Researchers are training autonomous flying drones to reach high speeds—up to 25 mph—with agility to fly through dense forests and collapsed buildings by using a fast perception system that can process changing light conditions and sensor noise.
The University of Zurich and Intel Labs published their research in Science Robotics Journal on Wednesday describing how they enabled quadrotors to fly autonomously through unknown and complex environments at high speeds using on-board sensing and computation. The planner in the autonomous system must find a path without crashing based on partial observations of the environment, which can be a dense forest or a partially collapsed building in a disaster zone.
Researchers trained a drone only in simulation by imitating how an expert is able to perform in challenging situations that weren’t introduced during the training of the convolutional network. A short video (above and below ) from Science Robotics shows a quadrotor zipping past trees and avoiding crashes with a segment describing the simulation used:
By only relying on on-board cameras and on-board computation, the drone was able to demonstrate a decrease in the latency between perception and action while also withstanding motion blur, missing data and sensor noise, the researchers said. The drone being used relied on a neural network that learned to fly by watching a simulated expert, an algorithm that flew a computer-generated drone through a simulated environment full of obstacles.
Existing systems have relied on using sensor data to create a map of the environment and the planning trajectories within the map, but those steps take time and mean it is impossible to fly at high speeds.
Applications of the technology could be construction sites, agriculture, search and rescue and more. A project web site describes more about the research.