AI-trained vehicles can adapt to extreme turbulence on the fly

In the wild, flying animals sense impending changes in their environment, including the onset of sudden turbulence, and quickly adapt to stay safe. Engineers who design airplanes would like to give their vehicles the same ability to predict incoming disturbances and respond to them appropriately.

Indeed, disasters such as the fatal Singapore Airlines flight last May, in which more than 100 passengers were injured after the plane encountered severe turbulence, could be avoided if the plane had such capabilities. automatic detection and prediction combined with vehicle stabilization mechanisms.

Now, a team of researchers from the Center for Autonomous Systems and Technologies (CAST) at Caltech and Nvidia has taken an important step toward such capabilities. In a new article in the journal npj roboticsThe team describes a control strategy it developed for unmanned aerial vehicles, or UAVs, called FALCON (Fourier Adaptive Learning and CONtrol).

The strategy uses reinforcement learning, a form of artificial intelligence, to adaptively learn how turbulent winds can change over time, then uses that knowledge to control a drone based on what it experiences over time. real.

“Spontaneous turbulence has major consequences on everything from civilian flights to drones. With climate change, extreme weather events that cause this type of turbulence are on the rise,” says Mory Gharib, Professor Hans W. Liepmann of Aeronautics and Medical Engineering, Booth-Kresa Leadership Chair at CAST and author of the new paper.

“Extreme turbulence also appears at the interface between two different shear flows, for example when high-speed winds encounter stagnation around a tall building. Therefore, drones in urban environments must be able to compensate for such sudden changes. FALCON provides these vehicles with a way to understand upcoming turbulence and make necessary adjustments. »

FALCON is not the first drone control strategy to use reinforcement learning. However, previous strategies have not attempted to learn the underlying model that truly represents how turbulent winds work. Instead, they have all been model-free methods. Such methods focus on maximizing a reward function that cannot be used to address different parameters, such as different wind conditions or vehicle configurations, without retraining, because they focus on a single environment.

“It’s not so good in the physical world, where we know that situations can change dramatically and quickly,” says Anima Anandkumar, Bren Professor of Computer Science and Mathematical Sciences at Caltech and author of the new paper. “We need the AI ​​to learn the underlying pattern of turbulence well so it can act based on how it thinks the wind is going to change.”

“Advancements in fundamental AI will change the face of the aviation industry, improving safety, efficiency and performance across a range of platforms, including airliners, drones and transport aircraft. These innovations promise to make air transport and operations smarter and safer. and more streamlined,” says Nvidia co-author Kamyar Azizzadenesheli.

As the acronym FALCON indicates, the strategy is based on Fourier methods, meaning it relies on the use of sinusoids, or periodic waves, to represent signals, in this case wind conditions. Waves provide a good approximation of standard wind movements, thus reducing the necessary calculations to a minimum. Within these waves, when extreme turbulence appears, the instability manifests itself by a notable change in frequency.

“If you can learn to predict these frequencies, then our method can give you an idea of ​​what to expect,” says Gharib, who is also director of Caltech’s Graduate Aerospace Laboratories.

“Fourier methods work well here because turbulent waves are better modeled in terms of frequencies, with most of their energy in the low frequencies,” says co-lead author Sahin Lale, now a senior research engineer at Neural Propulsion Systems, Inc. who completed the work at Caltech. “Using this prior knowledge simplifies both learning and controlling turbulent dynamics, even with a limited amount of information.”

To test the effectiveness of the FALCON strategy, the researchers created an extremely difficult test setup in the John W. Lucas Wind Tunnel at Caltech. They used a fully equipped airfoil system as a representative drone, equipping it with pressure sensors and control surfaces capable of making online adjustments to things like the system’s altitude and yaw. They then positioned a large cylinder with a movable attachment in the wind tunnel. As the wind passed over the cylinder, it created random and large fluctuations in the wind hitting the airfoil.

“Training a reinforcement learning algorithm in a turbulent physical environment presents all sorts of unique challenges,” says Peter I. Renn, co-lead author of the paper and now a quantitative strategist at Virtu Financial. “We couldn’t rely on perfectly clear signals or simplified flow simulations, and everything had to be done in real time.”

After approximately nine minutes of training, the FALCON-assisted system was able to stabilize in this extreme environment.

“With each new observation, the program improves because it contains more information,” explains Anandkumar.

“The future really depends on the power of the software, because it will require less and less training,” says Gharib. “Rapid adaptation will be the challenge, and we will push, push, push.”

Looking to the future, he adds that researchers are considering giving drones and even passenger planes the ability to share sensed and learned information about conditions with each other. Such sharing of sensor readings and artificial intelligence-based learning between aircraft, especially near a disturbance, could help keep aircraft safe.

“I believe it’s going to happen,” Gharib says. “Otherwise, things will become very dangerous as extreme weather events increase in frequency.”