Researchers at the University of Zurich have built an autonomous flying device that is sure to change the way we think about drones. An autonomous drone from researchers at the University of Zurich zips through the trees of a forest at high speeds with the ease of a large dragonfly dodging reeds. The researchers’ four-rotor machine has the distinction of being able to cope autonomously in any unfamiliar and complex environment. It uses only the data collected in real time by its on-board sensors to find its way around.
As well as being agile, it is also very fast, capable of travelling at speeds of up to 40 km/h in any environment, even through trees. This makes it capable of detecting any complex environment, from buildings to caves, and can be used for emergency operations.
If the researchers’ prototype proves successful, it will mark a major step forward for the drone industry. Currently, controlling drones is much more complicated. Either a map is loaded into the drone and it follows the path it is given (with a collision avoidance mechanism, of course), or drone pilots are deployed to control the flying devices from a distance in unfamiliar environments. But even the latter face serious difficulties in interpreting the environment in a fraction of a second to guide the drone in the right direction to avoid, for example, a collision. Experienced pilots can only do this after years of training,” a member of the research team is quoted as saying in a report published on the university’s website.
A special simulation method was used to teach the drone. They taught the quadcopter’s neural network by flying a computer-generated drone through a simulated environment full of complex obstacles. The algorithm had all possible information about the drone in real time to calculate the best trajectory.
From these simulated flights, the neural network learned how to predict the optimal trajectory based solely on sensor data. This allows the plane to react much faster to unknown objects than if the trajectory were determined in the traditional two-step process. Most people are experimenting with autonomous systems that first collect environmental parameters from the sensors, then use the data to create a map, and then use the map to plan the best route. Although this method is fast on a human scale, it does not allow high-speed flight.
The Zurich team simply recognised that an exact replica of the real environment is not necessary for a neural network to learn to interpret a complex environment. One of the team’s researchers was even more radical: if the right design approach is taken, even simple simulators are sufficient for learning.
After teaching in simulation, the system also excelled in the real world. It was able to maintain flight speeds of up to 40 km/h without crashing in a wide range of environments. What’s more, it only took a few hours or days to master this skill compared to human drone pilots.
The new autonomous control mode is not limited to drones. This approach could be a major help in the development of any autonomous vehicle. It could also lead researchers to new ways of training artificial intelligence systems. For example, Zurich’s method could be used to train a spacecraft to navigate with a high degree of autonomy on the surface of an alien planet about which we have little prior information.
The team wants to deepen the drone’s knowledge and speed up its sensors. Indeed, if they can collect more data about their environment in a shorter time, the drone’s speed could be increased to over 40 km/h.