Robotic gliders train themselves to soar on thermals
Researchers from the University of California San Diego have created robotic gliders that literally fly like an eagle. Using a form of machine intelligence called reinforcement learning the gliders learned through experience how to exploit updrafts of warm air to soar to altitudes of up to 2,300 ft (700 m).
In the past century, there have been incredible advances in aviation, but flying like the birds has proven surprisingly elusive. Worse than that, it's galling that the birds can do what we can't, yet none of them has ever spent a second in a classroom or cracked a book. When it comes to soaring, they are the perfect example of learning by doing.
Birds soar by using thermal plumes rising from hot spots on the Earth's surface, but how do they know to find them? How do they know how to use them? More important, how do you get a robotic glider to do the same thing? The UC team's answer was to program the AUV to think somewhat like a bird.
To soar, birds anticipate local conditions by sensing vertical wind accelerations and roll-wise torques. In other words, how the air moves about them. This allows them to develop a navigational strategy, so they know when to turn, bank, dip, and glide to take the best advantage of the updrafts.
What the team discovered was that surprisingly little learning was needed to turn their robotic glider from a neophyte to an expert. Derived from behavioral psychology, reinforcement learning lets the subject learn by means of feedback from doing things and then assessing the results. This lets the glider judge its own performance and develop a navigational strategy for responding to changes in wind conditions.
The tests, conducted in collaboration with the Salk Institute and the Abdus Salam International Center for Theoretical Physics in Trieste, Italy, used gliders with 2-meter (6-ft) wingspans and a flight controller to set them at a precise pitch and banking angle.
The onboard computer recorded its experiences over the course of several days, and the programming identified navigational cues by estimating the gliders' local vertical wind accelerations and roll-wise torques as well as vertical wind velocity gradients across the gliders' wings. This was then expanded to gliders with different wingspans.
"Our results highlight the role of vertical wind accelerations and roll-wise torques as viable biological mechanosensory cues for soaring birds, and provide a navigational strategy that is directly applicable to the development of autonomous soaring vehicles," says Terry Sejnowski, a member of the research team from the Salk Institute for Biological Studies and UC San Diego's Division of Biological Sciences.
The research was published in the journal Nature.
Source: UC San Diego