A specially designed algorithm allows drones to learn acrobatics through simulations.
A group of researchers from the University of Zurich have developed a new type of quadcopter capable of learning how to fly acrobatics on its own. The technology was developed in partnership with microprocessor company Intel with the goal of autonomous drones recognizing how to best use their agility and speed at the highest limit. This way, they can cover more distance within their battery life. While some argue that it’s not really necessary for a drone to know how to do fancy power loops or barrel rolls, having the capacity for more maneuvering is ultimately advantageous and more likely to be efficient.
The quadcopter drone operates on a specially designed navigation algorithm developed by the research team. Inside it is an artificial neural network that combines input from onboard cameras and sensors that is then communicated as control commands. This neural network is trained through acrobatic maneuver simulations the user wants performed. These are simulated through reference trajectories instead of the usual expensive ones demonstrated by human pilots. As a result, simulation trainings can also be easily scaled to include a diverse set of maneuvers without the risk of damaging the quadcopter—even the ones usually performed by the most skilled human pilots.
The team stresses that only a few hours of simulation training are sufficient before launching the quadcopter for whichever particular use. This does not require additional fine-tuning with real data. In particular, the algorithm makes use of abstractions produced by the sensory inputs during simulations, which are then used in real-life flight.The researchers demonstrated the quadcopter to test the efficiency of the algorithm. It was flown to complete complex maneuvers, such as a power loop, barrel roll and matty flip, among others. During the demonstration, the drone was subject to high thrust and extreme angular acceleration.
“Our algorithm learns how to perform acrobatic maneuvers that are challenging even for the best human pilots,” said Daivde Scaramuzza, a robotics professor and head of the robotics and perception group at the University of Zurich.
However, the team does acknowledge that human pilots still have a larger advantage. “Human pilots can quickly process unexpected situations and changes in the surroundings, and are faster to adjust,” Scaramuzza said. But he was also quick to add that this technology will be particularly useful in search and rescue scenarios or even for delivery services in which drones can cover longer distances much more quickly and efficiently.
The complete study is available here.
For more news and stories, check out how this study explores using swarms of tiny drones to transform search and rescue operations.