As the world quickly adapts to last-minute delivery, there is a need for a humanless solution. Actually, there is already a solution. Since the early 2000s, we’re rapidly advancing into a new era with the push into drones. From pizza to donor organs, the possibilities of their cargo are endless.
Nothing comes easy
Drones are now flown in controlled environments with no wind or by people using remote controls. Although these flights are carried out under ideal conditions, we’ve taught drones to fly in formation in open skies. However, drones must be able to adapt to wind conditions in real-time to do critical tasks. From delivering packages or airlifting injured drivers from traffic accidents, it’s crucial to be able to adjust.
To address this challenge, a team of engineers developed Neural-Fly. This is a deep-learning technology that can let drones cope with new and unexpected wind conditions in real-time by tweaking a few key settings. First, engineers used the Real Weather Wind Tunnel. This 10-foot-by-10-foot array of more than 1,200 tiny computer-controlled fans allowed engineers to simulate everything from a modest gust to a storm.
Not just an equation
The issue is that a simple mathematical model cannot accurately capture the unique effects of various wind conditions on aircraft performance or stability. They don’t attempt to qualify and quantify every effect of the turbulent and unpredictable wind conditions. The engineers use a combination of deep learning and adaptive control. This allows the aircraft to learn from previous experiences and adapt to changing conditions on the fly while maintaining stability and robustness.
Several models are derived from computational fluid dynamics, but achieving the proper method reliability and adjusting that model for each vehicle and wind condition is challenging. Existing machine learning approaches require a lot of data to train. Nor can they compete with state-of-the-art flight performance attained using traditional physics-based methods. Furthermore, a real-time adaptation of a whole deep neural network is a tremendous, if not currently impossible, undertaking.
Researchers claim that Neural-Fly overcomes these obstacles by employing a “separation technique.” Here just a few variables of the neural network must be changed in real-time. This is possible by their unique meta-learning technique, which pre-trains the neural net. Only these critical parameters need to be updated to capture the changing environment efficiently.
Autonomous drones with Neural-Fly learn to adapt to severe winds so well after only 12 minutes of flying data. This improves their performance dramatically. Compared to current state-of-the-art drones equipped with similar adaptive control algorithms, the error rate following that flight path is roughly 2.5 to 4 times lower.
At the Real Weather Wind Tunnel, test drones were tasked with flying in a figure-eight pattern with 27 mile per hour winds. This is defined as a “strong breeze” that makes it difficult to use an umbrella. It’s just less than a “strong gale,” making moving difficult and forcing entire trees to tremble. This wind speed is twice as fast as the drone encountered during neural network training. As a result, neural-Fly may easily extrapolate and generalize to unexpected and extreme weather situations.
As is prevalent in drone research and enthusiast groups, the drones were installed with a conventional off-the-shelf flight control computer. The Raspberry Pi 4 computer, which is roughly the size of a credit card, houses Neural-Fly. If you are interested in starting your project, these computers can be purchased for approximately $35 from the Raspberry website.