MSGID: 1:317/3 6274a4ac   
   PID: hpt/lnx 1.9.0-cur 2019-01-08   
   TID: hpt/lnx 1.9.0-cur 2019-01-08   
    Rapid adaptation of deep learning teaches drones to survive any weather   
      
      
     Date:   
         May 5, 2022   
     Source:   
         California Institute of Technology   
     Summary:   
         Neural-Fly technology could one day build the future of package   
         delivery drones and flying cars.   
      
      
      
   FULL STORY   
   ==========================================================================   
   To be truly useful, drones -- that is, autonomous flying vehicles --   
   will need to learn to navigate real-world weather and wind conditions.   
      
      
   ==========================================================================   
   Right now, drones are either flown under controlled conditions, with   
   no wind, or are operated by humans using remote controls. Drones have   
   been taught to fly in formation in the open skies, but those flights   
   are usually conducted under ideal conditions and circumstances.   
      
   However, for drones to autonomously perform necessary but quotidian tasks,   
   such as delivering packages or airlifting injured drivers from a traffic   
   accident, drones must be able to adapt to wind conditions in real time --   
   rolling with the punches, meteorologically speaking.   
      
   To face this challenge, a team of engineers from Caltech has developed   
   Neural- Fly, a deep-learning method that can help drones cope with   
   new and unknown wind conditions in real time just by updating a few   
   key parameters.   
      
   Neural-Fly is described in a study published on May 4 in Science   
   Robotics. The corresponding author is Soon-Jo Chung, Bren Professor of   
   Aerospace and Control and Dynamical Systems and Jet Propulsion Laboratory   
   Research Scientist. Caltech graduate students Michael O'Connell (MS '18)   
   and Guanya Shi are the co-first authors.   
      
   Neural-Fly was tested at Caltech's Center for Autonomous Systems   
   and Technologies (CAST) using 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.   
      
      
      
   ==========================================================================   
   "The issue is that the direct and specific effect of various wind   
   conditions on aircraft dynamics, performance, and stability cannot   
   be accurately characterized as a simple mathematical model," Chung   
   says. "Rather than try to qualify and quantify each and every effect   
   of turbulent and unpredictable wind conditions we often experience   
   in air travel, we instead employ a combined approach of deep learning   
   and adaptive control that allows the aircraft to learn from previous   
   experiences and adapt to new conditions on the fly with stability and   
   robustness guarantees."  O'Connell adds: "We have many different models   
   derived from fluid mechanics, but achieving the right model fidelity and   
   tuning that model for each vehicle, wind condition, and operating mode is   
   challenging. On the other hand, existing machine learning methods require   
   huge amounts of data to train yet do not match state-of-the-art flight   
   performance achieved using classical physics-based methods. Moreover,   
   adapting an entire deep neural network in real time is a huge, if not   
   currently impossible task."  Neural-Fly, the researchers say, gets around   
   these challenges by using a so- called separation strategy, through which   
   only a few parameters of the neural network must be updated in real time.   
      
   "This is achieved with our new meta-learning algorithm, which pre-trains   
   the neural network so that only these key parameters need to be updated   
   to effectively capture the changing environment," Shi says.   
      
   After obtaining as little as 12 minutes of flying data, autonomous   
   quadrotor drones equipped with Neural-Fly learn how to respond to   
   strong winds so well that their performance significantly improved (as   
   measured by their ability to precisely follow a flight path). The error   
   rate following that flight path is around 2.5 times to 4 times smaller   
   compared to the current state of the art drones equipped with similar   
   adaptive control algorithms that identify and respond to aerodynamic   
   effects but without deep neural networks.   
      
   Neural-Fly, which was developed in collaboration with Caltech's   
   Yisong Yue, Professor of Computing and Mathematical Sciences,   
   and Anima Anandkumar, Bren Professor of Computing and Mathematical   
   Sciences, is based on earlier systems known as Neural-Lander and   
   Neural-Swarm. Neural-Lander also used a deep- learning method to track   
   the position and speed of the drone as it landed and modify its landing   
   trajectory and rotor speed to compensate for the rotors' backwash from   
   the ground and achieve the smoothest possible landing; Neural- Swarm   
   taught drones to fly autonomously in close proximity to each other.   
      
   Though landing might seem more complex than flying, Neural-Fly, unlike   
   the earlier systems, can learn in real time. As such, it can respond to   
   changes in wind on the fly, and it does not require tweaking after the   
   fact. Neural-Fly performed as well in flight tests conducted outside the   
   CAST facility as it did in the wind tunnel. Further, the team has shown   
   that flight data gathered by an individual drone can be transferred to   
   another drone, building a pool of knowledge for autonomous vehicles.   
      
   At the CAST Real Weather Wind Tunnel, test drones were tasked with   
   flying in a pre-described figure-eight pattern while they were blasted   
   with winds up to 12.1 meters per second -- roughly 27 miles per hour,   
   or a six on the Beaufort scale of wind speeds. This is classified as a   
   "strong breeze" in which it would be difficult to use an umbrella. It   
   ranks just below a "moderate gale," in which it would be difficult to   
   move and whole trees would be swaying. This wind speed is twice as fast   
   as the speeds encountered by the drone during neural network training,   
   which suggests Neural-Fly could extrapolate and generalize well to unseen   
   and harsher weather.   
      
   The drones were equipped with a standard, off-the-shelf flight control   
   computer that is commonly used by the drone research and hobbyist   
   community. Neural-Fly was implemented in an onboard Raspberry Pi 4   
   computer that is the size of a credit card and retails for around $20.   
      
      
   ==========================================================================   
   Story Source: Materials provided by   
   California_Institute_of_Technology. Original written by Robert   
   Perkins. Note: Content may be edited for style and length.   
      
      
   ==========================================================================   
   Journal Reference:   
      1. Michael O'Connell, Guanya Shi, Xichen Shi, Kamyar Azizzadenesheli,   
      Anima   
         Anandkumar, Yisong Yue, Soon-Jo Chung. Neural-Fly enables rapid   
         learning for agile flight in strong winds. Science Robotics, 2022;   
         7 (66) DOI: 10.1126/scirobotics.abm6597   
   ==========================================================================   
      
   Link to news story:   
   https://www.sciencedaily.com/releases/2022/05/220505085644.htm   
      
   --- up 9 weeks, 3 days, 10 hours, 50 minutes   
    * Origin: -=> Castle Rock BBS <=- Now Husky HPT Powered! (1:317/3)   
   SEEN-BY: 15/0 106/201 114/705 123/120 129/330 331 153/7715 218/700   
   SEEN-BY: 229/110 111 317 400 426 428 470 664 700 292/854 298/25 305/3   
   SEEN-BY: 317/3 320/219 396/45   
   PATH: 317/3 229/426