robot helicopter that learns

That's interesting.

I wonder what "watching" in that context actually means?

The article said the movement of the controls were recorded, but that alone wouldn't be enough. It would need input on what the helicopter was actually doing as well. That's consistent with what the article said when it talked about the software determining the optimal path the pilot was trying to achieve.

But how was that path collected? Did they simply have some sort of high resolution inertial measurement unit collecting the data?

The way the article was written, you half expected the autonomous helicopter to be flying around, watching (with a video "eye") what another helicopter was doing, and then copy it. But my best guess is that they got the expert pilot to fly the AI helicopter, and they recorded the commands, as well as how the helicopter responded, and then used that data off-line, as a training set for their code.

But, what did they do from there? Did they simply extract the entire path of the air-show flight, and program that into the AI code as the desired result, and then use some type of learning code to allow the AI system to figure out on it's own how to fly the helicopter to produce that path?

And what did they do with the recorded control movements of the human pilot?