How do the phones know that they are close? The basic method is simple enough the phones beep at each other and time how long it takes the pulse to arrive. They then swap data and both compute the distance that they are apart. The difficulties in getting it accurate and fast enough make this a big signal processing problem.
However they seem to have solved it well enough to get a 2cm accurate position 12 times a second — which is enough to play SwordFighting or ChaseCat (don't ask).
Is this a good idea?
It only needs standard phones and no extras and Microsoft Research think that it could be the basis of a whole new genre of games — Motion Mobile Games (MMGs). They don't seem to have learned the lesson of the flying WII controller however and is seems reasonable to predict lots of broken phones due to advanced sword play...