Three technologies in which immobots are already taking hold-spacecraft, copiers, and cars-are roughly on the same size scale as the human body. But as Struss is demonstrating in another project, there’s nothing preventing immobots from tackling challenges much larger in size and scope.
In Porto Alegre, Brazil, an industrial metropolis with a population of about 1.3 million, the water utility lacks qualified operators to run all five of the city’s water treatment plants. Struss says that many of “the people looking after the drinking water have limited education about how to do that.” To support the operators, Struss and his university colleagues are collaborating with the utility to build an advisory immobot that will monitor such water quality measures as turbidity, color, acidity, and alkalinity. The goal is to spot trouble as early as possible and automatically propose therapies.
Although making water safe for human consumption may seem a simple matter, Struss says that it is much more difficult to translate that task into models than it is to deal with the workings of an automobile. That’s because a reservoir is more an ecosystem than a machine, and that means that modeling “cannot be limited to looking for faulty components, but [must also look for] unanticipated influences and interactions,” he explains.
If there’s a change in sensor readings, for example, an immobot-controlled water-treatment plant can reason its way back to the most likely cause only if it has enough interlinked models of the physical and biological processes that affect water quality. Say the water’s concentration of iron is hazardously high. The system should hypothesize that an algal bloom is to blame because dying algae change the acidity of water, which in turn redissolves iron in the sediments. Ideally, under such conditions, the immobot would advise operators about short-term and long-term responses: treating the water immediately to remove the iron and investigating the most likely cause of the algal bloom, excess nutrients from fertilizer in agricultural runoff.
A prototype advisory system is being tested in Porto Alegre. But building complete and accurate models of processes rather than just parts-and making all those models work together-remains a bit beyond the state of the art, Struss says. It’s “difficult because you cannot enumerate all the components. Sometimes you don’t know what they are.”
While these challenges are being worked out, immobot software is making its way into smaller, self-contained systems such as copiers and cars. Before large-scale infrastructure technologies can be endowed with model-based reasoning abilities, however, researchers must overcome another barrier: the skepticism of old-school engineers who are accustomed to keeping their machines on the short leash provided by heuristic control software.
At NASA, for example, cautious mistrust of truly autonomous software almost killed Remote Agent even before Deep Space One left the ground. “Mission managers are by their nature extremely risk-averse people,” says Ken Ford, a computer scientist who directs the Institute for Human and Machine Cognition at the University of West Florida. Ford, who was associate director at NASA Ames Research Center at the time Williams was developing Remote Agent, says, “In fact, it was hard to get them to fly it at all, even as an experiment.”
But the success of Remote Agent, along with other demonstrations of model-based reasoning, may be slowly swaying the opinions of even conservative engineers. The Jet Propulsion Laboratory’s Rasmussen, for instance, is working with Williams’s group to develop a model-based program that has been tentatively designated as the main control software for the Mars Science Laboratory, scheduled for launch in 2009. In Edinburgh, Scotland, a small company called Intelligent Applications is working with General Electric International to sell its model-based software for monitoring and diagnosing the behavior of power-producing gas turbines. And the European Commission has launched a project to develop model-based failure-management software for commercial aircraft.
Eventually, researchers say, immobots could become pervasive, helping to control some of our most important infrastructure technologies. In air traffic management, for example, Williams suggests building immobotic systems that would use sensor data from satellites and ground stations to assess local weather conditions, automatically identify and select the safest, most efficient flight paths, and redirect air traffic.
“At the grand scale,” Williams says, “you can address a set of problems people have never tackled before.” If he and his colleagues are right, the only way to make infrastructure technologies autonomous without increasing the risk of massive software-related failures may be to program them with distinctly human qualities, such as the ability to plan ahead and solve problems creatively. And in a world that seems increasingly dangerous, knowing that immobots are looking out for themselves-and for us-could be a source of comfort.