Stanford's 'Autonomous' Helicopters Teach Themselves to Fly
Stanford University (08/29/08) Stober, Dan

Stanford
University computer scientists have developed an artificial intelligence (AI)
system that enables robotic helicopters to teach themselves how to fly and
perform difficult stunts by watching other helicopters perform the same
maneuvers. Professor Andrew Ng says the stunts are by far the most complex
aerobatic maneuvers flown by any computer-controlled helicopter. The helicopters
learned how to perform the stunts by watching a helicopter controlled by expert
radio control pilot Garett Oku. After observing the human-controlled helicopter,
the AI-controlled helicopter performed a variety of stunts on its own. The air
show is an important demonstration in "apprenticeship learning," in which robots
learn by observing an expert instead of having software engineers attempt to
write the instructions from scratch. "I think the range of maneuvers they can do
is by far the largest," says Georgia Institute of Technology professor Eric
Feron. "But what's more impressive is t he technology that underlies this work."
To teach the helicopter to fly, the researchers had Oku and other pilots fly
entire air show routines while recording the movements of the helicopter. As
maneuvers were repeated several times, the trajectory of the helicopter varied
slightly with each flight, but the learning algorithms were able to discern the
ideal trajectory the pilot was seeking, enabling the autonomous helicopter to
learn to fly the same routine better and more consistently than the human
pilots. During autonomous flight, a ground-based computer processes the data,
makes quick calculations, and sends new instructions back to the helicopter 20
times per second.
Web Link