None of the hardware construction looks all that expensive. (I've watched the video a few times, tried to catch a few details.) They're pretty high-speed motors, obviously, and otherwise lightweight construction.

The real trick is how they're getting them to calculate vectors. They're certainly not just remote controlled -- not unless they had an awful lot of takes of that 3" clearance shot -- so that raises the question, are they precalculating the vectors, or are they using some really advanced method to calculate them on-the-fly (so to speak)?

If they're precalculating, then the impressive thing here is the method they've developed to get these things to do tricks with rotors. If they're doing on-the-fly calculation, then I really want to know what kind of sensor they're using. I don't know of anything available on the hobbyist market with that kind of resolution and speed.

EDIT: Well, this is embarrassing. It spells out the answers to my questions right in the linked article. :-( Hooray for not paying attention.

$100 in parts, according to this guy: http://vicacopter.com

I tried (unsuccessfully) to make his design a year ago and it was more like $300 though, and that assuming you have the RF equipment.

Guy works at NASA I think. The claim that his is the only fully open source autopilot is to my knowledge true. No docs or schematics though, but these could be adapted from other projects, http://mikrokopter.de http://diydrones.com

It sounds like the control system is the expensive part. It uses a Vicon motion capture system to track the quadrotor using 20 cameras arranged around the room, plus custom control software running on a laptop, and an XBee wireless link from the laptop to the quadrotor to send the control signals.

The interesting part to me is the overshoot you see when the quadrotor shoots through the window, then pulls back nearer into a resting position. This implies that the control software can work out a dynamic trajectory (one where the quadrotor wouldn't be able to hover at every point along the trajectory), then reconcile it with a desired end point that won't fit onto the trajectory.

That sounds simple, but many walking robots still can't do that -- their gaits are designed so they could freeze in place at any time and still not fall over.

I work with the exact same setup. You're looking at 100-250k for the "Vicon" (motion capture) setup and 4k+ for the quad rotor (an Ascending Technologies Hummingbird).

That sounds like a lot, but the motion capture system can accurately track things with sub-millimeter precision at, at least, 100fps. The Hummingbird is very sophisticated, with a 3-axis accelerometer, 3-axis magnometer, 3-axis gyro and high performance brushless motors, and of course, the R&D cost for developing all of it.

DIYDrones has a lot of good info on hacking these together, but you won't get the performance you're seeing in that movie.

I've talked to people who are trying to build a poor man's version of these and they're aiming for 600-800, but it's not going to be as high performance.