Trained drones whiz into the unknown

Trained drones whiz into the unknown

An autonomous drone flies through the forest at 40 kilometers per hour. (Image: University of Zurich)

So far, caution has slowed autonomous drones considerably – but now researchers are cleverly ensuring acceleration: They have developed a system based on artificial intelligence that enables quadrocopters to fly safely at up to 40 kilometers per hour through unknown environments full of obstacles. A neural network calculates the optimal flight maneuvers from sensor data, which was previously trained using simulations. The new capabilities could significantly improve the use of drones in the exploration of complex environments, say the developers.

They are small, agile and can be equipped with all kinds of ingenious technology: Drones are ideal scouts that can also be used in confusing surroundings such as forests, caves or rubble structures. But without prior information about the location of obstacles, the navigation capabilities of the aircraft leave much to be desired. Above all, in order to exploit their speed potential during operations, they currently still have to be remote-controlled by experienced human pilots.

“When maneuvering a drone quickly, the environment has to be captured in fractions of a second in order to steer it onto collision-free paths. This is very difficult for both people and machines, ”says Davide Scaramuzza from the University of Zurich. “Experienced pilots can reach this level after a long training session. But machines still have a hard time doing that ”. That is why Scaramuzza and his team have now devoted themselves to the task of increasing the response speed of drones during autonomous navigation.

No more lame mapping necessary

As they explain, the previous AI systems were based on the processing of sensor data into maps of the surroundings, which the drone then uses for orientation. However, this process is time-consuming and thus prevented movement at high speeds. The researchers’ new concept, however, is now based on training the drone, which enables it to react spontaneously to obstacles with confident evasive maneuvers. The “brain” of the device, which enables this teaching ability and the subsequent data processing, is a so-called neural network.

This form of artificial intelligence is trained by a simulation system: the drone is controlled through a virtual environment full of complex obstacles. The simulation system constantly records the position of the quadrocopter in the computer world and generates measured values ​​for the virtual obstacles. This information is then converted into trajectories in fractions of a second, which guide the aircraft optimally around various obstacles. In the course of the training, the “brain” of the drone acquires confident and fast navigation skills that can be transferred to the real world, the scientists explain.

Clever flight maneuvers

They were able to document this through subsequent tests: They let a trained drone whiz through unknown surroundings such as forests, buildings, ruins or trains. Without colliding with trees, walls or other obstacles, the quadrocopter reached speeds of up to 40 kilometers per hour, the scientists report. “While people sometimes need years for training, artificial intelligence can achieve comparable navigation skills much faster, almost overnight, with the help of high-performance simulators,” says first author Antonio Loquercio from the University of Zurich.

The scientists are now working on increasing the performance of the system even further: They want to develop sensors that can provide more information about the surroundings in a shorter period of time so that the drone can reach speeds of over 40 kilometers per hour. According to the team, the system could be used in the future to make drones more powerful for different applications. Because, for example, when deployed in disaster areas, their speed can save lives. As the developers finally emphasize, the system also has potential that goes beyond its use in drones: The concept could also be useful for other AI-based systems that require fast reaction times.

Source: University of Zurich, specialist article: Science Robotics, doi: 10.1126 / scirobotics.abg5810

Recent Articles

Related Stories