AI-powered drone continues to fly during tornadoes

Most drones can only operate effectively in ideal weather conditions. However, in order for drones to be able to perform necessary tasks, such as delivering packages or airlifting injured drivers from a traffic accident, they must be able to adapt to wind conditions in real-time.

A team of engineers from the California Institute of Technology (Caltech) have found a solution. They have developed 'Neural-Fly', a deep-learning method that can help drones cope with new and unknown wind conditions, allowing them to fly even in extreme events such as hurricanes.

The five-pound drone can recalculate weather conditions around it five times per second, and adapt its path accordingly, as described in a study published in Science Robotics.

“We can make sure that this drone can land under any weather conditions because of our neural flight method,” said Soon-Jo Chung, a Caltech professor and part of the research team.

The drone was tested at Caltech's Center for Autonomous Systems and Technologies and put through its 'Real Weather Wind Tunnel', a custom 10-foot-by-10-foot array of more than 1,200 tiny computer-controlled fans that allows engineers to simulate everything from a light gust to a gale.

During the demonstration, the drone managed to stay upright with little effort and performed a series of figure-of-eight movements between tight obstacles while being blasted with winds up to 12.1 metres per second without veering off course.

The researchers say the technology could revolutionise everything from air travel to medical rescues.

“In the future, if we have 3D flying cars flying in the air, it's very important to have precise control because you want to make sure the flying car can track the trajectory very accurately,” said Guanya Shi, co-first author on the paper and a researcher at Caltech.

“Our long-term mission is that we want to generalise this technique to not only quadcopters, to flying cars, to any area of robots, such that they can accurately, safely, do complicated tasks.”

The team is developing different types of drones. One of them more closely resembles a plane, but is in reality an autonomous flying ambulance, which uses the same software to reach sites of medical emergencies. The team expects to be able to have this design in use within a year.

“Thanks to our Neural-Fly AI-based control method, basically we can send this flying ambulance to medevac the injured passengers from the traffic accident,” said Chung.

The scientists' meta-learning algorithm is able to pre-train the neural network so that only key parameters need to be updated to effectively capture the changing environment.

After obtaining as little as 12 minutes of flying data, the drones equipped with Neural-Fly learn how to respond to strong winds so well that their performance significantly improved. So far, tests have demonstrated an error rate that is between 2.5 times to 4 times smaller compared to drones currently on the market that don't use neural networks. 

The team has also shown that flight data gathered by an individual drone can be transferred to another drone, building a pool of knowledge for autonomous vehicles.