Building a vehicle that can learn to fly from scratch

“If something can learn to do something from scratch then it can learn to change as well… Maybe it’s been damaged being shot or its bumped into something—we can change its control parameters, so it can adapt to that and not fly out of control.”

Associate Professor Matt Garratt from UNSW Canberra’s School of Engineering and Information Technology is part of a team undertaking an ambitious project—building a machine that can teach itself to fly from scratch.

The flight platform uses flapping-wing flight, but rather than being based on a bird, it is a four-wing vehicle operating more like a dragonfly.

“It has four wings, it can propel itself and can take off, fly around sideways, backwards and forwards and hover,” said Garratt.

“It’s also going to be quite small.”

The four wings as well as the complexity of flapping-wing flight in general makes simulating the possibilities of flight for the platform incredibly challenging.

This complexity has led Associate Professor Garratt and his team to work with real models rather than a simulation.

“It’s very complicated, you’ve got wing flapping altitude, wing flapping frequency, there are lots of different parameters to deal with – the system is very non-linear,” said Garratt.

“We’re testing some of these concepts on multi-rotor drones, like hexacopters and octacopters—eight rotors or six rotors—and we’re about to begin testing the flapping wing stuff on real hardware. We’re aiming for this thing to be able to teach itself how to fly from nothing.”

Applying “lifelong learning” to a flapping-wing vehicle involves working with complex artificial intelligence to determine how the vehicle will respond to its environment. Using neural network development modelled on the human brain, the vehicle adjusts the way it responds to its environment based on whether its actions—e.g. how fast its wings flap, the height of the flapping—achieve a desired result.

But learning to fly from scratch isn’t the actual goal, it’s a means to an end. A flapping-wing vehicle that can teach itself how to operate in its environment can also adapt to new environments, or to changes to the machine itself.

“We’re looking at how robots and autonomous systems can be more adaptive, so more resilient when things change,” said Garratt.

“Robots can do pre-programmed things, which is all very good, but when things go wrong, they fail, they can’t cope with the change.”

Garratt gave the example of an 8-rotor drone attempting to keep flying after one of the rotors has been damaged. The lifelong learning processes his team is working with would allow the remaining seven rotors to adapt so that it can continue to fly.

“So, if something has learnt to fly at sea level and it then needs to fly on a mountain, it needs to be able to adapt to the changes in air pressure,” he said.

“Or, for example, if its battery voltage drops down. Maybe it’s been shot down, or maybe it’s bumped into something—we can change its control parameters so it can adapt to that and not fly out of control.”