The U.S. military is looking for a scout that is as quick, maneuverable and smart as a hawk zipping through thick woods to find prey, but looks instead for enemy troops or threats to U.S. ground troops.
The scout has to be a drone, though. Birds won't do. Neither will slow, remote-controlled unmanned aerial vehicles (UAV) that are noisy, slow and have to be flown by a pilot who pays attention to the video screen in the remote and controls both the flyer and its camera.
This scout should be autonomous enough that a military foot patrol or team of emergency responders can turn on the power, point the drone in the right direction and let it go. The drone has to be able to figure out how to search an area, carry cameras, microphones and other sensors and be able to navigate indoors, with no assistance, a "labyrinth of rooms, stairways and corridors," according to an announcement late yesterday from the Defense Advanced Research Projects Agency (DARPA).
In its announcement of the Fast Lightweight Autonomy (FLA) program, DARPA asked for ideas or proposals on how to write the algorithms that would allow a small UAV to fly up to 45 miles per hour while navigating unfamiliar indoor spaces without the help of GPS coordinates or guidance from operators outside.
It's goal is to evolve drones beyond their current status as awkwardly fault-prone remote-controlled gadgets whose intelligence is focused almost entirely on making them easier to maneuver using remote controls.
The FLA drones are supposed to be quick, independent robots who can create and follow their own search patterns, react correctly to the things they find and report back to operators who may have had to do nothing but watch the video feed as the FLA 'bot searches.
That sounds as if DARPA is looking for more than simply real-time sensor-data processing. Building a bot that can read a situation, navigate through it and respond to particular stimuli – identifying flood survivors clinging to debris, for example, and signaling for help – requires something very close to a low-order artificial intelligence comparable to the birds or bugs DARPA uses as examples of what it wants, but without the self-awareness and self-determination obvious even in birds.
What DARPA is going for is a UAV that would be able to dodge trees at impossible speeds while hunting mice through the woods – at least according to the announcement and this really cool POV video of a Goshawk casually threading needles no human pilot could manage at that speed or altitude.
"Birds of prey and flying insects exhibit the kinds of capabilities we want for small UAVs. Goshawks, for example, can fly very fast through a dense forest without smacking into a tree," according to a statement in the announcement from Mark Micire,DARPA program manager. "The goal of the FLA program is to explore non-traditional perception and autonomy methods that would give small UAVs the capacity to perform in a similar way, including an ability to easily navigate tight spaces at high speed and quickly recognize if it had already been in a room before."
The focus is so firmly on developing a workable, independent flight intelligence, in fact, that DARPA doesn't even want to talk about what that intelligence would be flying in.
"The program focuses on autonomy and not on the flight platform, where “autonomy” includes sensing, perception, planning, and control," according to a nother statement from Micire. All the systems developed to satisfy the autonomy and mobility requirements will be tested in the same, DARPA-selected UAV, to keep developers' focus on the ability to fly rather than the mechanism to allow it.
Stefanie Tompkins, director of DARPA's Defense Sciences Office described the goals of the program in terms that were a little more down to earth, however:
"By enabling unmanned systems to learn ‘muscle memory’ and perception for basic tasks like avoiding obstacles, it would relieve overload and stress on human operators so they can focus on supervising the systems and executing the larger mission," she said in the announcement.
The Defense Sciences Office is "DARPA's DARPA," by the way, the division within DARPA specifically responsible for "high-risk/high-payoff research initiatives" that will eventually deliver "radically new military capabilities."
That sounds less like practical artificial intelligence than like a kind of driver-assist system for inexperienced drone pilots who might have trouble flying a consumer drone the size of a gallon of milk into a barn without scraping both sides of the door.
Tomkins – who got her Ph.D. in geology from Brown, and was deputy director of the Strategic Technology Office – was a leader of the "mad-science wing of the Defense Department," whose special interest was augmenting GPS and vision systems to give soldiers the ability to see beyond the visible spectrum and know where they were when GPS craps out using methods like triangulating the location of lightning strikes, according to IEEE Spectrum.
That doesn't sound like someone who would downplay a really ambitious goal. So it's more likely that the so-cool raptor's-eye-view of a high-speed flight through the woods is just an attention-getting, gadget-porn image of the kind of UAV autonomy DARPA actually expects, which is probably something smarter than a Roomba that can fly, but not smart enough to hunt a mouse through the woods.