To complete the system, they constructed neural-network circuitry that mimics the behavior of nerve cells. Redowl is not so much programmed as it is trained to recognize gunshots, Deligeorges explains. When it’s exposed to a sound, it guesses the location of its source. The researchers feed in the difference between the guess and the correct location, and the trial is run again. Each time, the “neural” connections in Redowl change slightly until the robot can always guess correctly.
“As long as we know how the processing works in a biological system and what’s important, we can take the best part of the biology and the best part of the electronics and merge them,” says Deligeorges. Redowl’s electronics, for example, make it capable of reacting much faster than the human brain.
To account for the intricate and confusing surfaces that reflect sound in an urban environment, Deligeorges has built echo suppression into Redowl. The system recognizes the distinct soundprint of a gunshot – both the initial blast and the shockwave from the bullet – and stores it in memory. Since the echoes that follow will have a similar print, the system can ignore them.
In addition to suppressing echoes, this soundprinting capability can also reveal the difference between an AK-47, an M-16, and city background noise, such as a car backfiring, says Glenn Thoren, Deputy Director of the BU Photonics Center.
In fact, Thoren has a more ambitious system in mind. Already, Redowl can illuminate the target, something other devices don’t do. But Thoren wants to integrate Redowl’s acoustic sensors with optical sensors and other types of detectors. His device would include multiple infrared lasers for pointing to the target, along with a 300x zoom lens and a laser rangefinder. An onboard GPS unit would translate a shooter’s calculated position into geographical coordinates. Such a robot “scout” could move ahead of troops into dangerous locations, such as buildings and open intersections.