Inside Unmanned Systems

APR-MAY 2018

Inside Unmanned Systems provides actionable business intelligence to decision-makers and influencers operating within the global UAS community. Features include analysis of key technologies, policy/regulatory developments and new product design.

Issue link:

Contents of this Issue


Page 59 of 67

AIR SOFTWARE 60   April/May 2018 unmanned systems inside incorporate other pieces of information, such as uncertainty in the data from the drone's depth sensors, Florence said. He and his col- leagues will present their latest findings in May at the IEEE International Conference on Robotics and Automation in Brisbane, Australia. MIMICKING CARS AND BICYCLES Drones f lying through cities have more than trees to contend with. There are also pedestri- ans, cyclists and automobiles that can all move unpredictably. Nevertheless, people routinely navigate such dynamic environments without the aid of complex, expensive devices. "When we drive a car or ride a bicycle, we are al- ready pretty good at this task, and we don't require any costly sensor, but just rely on what we see with our eyes," Loquercio said. "From there, we had the idea of imitating cars and bicycles to let a drone f ly through cities." Loquercio and his colleagues developed an artificial intelligence system known as DroNet, which is a kind of neural network. In such a system, components dubbed neurons are fed data and cooperate to solve a problem, such as recognizing an obstacle. The neural network then repeatedly adjusts the connec- tions between its neurons and sees if the new patterns of connections are better at solving the problem. Over time, the neural net discov- ers which patterns are best at computing solu- tions and adopts them as their defaults, mim- icking the process of learning in the human brain. The Swiss researchers previously used neural networks to help drones autonomously recognize and follow forest trails to help them find injured or lost hikers. In their latest work, the scientists trained DroNet on images collected by cars and bi- cycles traveling in urban environments. The aim was to have DroNet analyze video col- lected from a drone's forward-facing camera and predict where to steer the drone "and, if something dangerous is happening, whether to stop," Loquercio said. Specifically, DroNet learned what angles to steer at to avoid obstacles by analyzing more than 70,000 images captured from cars. The data was gathered by Udacity, a firm working on the world's first open-source self-driving car—a project since spun off into the com- pany Voyage. DroNet also learned how to estimate the probability of a collision by ana- lyzing roughly 32,000 images the scientists collected by mounting a GoPro camera on the handlebars of a bicycle and riding to different parts of Zurich. The researchers tested DroNet using a commercia l Pa r rot Bebop 2 quad-rotor drone. The system ran on an Intel Core i7 2.6-gigahertz CPU, receiving pictures from and sending commands to the drone through WiFi. The artif icial intelligence relied on greyscale images only 200 by 200 pixels large from the drone. By imitating cars and bicycles, DroNet au- tomatically learned to respect traffic rules, such as how to move down a street without drifting off into the oncoming lane, or how to stop when obstacles such as pedestrians, other Photo courtesy of University of Zurich. By imitating cars and bycicles, the drone automatically learned to respect the safety rules. "WORKING ON ROBOTS RACING THROUGH FORESTS IS JUST INTRINSICALLY SUPER FUN TO ME." Peter Florence, lead researcher, NanoMap

Articles in this issue

Archives of this issue

view archives of Inside Unmanned Systems - APR-MAY 2018