Robo-choppers learn tricks from algorithm


When Stanford scientists say this helicopter has a mind of its own, they mean it. A team of researchers created a robotic aircraft capable of learning complex aerobatic maneuvers.

"I think helicopters are one of the most difficult robots to control and fly and what we have today is a new type of machine learning technology that enables a helicopter to learn to fly by itself," says Stanford assistant professor Andrew Ng.

Instead of trying to program a very unstable aircraft, Pieter Abbeel and Adam Coates gave the helicopter artificial intelligence and on board sensors. Then, they introduced expert radio control pilot Garett Oku to teach the helicopter how to fly.

"In the back of my mind, I thought, 'it's impossible,'" says Oku.

But for four years, the computer watched and learned. The result of all this has the robot making rolls, loops and this dazzling maneuvers.

"It can figure out when the pilot moves one of the control sticks, where does the helicopter go and how does it behave and from that can decide how to control the helicopter itself when we put the computer in charge," says graduate student Adam Coates.

The robotic helicopter knows what to do for any given stunt and can use it's sensors to adjust to variables such as wind.

"So, you just can't play back what the pilot did. The machine learning algorithm has to infer how to fly and how these difficult maneuvers and then replay them reliable," says graduate student Morgan Quigley.

The team is already talking to people about practical uses for a helicopter that can fly itself.

"So, applications like search and rescue, mapping out forest fires and where land minds are located," says Pieter Abbeel.

Until it's talents have a mission, this is the most sophisticated, self learned, computer-controlled air show in the world.

Copyright © 2024 KGO-TV. All Rights Reserved.