Machine learning helps morphing-wing UAV land in cramped space

An unmanned aerial vehicle has carried out a perched landing, controlled by machine learning algorithm, for the first time.

The achievement, by a team at BMT Defence Services and Bristol University, could lead to the development of efficient, morphing wing UAVs that can land in small or confined spaces, to deliver aid or gather intelligence.

Existing aircraft are either fixed wing, which are very efficient but have limited manoeuvrability, or multi-rotor, which are very good at landing in small locations, but are inefficient, according to Antony Waldock, principal systems analyst at BMT Defence Services.

So in a project funded by the Defence Science and Technology Laboratory (DSTL), as part of its Autonomous Systems Underpinning Research (ASUR) programme, the team set about designing a UAV with morphing wing structures inspired by birds.

“The approach we took isn’t limited to perching, it could be applied to just about any type of manoeuvre, what we were interested in was being able to control the aircraft in a flexible way, similar to the way birds fly,” said Waldock.

The UAV is equipped with wings that can sweep forwards and backwards, with hinges in the middle, allowing the aircraft to pitch up and down very quickly. The wings can also twist at the end to allow for roll control.

Modified Bixler aircraft with variable wing sweep and tip twist (Credit: Colin Greatwood, Bristol University)
Modified Bixler aircraft with variable wing sweep and tip twist (Credit: Colin Greatwood, Bristol University)

The researchers generated a numerical model of the aircraft using wind tunnel data, in which they could move the wings forward and back to generate different configurations and monitor the effect, in a similar way to a computer game.

They then encoded the particular objective they hoped to achieve, for example to arrive at a certain location, with minimal velocity, in order to avoid a crash landing.

“Next, we allowed the machine learning algorithm to determine how best to achieve that objective, from a range of different starting conditions,” said Waldock.

The model determined which trajectory to take, in order to achieve the best “score” in the computer game.

“We trained the algorithm, by letting it play the game over and over again, to learn the model, and then we could query the model, for any given starting condition, to find out the best trajectory or approach to follow to land at the desired location,” said Waldock.

When the researchers tested the UAV at altitude, they found it was able to carry out a perched landing at the same location multiple times, with minimal error.

One of the UK’s top zoologists is leading an effort to develop a flapping wing military drone inspired by the natural world’s most accomplished aviator. Click here to read more.