By JUDY WEST
Vijay Kumar, a professor of mechanical engineering, spends a lot of time thinking about squid. And harvester ants. And killer whales. The director of Penn Engineering’s GRASP (General Robots Automation Sensing Perception) Lab, Kumar believes nature may hold the answers to some of the most complex challenges in the world of robotics.
Specifically, he’s interested in the way some animals exhibit collective behaviors—swarming, flocking and the like—to accomplish a task, and how studying those behaviors could inform the design of large networked groups of robots.
Recently, he and his colleagues at Penn, along with partners at UC Berkeley, MIT, UC Santa Barbara and Yale, received a $5 million grant from the Department of Defense to pursue this research.
“Our hope is that by doing analysis we’ll discover some simple motifs that we can then apply to the problem,” says Kumar. What interests him about the animal behavior is how they communicate and act without a centralized infrastructure. Ants, he says, are particularly intriguing. They have an interesting way, he says, “of identifying potential nests and figuring out which are best and going back and telling everyone else, ‘Hey, guys, we need to live here.’ And remember, their nervous systems are known to be primitive. They don’t have the capacity to do all the calculations for themselves, and they’re not communicating with some central supervisor.”
Kumar can envisage a network of 100, or even 1,000, robots capable of sharing information and arriving at logistical decisions, just like the ants. “The question,” he says, “is can I look at insects and how they build nests and use that to figure out how robots could distribute themselves and look for targets?”
Kumar can see many scenarios, such as a chemical spill or other hazardous emergency situation, where a large team of robots would be preferable to human manpower. And in these circumstances, it would be impossible to control each unit separately.
Already, in a project called MARS (Multiple Autonomous Robot Systems), Kumar and his team have figured out how to get a team of 10 robot vehicles to share information and act as a network. A demonstration at Fort Benning, Georgia, last December, showed how the network could be commanded to look for a target—in this case, anybody dressed in an orange flak jacket—in an urban setting. During the exercise the robots took in sensory information about the environment, stayed in contact with the operator through a wireless link and made decisions for themselves.
In this sense, says Kumar, they were acting autonomously, with little human supervision.“Obviously one of us wrote the code, but it’s hard to predict at any point what the robot is going to do because the decision making process is so complicated.
The vehicles used for that project resemble miniature monster trucks loaded up with high tech equipment including a laptop computer, a global positioning system, radio antennae and a set of cameras. The price tag for each robot vehicle is around $5,000.
The challenge for Kumar and his team—which includes experts in
biology, control theory and algorithms—is how to scale up from
the MARS project. Right now, he says, unmanned vehicles used by the military
still use a lot of manpower. The Predator, for example, needs 10 operators.
“What you want to do is reverse that,” he says. “We
want one operator for every 10 vehicles, and then one for every 100 vehicles.”
Originally published on April 28, 2005