En, I've asked a big bull in the field who published similar stories. It's actuall just an interface on which users can train. It's the same as gamer controllers, but the input signals are fancier.
It is a quite established field for both neuroscience and engineering research now. The fundamental was paved by Apostolos P. Georgopoulos (U Minnesota, http://www.neurosci.umn.edu/faculty/georgopoulos.html) in their papers published in early 1980's about population coding of movement direction in Monkey's motor cortex. Another breakthrough was made by Miguel Nicolelis (Duke U, http://www.nicolelislab.net/) in early 2000's -- manipulate robotic arms/effectors using multi-unit recording in mice and monkeys' motor cortex, and then John Donoghue (Brown U, http://donoghue.neuro.brown.edu/) used it in human being for paralyzed patients (as being showed in the video posted below, well, not by me :-)). Above mentioned approaches are quite accurate but all invasive which everyone of them requires implantation of sensors in the brain, which drew many suspects of risk to patients (though I believe the paralyzed patients rather take this risk for benefits like freedom at home). Another concern of this line of approaches is the inactivation of the sensors due to tissue building up around the site along time (Someone proposed 'rejuvenating' methods such as brief high-voltage impulse or quick heating to remove the build-ups, etc). Another line of approaches are using EEG/MEG signals to control robotic arms or cursors on the screen, which is noninvasive and quite safe. There are many labs working on it since the relative low-cost and more convenient setups, as being showed in this link. A company has exhibited their 'brain-control game' -- people compete by moving the balls towards a target on a platform in sfn for at least three-years in a row. I don't know whether they will do it again this year in New Orleans, but it's quite fun to play and watch. Though I doubt whether it relies mainly on brain signals or EMG on the forehead. There are also people using other signature EEG/MEG signal, such as P300 on other applications of BMI (mind spelling of words, etc.). People have posted news and studies about those many times on this board. In my opinion, there are at least a couple of major challenges scientists and engineers are currently fighting on in this field: 1. How do we go beyond linear coding/decoding methods we are using and make more progress in more complicated movements (i.e. drawing spiral circles and playing instruments) that requires non-linear coding/decoding. There are many studies on this approach, such as Georgopoulos himself and many other groups (i.e. Gert Pfurtscheller in Austria) but I haven't seen a real breakthrough yet (well, there might be some progress in grasping, which I am not familiar with). 2. How do we get better performance by using the noisy EEG/MEG signal. For 2 -D encoding, it's pretty good now, but the performance still needs improvement for the 3-D coding.