Incentives make robots soar like eagles
Researchers report a bot that learns how to glide, giving clues to how birds do it. Andrew Masterson reports.
A type of punishment-and-reward approach is teaching robotic gliders how to fly like birds, a new report reveals.
Gliding birds – such as albatrosses, or eagles – maintain their flight by shifting constantly between upwellings of warm air, known as thermal plumes. These plumes form and decay constantly, often over timescales measured in mere minutes, and how birds manage to detect and shift between them is one of the great unanswered questions of ornithology.
A team of robotics scientists led by Gautam Reddy from the University of California, San Diego, however, have engineered an end-run around the problem by using an incentive-based approach to encourage a fixed-wing glider to move between air currents.
In a paper published in the journal Nature, Reddy and colleagues describe a glider with a two-metre wingspan. The machine was fitted with a flight controller that precisely controlled its bank angle and pitch. The aim was to adjust the two variables in such a way that the glider was able to exploit available thermals and stay aloft.
The machine as not remote-controlled. Instead, its success or failure was determined onboard using information gained over repeated flights. To maximise efficiency, and point the AI system in the right direction, the researchers used a robotics approach known as “reinforcement learning”.
First formulated in the late nineties, the approach works by using electronic rewards and punishments to emphasise the value of choices. It has previously been used very successfully in other landmark robotics projects – notably in creating an AI model that was able to learn and master the fiendishly difficult board game, Go.
Using this approach, Reddy and colleagues sent the glider aloft for several days of learning experiences. They induced it to accurately measure vertical wind accelerations – the fundamental property of thermals – as well as the forces acting to push the glider sideways, known as “roll-wise torque”.
Using these inputs as navigational cues, the glider was able to stay aloft and undertake test journeys.
The researchers suggest that their findings may indicate that birds also use similar cues to achieve gliding success.
Although more research needs to be conducted – in particular, in modelling and accounting for thermal turbulence – Reddy and colleagues conclude that their work points towards a near-term future in which AI flying craft will be able to make long, safe, journeys with minimal energy expenditure.