Personally, I've operated accelerometers and gyros at 200 Hz after some low-pass filtering, which is really all you need for this application. I'm sure there are sensors that operate at even higher frequencies if need be. The GPS module doesn't have to be updated too fast (maybe 10 Hz?) because it only serves to correct for horizontal drift.

I run the 100 Hz flight control loop for my tricopter on an ATmega1280 clocked at 16 MHz. A _lot_ more could be done with something like the Beagleboard, including generating a 3D map of the surroundings if I mounted a Kinect on it.

I want to agree about the decentralized operation because there would be so many wireless signals that could interfere, but those guys seem to be controlling 20 quadrotors without problems. I know XBees can be configured to communicate on different channels, but decentralized operation will most likely require something beefier like wifi (and that, I don't know how to configure to work over so many different channels).

Even in best case scenario, I'm not sure GPS would be accurate enough, as it's only accurate down to few centimeters, you'd probably need more if you'd like to perform that kind of acrobatics as shown in the video. And what about the height? You could of course use GPS, but apparently (http://users.erols.com/dlwilson/gpsvert.htm), average consumer devices have error of few meters. Maybe we could use laser range finders pointed at the ground, but what if the quadcopter tilts? Or what if another quadcopter flies underneath it? Of course it's not unsolvable, but it'd probably require more computing power in the quadcopter.

What do you think?

I think LIDAR is the most likely solution. With enough resolution, you could get a reasonable 3D map of the vehicle's surroundings with a very accurate relative model of where the vehicle is located.

The wireless becomes the issue then, as the network effect quickly takes over. Sensing at high speed with bad sensors and low power/computation is really tough.