Thailand’s DEPA seeks to meet industry’s digital technology needs – OpenGov Asia
Researchers at the California Institute of Technology (Caltech) have discovered that a deep learning technological beacon, known as Neural-Fly, could help flying robots called “drones” adapt to all weather conditions .
Drones are now flown in controlled conditions, with no wind, or by people using software or remote controls. Flying robots have been trained to take off in formation in the open air, although such flights are usually conducted in perfect conditions.
However, for drones to autonomously perform important but mundane tasks, such as delivering packages or airlifting injured drivers in traffic accidents, they must be able to adapt to windy conditions. in real time.
With this, a team of Caltech engineers created Neural-Fly, a deep learning technology that allows drones to adapt in real time to new and unexpected wind conditions by simply adjusting a few key parameters. Neural-Fly is discussed in a recently published research titled “Neural-Fly Enables Rapid Learning for Agile Flight in Strong Winds” in Science Robotics.
The problem is that the direct and specific effect of various wind conditions on aircraft dynamics, performance and stability cannot be accurately characterized by a simple mathematical model.
– Soon-Jo Chung, Bren Professor of Aerospace and Control and Dynamic Systems and Research Scientist of the Jet Propulsion Laboratory
Chung added that they use a combined approach of deep learning and adaptive control that allows the aircraft to learn from past experiences and adapt to new conditions on the fly, with guarantees of stability and robustness, instead of trying to qualify and quantify every effect of the turbulent and unpredictable wind conditions they frequently encounter in flight.
Neural-Fly was evaluated at Caltech’s Center for Autonomous Systems and Technologies (CAST) using its Real-Time Wind Tunnel, a 10-foot-by-10-foot array of more than 1,200 small computer-controlled fans that allows engineers to mimic everything from a light breeze to a gust of wind.
Many models derived from fluid mechanics are available to researchers, but it is difficult to get the right model quality and to fine-tune that model for every vehicle, wind condition, and mode of operation.
Existing machine learning methods, on the other hand, require massive amounts of data for training, but cannot match the flight performance achieved by conventional physics-based methods. Adapting a complete deep neural network in real time is a monumental, if not impossible, undertaking.
According to the researchers, Neural-Fly addresses these challenges using a technique known as splitting, which requires only a few neural network parameters to be modified in real time. This is accomplished using their innovative meta-learning technique, which pre-trains the neural network so that only those critical parameters need to be changed in order to successfully capture the changing environment.
After just 12 minutes of flight data, Neural-Fly-equipped autonomous quadcopter drones learn to respond to high winds so well that their performance improves dramatically judging by their ability to accurately follow a flight route.
Compared to drones equipped with state-of-the-art adaptive control algorithms that identify and react to aerodynamic effects but lack deep neural networks, the error rate following this flight path is between 2.5 and 4 times lower.
Landing may seem harder than flying, however, Neural-Fly can learn in real time, unlike previous systems. As a result, it can react on the fly to wind variations and does not require post-processing.
The flight tests were conducted outside of the CAST facility; Neural-Fly performed as well as in the wind tunnel. Additionally, researchers have shown that flight data collected by one drone can be transferred to another, creating a pool of knowledge for self-driving cars.
The drones were equipped with a typical, standard flight control computer used by the drone research and enthusiast communities. Neural-Fly has been integrated into a Raspberry Pi 4 embedded computer, which is the size of a credit card and costs around $20.