I think that's worded a bit wrong. Can detect a user "up to 3 feet from the sensor"... Should probably be something like "from as close as 3 feet away from the sensor".
Anyway, the original Kinnect is designed to work with users from 1.2m to 3.5m (~4ft - 11ft).
I think there's been some attempts at using Kinect (or other motion tracking devices) with real human models - or dolls, if you must - to collect animation data and filter it down to something that can be fed into a video game pipeline, for instance see here but apparently it's still too noisy to be really usable for anything beyond very low-detail animations, sadly.
Yeah, there's a motion builder plugin for Kinnect capture that's in wide use for very cheap mocap
It works pretty well!
BTW, all mocap is too noisy. Mocap data is always used as a starting point for an animator, and always has to get cleaned up before being used in a game. Yeah, the kinnect data will be more noisy than data from a $50,000 mocap rig, but noisy data isn't just a kinnect-specific problem ;)
The real disadvantages of kinnect compared to larger (more expensive) systems is that the performance space is much smaller, the number of bones captured is smaller (e.g. no finger movements), it only captures from one angle (movement of a hand behind your back can't be captured), and can't capture more than a couple of actors at once.