Attention, there is a ball pounding! Researchers have developed a super-fast perception system through which drones can even avoid targeted ball throws. With the help of so-called event cameras and sophisticated data processing, they were able to reduce the previous response times to a few milliseconds. The concept could help prevent collisions with birds or between drones, the developers say.
The future belongs to the buzzing flying robots, it is said: One day, drones will deliver goods, transport people or be used in tricky exploration and rescue operations. Technological progress in recent years has made autonomous aircraft faster and more agile – but one aspect of their performance has not developed accordingly: their vigilance. Especially when drones are moving fast, they can hardly avoid obstacles that suddenly appear.
Still too sluggish
Previous camera and data processing systems require 20 to 40 milliseconds to initiate a flight adjustment. That sounds fixed – but it is not enough for real spontaneity: To avoid a bird or another drone, a significantly shorter reaction time is necessary. “Fast-navigating drones have a particularly high risk of colliding with obstacles, especially if they move,” says Davide Scaramuzza from the University of Zurich. He and his colleagues are therefore devoted to developing concepts that can make drones more responsive.
They used a fairly recent invention for their project: the so-called event camera. Their performance differs significantly from conventional systems. Video cameras, such as in smartphones, continuously take snapshots of an entire scene by simultaneously exposing all pixels. In order to capture a moving object, all pixel information must be analyzed. Event cameras, on the other hand, are more selective: They evaluate the pixels whose light intensity changes in a targeted and almost instantaneous manner. An overall image acquisition is not necessary. This enables significantly faster information processing, the scientists explain.
However, the researchers had to overcome a challenge when using event cameras for drones: In order to correctly record the movement of obstacles, the aircraft’s own movement must be factored in. The researchers developed an algorithm for this. In simple terms, the calculation method gives the system the ability to subtract movements recorded by the camera from the effects of the drone’s own movement. The difference signal describes the moving foreign object. Evasion maneuvers can then be triggered on the basis of this information.
Dodgeball with drone
The researchers demonstrated what their system can do with “drone dodgeball”: They equipped a quadrocopter with special cameras and algorithms and tested the drone’s ability to avoid targeted ball throws. As they report, the flying object was able to respond successfully in over 90 percent of the cases – even if a ball came flying from a distance of only three meters at ten meters per second. The perception system reacts in up to 3.5 milliseconds. It is about ten times faster than conventional concepts, the scientists sum up.
As they explain, the concept could significantly expand the uses of drones. In addition, other robot systems could also benefit: “The ability of robots to reliably recognize obstacles that are coming up to them also plays a crucial role in other areas, such as the automotive industry, mining and remote inspection with robots,” says Scaramuzza.
However, he and his colleagues now want to focus on the potential in drone technology: They are currently exploring the extent to which the new concept can further increase the reaction speed of drones. To do this, they test quadrocopters that are capable of even faster flight maneuvers, so that they may not actually be able to catch a throw in drone volleyball.
Source: University of Zurich, technical article, Science Robotics, doi: 10.1126 / scirobotics.aaz9712