A team of computer scientists from ETH Zurich, Switzerland, and the University of Bologna, Italy, have created what is thought to be the world’s smallest fully autonomous drone – just 4 x 3in – capable of flying without a remote pilot.
The team built an autonomous system on top of the 27-gram Crazyflie 2.0 Nano Quadcopter.
The deployment of fully autonomous flying robots is being held back on a number of fronts. For one, until there are computer vision systems that inspire confidence in the minds of regulators, kinetic energy will continue to be a factor.
In other words, size, weight, and power matter when it comes to flying safely in the vicinity of human beings, buildings, and other vehicles, which is why the notion of tiny, intelligent drones could be appealing.
But navigating around dynamic urban environments presents greater challenges to pilotless systems. The ability to deploy the computer vision required to do that demands plenty of processing power, which adds to drones’ weight.
The research team’s solution is the first drone to use a low-power, vertically integrated system for fully autonomous, deep-neural-network-based navigation. The system only consumes around 94 milliWatts of energy, which amounts to just one percent of the drone’s total power.
The platform is able to run the research team’s DroNet lightweight convolutional neural network. DroNet was trained using urban footage taken from cyclists and cars, and allows the drone to choose movement and direction by predicting the correct steering angle and accounting for collision probability.
The drone gathers enough information about its environment by processing incoming video at 20 frames per second. But because it was trained with images from fixed positions on cars and bicycles, it’s not yet capable of safely determining the correct altitude.
The result is a system that can get to where it needs to go and perform basic tasks – just not in the most elegant way possible.