Intel eyes robotic-controller sockets for Xscale

Intel eyes robotic-controller sockets for Xscale
HILLSBORO, Ore. — Intel Corp. may have uncovered a new direction for its chip-making expertise: low-power, high-performance robotic controllers. The interest generated by such products surfaced at a recent symposium that Intel held on the possibilities of robotic uses for its Xscale microarchitecture.

Experts in robotics from eight major U.S. research labs were invited to present the state of the art in autonomous robots to Intel's engineers here at Intel Research. The presentations revealed a common need for lower power combined with higher performance for robots and wireless appliances.

"Xscale is designed to provide optimal Mips/watt operation for embedded devices like wireless Internet appliances, but we believe that robotics is a good application area for it too," said Patrick Gelsinger, vice president and chief technology officer for Intel.

"Currently, we have zero research efforts in robotics, but we are already creating the software development tools for Xscale applications developers, and we plan to also start a robotics effort very soon — we just wanted to have this conference first to help guide us," Gelsinger explained. Gelsinger worked his way up through the engineering ranks at Intel, literally writing the book on the 386 processor, architecting the 486 and championing "design for testability" across all of Intel's designs.

"We never needed a CTO until recently, but without Intel cofounders Robert Noyce and Gordon Moore, we felt now was the time to get a CTO to help lead the way," said Intel research sector director Eric Hannah.

As the leader of Intel's new Corporate Technology Group, Gelsinger spelled out long-term goals for the Intel Research and Intel Architecture Labs. One is to push "intelligent" RISC performance that optimizes Mips (million instructions per second) per watt by combining low power consumption and long battery life with increased device density and reduced thermal constraints.

"What Intel needs as its basic strategy for the future is to shift our focus from individual components to their intelligent integration, as our Xscale microarchitecture already does," Gelsinger said.

Intel's Xscale architecture expands the ARM embedded microprocessor family by combining low-power "intelligence," the popular Thumb instruction set and extensions for high-speed digital signal processors (DSPs).

Good candidate

The overall architecture lays claim to innate intelligence that combines high performance for video and audio, with low-power consumption and reduced operating temperatures. Xscale accomplishes these simultaneously, rather than trading off speed against power and temperature, by intelligently scaling back wattage in real-time to match instantaneous computing needs. Typical Xscale chips can run between 0.1 watt to 1 watt as their clock speeds are tuned to instantaneous computing requirements, making them a good candidate for autonomous robots, according to Intel.

At the conference, researchers concurred with Gelsinger's vision, namely the convergence of higher computational requirements and lower power. According to the Gelsinger vision, processing audio and video streams, whether within wireless Internet appliances or in autonomous robots, requires ever-higher computational capabilities.

These faster processors, however, must simultaneously consume less power and operate at lower temperatures in order to extend battery life and to increase component density by reducing cooling requirements.

"We are finding that developing cooperating autonomous robotic applications involves very similar problems to those being studied by networked embedded system designers. There is a definite convergence between the distributed robotics community and the sensor net community," said Gaurav Sukhatme, a professor at the Robotic Research Lab at the University of Southern California in Los Angeles.

Seek and find

Sukhatme described a robotic police helicopter at the conference, which could autonomously track fleeing "suspects" by air while reporting their whereabouts wirelessly to its human supervisors. When the suspect tried to elude the shadowing helicopter by crawling underneath a parked car, the helicopter landed and deployed a smaller wheeled robot that it was carrying to flush the suspect out from beneath the car.

Likewise, Georgia Institute of Technology professor and director of GIT's Mobile Research Laboratory, Ronald Arkin, described what he called "marsupial" robots, because they each carry smaller robots inside themselves that can be deployed to fit cramped quarters. According to Arkin, eight years of development work for the Defense Advanced Research Projects Agency (Darpa) on its MissionLab software for coordinating multirobot teams has resulted in marsupials inside marsupials inside marsupials.

"They are almost like those Russian dolls, where a smaller one is always inside the larger ones. The large robots move to a locale and then throw the smaller robot at suspicious locations for up-close surveillance," Arkin said.

According to participating Darpa researchers like Arkin and Sukhatme, the high computational requirements of autonomous robots, which must process audio and video often, makes their CPUs burn more watts than the much larger motors and actuators that provide the kinetic energy of the robot. The motors and actuators may consume tens of watts of power when they are used, but they consume zero power when not being used, whereas the high-speed processors consume watts of power constantly, even when not computing. Intelligent power management, like that supplied by Xscale architectures and other processors, could keep the "marsupials" and their DSPs in sleep mode except when needed.

Even non-defense contractors, like Carnegie Mellon University professor Illah Nourbakhsh, confide that intelligent means for controlling power consumption has been the long-sought solution for autonomous robots. Nourbakhsh refuses defense work from Darpa and the like for ethical reasons, opting to instead market products to fund research, such as the university's popular Palm Pilot Robot. He contends, however, that the low-power problem is endemic regardless of the area to which the robotic product is applied.

As a novel solution, Nourbakhsh added intelligence to offset power consumption of the robots' motors themselves instead of the processors. For instance, his "pogo stick" robot uses gyroscopes to maintain its balance while "hopping" over obstacles. The pogo stick robot recovers over 90 percent of the energy it expends for locomotion by compressing a spring each time it comes back down to earth. By intelligently aiming its "foot" so that it not only recompresses its spring but also rebounds in the correct direction, the pogo stick robot can hop upstairs and over other obstacles that foil walking or rolling robots.

"We think we can refine the design of the pogo stick robot so that it will only need a couple of AA batteries to top off the compression of the spring after each hop," Nourbakhsh said.

The current prototype uses two C-size batteries to bring the compression of its spring up to 100 percent from the 90 percent or more it recovers when hitting the ground after each hop. Already, the computational energy expended to aim the foot is rivaling the power needed to replenish its spring compression, but Nourbakhsh is optimistic that further refinements will reduce the spring-compression requirements to become even smaller than its computational power consumption.