulouse, France).

Their "Shaper" algorithm solves combinatorial explosions in planning algorithms by first simplifying a multidimensional task state space into a "shape." Shaper then used this distilled version of the data when comparing competing task scenarios, thereby avoiding the "explosion" of different combinations that result from planning with a fully detailed state space, according to the authors.

Machine learning sessions had separate tracks for hardware-robot learning and software-agent learning, as well as tracks for reinforcement learning, knowledge acquisition and inductive logic.

Some research, such as the MinPath learning navigator for wireless personal digitial assistants (PDAs), is already leading to practical real-world applications. MinPath works by learning the behavior of wireless PDA visitors to a Web site. It then suggests short-cut links to new visitors that can drastically reduce their connection delays.

This artificial intelligence consists of an algorithm that solves the "deep links" problem for PDAs — that is, some Web sites make you click through a series of descending menus that bog down bandwidth-limited wireless PDAs. MinPath learns the short-cut links to deep-link locations that other wireless PDA users have found helpful, according to the authors, professors Pedro Domingos and Daniel Weld at the University of Washington, Seattle, and graduate student Corin Anderson.

Neural networks were cited in several papers as being instrumental in solving unstructured problems, such as the so-called road-sign problem. Robots often confront learning situations — passing a road sign, for instance — where they should filter irrelevant information from the event stream while waiting for whatever the road sign was warning about.

Professor Fredrik Linaker from the University of Skovde (Sweden) and professor Henrik Jacobsson of the University of Sheffield (England) described a method that enables a neural network to learn information from a road sign and then postpone responding to it until the cited event occurs.

The technique involved abstracting to a higher level, where only events grounded in sensory-motor interactions — rather than every perceived object — enter the higher-level event stream. This trimmer, more relevant set of events enabled a neural network not only to learn the road-sign problem, but also enabled the robot to deal with unrelated intervening events, like the car in front of it slamming on its brakes, without losing track of the sign.

Of the invited papers, one by researcher Wolfgang Wahlster from the German Research Center for Artificial Intelligence (DFKI; Saarbrücken, Germany) described a European initiative called SmartKom, which is designed to enable Europeans from one country to understand Europeans from any other country even though they are speaking different languages.

SmartKom follows on the heels of the VerbMobile initiative, which ended last year. VerbMobile was a seven-year, $80 million effort to enable Europeans to speak over the phone to each other with automatic language translation. But according to Wahlster, DFKI will up the ante for SmartKom by adding distinctive gestures to each of the English-, French-, German- and Spanish-spoken languages. Communications will then be "delegated" to a virtual assistant who will translate not only the speaker's language, but also the natural hand-waving gestures of the locale.