Drones when they fly have a "small" problem and that is that the weather conditions affect them a lot, since in general they are small and light. In adverse conditions and with strong gusts of wind, its flight becomes very complicated.
When talking about the possibilities of drones and tests are carried out, they are always carried out in optimal weather conditions, and in many cases there is a controller who directs them, but the reality when they have to carry out their tasks can be very different.
That is why Caltech technicians have developed a system called Neural Fly using deep learning technologies in real time to teach drones to fly in adverse atmospheric conditions (strong and changing winds, etc…)
With this system, after 12 minutes of flight, the drone obtains data that allows it to learn from its flight experience and be able to respond to the effect of strong winds on it.
Tests have been carried out on drones flying in the Caltech wind tunnel, and applying all kinds of gusts of wind and even almost a gale, with positive results.
A video can be seen at: https://youtu.be/y3Z5ZJK6FDg
As I have commented many times, Artificial Intelligence in its different forms is introduced in all kinds of activities to improve the functioning of machines.
Drones are going to be a very commonly used tool for various activities, and we must be sure that they carry out their functions with the greatest possible precision and with the minimum of supervision.
Only then will they be truly efficient.
EN 
ES