Leaders of the AI crowd used to dismiss problems of accurate vision and high-speed I/O as the simple low-level stuff, until practical experience with neural networks and fuzzy logic in the 1990s showed that the so-called hard stuff like Boolean reasoning was easy, while the so-called easy stuff, like interaction with the environment, was very, very hard.

In the 21st century, interest in all kinds of robotics has been rekindled by the growing recognition that new intelligent behaviors and perhaps simple sentience can arise from "emergent" networks of parallel, distributed control loops, operating at times without a host processor. This has led to a variety of projects focused on simple systems that respond in real-time to the environment.

For consumer and entertainment platforms, Sony's Aibo dog has triggered a host of imitators among large integrated Japanese conglomerates, small startups in North America, and even adhoc teams of undergraduates participating in RoboCup competitions worldwide. While a few researchers are exploring the potential of two-legged humanoid systems, RoboCup founder Hiroaki Kitano predicted in EE Times last year that humanoid robots would capture, at best, only 10 to 15 percent of the market. The flexibility of four-, six-, and eight-legged systems, as well as wheeled and semi-stationary systems for vertical markets, makes them of far greater interest for those exploring robot target markets in consumer, industrial and military fields.

The Aibo interest sparked a first wave of fascination in the market. Rodney Brooks, Fujitsu professor of computer science at MIT and a founder of iRobot Inc., described in his recent book, Flesh and Machines, how the designers of My Real Baby learned a hard lesson about the manufacturing constraints imposed by the Asian toy industry on robotic applications for toys. DSP giants like Texas Instruments Inc. and Analog Devices Inc. claim they are willing to march aggressively down a cost curve for volume applications, though the levels required to make intelligent dolls cheap may be difficult to achieve for several years.

The founders of iRobot Inc. (Burlington, Mass.) have used that hiatus to explore household applications. The company's Roomba, a room-cleaning robot offered at a relatively low price of $200 or so per dish-sized system, can perform rudimentary cleaning tasks as a hit-or-miss proposition, though a more efficient concept of $10 disposable room cleaners subservient to a $500 or $1,000 control-station robot, may make miniature maids too complex a proposition to replace a vacuum cleaner in the near term. Brooks of iRobot said that navigation remains a problem for room cleaners and lawn mowers: Without a frame of reference like GPS coordinates, maintenance systems perform a random walk in a given area and does not always perform an adequate job.

This points to the difference between emergent and programmed intelligence. Pioneers of top-down robotics in the 1980s, like Brooks' groups at Stanford and MIT, as well as teams under Hans Moravec at Carnegie-Mellon University, attempted to make a robot operate predictably within a pre-defined space by programming all the rules of that space in advance. While a robot's behavior in such a space was deterministic, it was not adaptable, since obstacles or unpredictable problems within the environment could cause the robot's central system to crash.

The notion favored by Brooks and Cynthia Breazeal of MIT in the last 10 years has been to create robots with no general rules or knowledge about a space, but plenty of adaptability. This allows mobile robots to survive and adequately accomplish simple mowing and sweeping tasks without comprehending the physical space, but it doesn't allow them to always do the best possible job for the task at hand.

While the new robot specialists like Science Applications International Corp. and iRobot continue to perfect household systems, the U.S. military and special government agencies can find immediate uses for such "bottom-up" robotic intelligence. Distributed-control mobile robots form the basis of UAV (Unmanned Aerial Vehicle) programs in the Air Force, undersea UAV programs in the Navy, and ground-hugging and earth-penetrating robots used by the Army. Agencies such as Department of Energy and the Environmental Protection Agency can use remote-presence systems for hazardous-materials handling and cleanup.

Robots like Roomba (left) are barely cleaning house; Aibo (center) is a poor replacement for the real schnauzer; but Coco is coming next-a small, ape-like robot that MIT researchers use to explore humanoid intelligence.

Sometimes, precision is obviously required when small robots go to war. But in general-purpose remote-sensing and exploration tasks, military groups and police agencies can make use of the random-walk robots that explore their territory in a haphazard way. In fact, said Ron Sega, the Defense Department's Director of Research and Engineering, in a recent speech, when several robots are used together on common tasks in a swarm, they exhibit what DOD officials excitedly call "emergent hive intelligence"-unexpected behaviors of joint activity, undertaken without the benefit of a host processor sending commands on the control plane.

Darpa, the Defense Advanced Research Projects Agency, is funding this through two separate programs within its Microsystems Technology Office, the MEMS (Microelectromechanical Systems) program, and the Distributed Robotics program. Several smaller Darpa programs out of the Special Projects Office, as well as Sega's DOD-level National Aerospace Initiative program, provide a unifying glue for the government's distributed-robotics mission. Darpa has worked with scores of companies, including TI, Rockwell, Analog Devices, Hewlett-Packard, Honeywell, Silicon Light Machines, Agere, CoreTek, Siemens and Raytheon, on such component fields as inertial measurement MEMS, optical MEMS, and RF technologies for robotics.

For semiconductor vendors, this leads to interesting prospects, particularly for those who have roots in DSP, MEMS-based sensors, and mixed-signal data acquisition devices. Suddenly, the cost of complex 64-bit integer host processors for control-plane tasks goes away, and interesting multilayer robotic intelligence can be implemented in parallel or hierarchical DSP devices priced as low as $5 a chip.

Christophe Lemaire, marketing manager for ADI's MEMS group, said that the Aibo dog was helpful in sparking a general interest in inertial measurement MEMS for consumer and special-purpose robots, though the automotive market remains the biggest single user of inertial MEMS. Such MEMS have found new applications in platform stabilization for antennas in high-end boats and special-purpose planes and UAVs, said Harvey Weinberg, applications engineer in the ADI MEMS group. That could lead to wider stabilizing applications in mobile robots, particularly in wheelchair-like systems for the handicapped or for intelligent follow-ons to the Segway system.

Finbarr Moynihan, product line manager for 16-bit DSPs at ADI, said that the code optimization for BlackFin is much higher than the melding of 320C54x DSP and ARM7 or ARM9 architectures seen at Texas Instruments.

TI, of course, concedes no such thing. Dave Figoli, chief technology officer for the C2000 DSP product line at the Dallas company, said that the new chip is ideal for both nodal control in distributed control loops, as well as for interprocessor communications for systems that use many such DSPs operating in parallel. "In some designs, the DSP can serve as both the central control and a node controller," Figoli said. "A control node must be able to do multiple loops at very high data rates for a distributed system that interacts with its environment."