The UZH pilot was flow against two state-of-the-art drone piloting programs. At low speeds performance was similar, but at higher speeds the other systems suffered from too much ‘latency’ – they were not able to calculate a safe route fast enough – and experienced multiple crashes. The previous software used separate modules for sensing, mapping and planning, whereas the machine-learning approach fuses all three functions and provides swifter routing.

The latency of the two other systems was 65 and 19 milliseconds for a sample task, the UZH system managed 10 milliseconds. While this may not be consistent for every piloting challenge, it shows the potential of the approach.

Scaramuzza notes that this is the first time such a system has been trained in a completely simulated environment and then used its learning in the physical world. While it could be useful for training delivery drones and air taxis to negotiate the urban landscape in all weathers and lightning conditions – cheaply and without the risk of accidents — there may also be far greater potential.

The technique suggests that guided machine learning in virtual environments could be a way of quickly and cheaply training all sorts of robotic systems. For example, a robotic system for unloading boxes from trucks could get the equivalent of years of experiencing with learning how to handle awkward shapes and sizes of package. Domestic cleaning robots, which sometimes struggle with odd-shaped rooms or particular items of furniture, could get smarter with virtual training.

For the immediate future, this approach is likely to be applied mainly to small drones, enabling them to carry out aggressive maneuvers faster than any human pilot. The paper indicates that even faster speeds could be achieved with better sensors and more accurate modeling; it is not clear what the absolute limits might be.

You can watch a two-minute video showing how 'Learning high-speed flight in the wild' was achieved here.