Drones Get a Lift From AI

Anyone who has ever been stuck in rush hour traffic has dreamed about pressing a button that puts their car into a flight mode that allows them to soar above the gridlock below. But while experimental flying cars have been touted by the popular media for several decades, nothing practical has ever emerged that an average driver could operate. The simple fact of the matter is that flying requires much more skill than driving, and a lot of things can go wrong in the air. It is not really that the technology is lacking, but rather that relatively few people take the time to learn to safely pilot an aircraft.

But could that lack of skill be made up for by technological advances that take the controls when things start going sideways? A research group at the California Institute of Technology believes that this is the path to putting a flying car in every driveway (or runway) in the future. They have developed a machine learning-powered solution called Neural-Fly for Fault Tolerance (NFFT) that leverages what they call “virtual sensors” to transparently make corrections before disaster can strike. And this technology is not just for the flying cars of the future — it could prevent costly crashes in drones, or any other sort of multirotor aerial vehicle.

Multirotor vehicles have many points of failure. A problem with any one of the rotors could spell trouble, so they all need to be monitored for safety. However, there are many factors that need to be measured, so each rotor requires many sensors. Not only is this setup expensive, but it also adds to the vehicle’s weight and adds the risk that the sensors themselves could fail and cause issues.

To address these issues, the team has developed a deep learning algorithm that can account for a number of adverse situations, like rotor failures and gusts of wind. This neural network has been trained on real-life flight data, which enables it to rapidly make corrections that can compensate for these conditions. And crucially, this system does not require a large constellation of sensors. Instead, certain behaviors of the aircraft, like its attitude and position as a function of time, are monitored by the algorithm. The knowledge encoded in the model can interpret these parameters and determine what type of failure or other adverse condition is being experienced as they change.

This system is not necessarily for the novice pilot only. Since it is capable of appropriately responding in a fraction of a second, it could also keep even the most experienced of pilots safer.

Currently, the researchers are working to develop a vehicle called the Autonomous Flying Ambulance, which incorporates NFFT technology, to safely transport patients to the hospital quickly. Looking further ahead, they hope to build NFFT into ground-based vehicles and even boats. This could be good news, as we might experience not only added safety, but also cost savings in the future.