Experimental AI Powers Robot Army
Wired News | September 14, 2006
By David Hambling
Darpa's Grand Challenge may have looked tough, but it was a piece of cake compared to the challenge facing robots currently being developed by the U.S. Air Force.
Rather than maneuver driverless through miles of rough desert terrain, these will have to find their way into underground bunkers, map unknown facilities in three dimensions and identify what's in them while avoiding detection -- all without any human control.
This is well beyond the capability of any existing system, but the Air Force Research Laboratory, or AFRL, is putting its hopes on new software that lets robots learn, walk, see and interact far more intelligently than ever before.
It's based on work by Stephen Thaler, who came to prominence 10 years ago with his brainchild the Creativity Machine. This is software for generating new ideas on the basis of existing ones, and it has already written music, designed soft drinks, and discovered novel minerals that may rival diamonds in hardness.
The software is a type of neural network with two special features. One introduces perturbations, or "noise," into the network so that existing ideas get jumbled into new forms. The second is a filter that assesses the new ideas against existing knowledge and discards those that are unsuitable. Current applications range from detecting intruders in computer networks to developing new types of concrete and optimizing missile warheads.
Recently, Thaler has been working for the AFRL on what he calls Creative Robots, which joins his brand of AI to robotic hardware.
â€œDr. Thaler's approach is clever and should have some interesting properties,â€ said Michael Rilee, a NASA researcher who is working on a neural networking project to use bot swarms in space and planetary exploration, known as Autonomous Nano-Technology Swarm, or ANTS. â€œThe chief novelty is in its use of neural nets to train other neural nets.â€
Self-learning and adaptability will be the key to success, and this is where the Creativity Machine excels. Give it any set of robotic limbs and it will master locomotion within minutes without any programming, swiftly finding the most efficient way of moving toward a goal. It will spontaneously develop new gaits for new challenges. (Thaler recounts how a virtual robotic cockroach adopted a two-legged gait and ran on its hind legs, not unlike basilisk lizards, when it needed to move faster.)
Perhaps the most impressive -- and spookiest -- aspect of the project is the swarming behavior of the robots. In computer simulations, they acted together to tackle obstacles and grouped together into defensive formations where needed, Thaler said. They also worked out how to deal with defenders, and spontaneously devised the most efficient strategy for mapping their environment, he added.
"This approach has less chance of getting stuck than any other" when dealing with unpredictable obstacles, according to Lloyd Reshard, a senior electronics engineer at AFRL.
Thaler declined to describe his results in detail, but said his system has produced unspecified "humanlike capabilities."
"I can relate the results of virtual-reality simulations, where swarms of Creativity Machine-based robots have deliberatively sacrificed one of their kind to distract a human guard, enabling the remainder to infiltrate a mock facility," he said.
Owen Holland, a researcher at the University of Essex who is building an â€œultraswarmâ€ of miniature Bluetooth-connected helicopters, said neural networks can be very effective for dealing with changing circumstances: "If you rip a leg off, they'll work out what's happened, and re-evolve a different gait that works."
Still, he admitted the swarm approach has its limits. "The fundamental problem with the swarm intelligence approach is that we cannot usually go from a knowledge of what we want the system to do to a knowledge of the simple rules that will automatically produce the desired result when loaded onto the robots,â€ he said. â€œI can't see us ever being able to do more than repeat the process using artificial evolution or some other open-ended search technique."
Thalerâ€™s current project, which should be completed over the next six months, will develop a piece of software called CSMARRT (for Creative, Self-Learning, Multi-Sensory, Adaptive, Reconfigurable, Robotics Toolbox). The software can be used to design and model virtual robots that can be placed in virtual environments to learn and develop. The user can then view the result to see how neural network modules have â€œknittedâ€ themselves into complex control architectures.
This toolbox can create software to control any robotic hardware, handling locomotion, sensors and intelligent behavior to carry out a mission including swarming. The AFRL will make this available to its customers -- other branches of the U.S. military.
â€œThe biggest challenge is computing power,â€ Reshard said.
Desktops PCs are more than powerful enough to run CSMARRT, but the aim is to get something that will run on as little computing power and fit into as small a package as possible. Exactly how small is anyone's guess, but AFRL left the impression of a scale of inches rather than feet.
The Air Force robots may look like cockroaches, or they may be "snakebots" like those currently in development. Thaler has carried out validation tests by using the software to control small H3 "robot cockroaches." The results are classified, he said.
Click Here to Read the Rest of the Article
Infowars.com is Copyright 2006 Alex Jones | Fair Use Notice