Oscillatory Neural Networks for AI Applications in Drones

Artificial Intelligence (AI) in drones has been utilized to perform autonomous operations, including navigation, search and rescue. However, these drones and similar edge devices are power-limited, and traditional deep neural networks require high processing power to process large amounts of data. Oscillatory Neural Networks (ONNs) are an alternative physics-based computing paradigm that enables compact, fast, and low-power computation on the edge via digital implementations. In this paper, a drone fitted with 14 laser-based ranging sensors and an ONN on an FPGA, capable of perceiving the environment, performing pattern retrieval, taking off, landing and avoiding obstacles is presented. An ONN structure comprising two cascaded ONNs for horizontal (lateral) obstacle avoidance and one ONN for vertical (altitude) obstacle avoidance is presented. A response time of 39 ms for the implemented drone is reported, while the average computation time of the ONNs was <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$139 \mu ~\mathrm{s}$</tex>. The power consumed by the ONNs was less than 4% compared to the power consumed by the drone during flight.