The two monkeys’ training programme consisted of one-dimensional left-right-movements of the arm like they are needed for displacing an object from one position to another. The other task was a three-dimensional grabbing exercise, in which food was presented to the monkeys at random positions.

While performing their training, the electrical activity of the nerve cells was continously recorded. This information was then used to test mathematical algorithms, either linear algorithms or artificial neural networks in order to find out which are suitable for the prediction of the monkeys’ arm movements in real time. What worked best in the end was an adaptable system which was able to change the algorithms during a training session.

Prediction is the key thing in the whole BCI business, because when you can predict a movement in real time, half a second or so before the monkey actually moves his arm, then you should also be able to drive a robot arm.

When it comes to possible therapeutic value for paralysed patients, though, there is a slight flaw in experiments like these: You obviously need to be able to do the training in order to “trim” the software. It is still an open question if – in case of a totally paralysed person - “remembering” motor activity is enough “training” of that sort.

What was encouraging about Wessberg’s experiments, though, was the fact that his software was also able to predict movements that had not been trained specifically. So when the training of linear hand movement was only performed for targets on the left, the software was nevertheless able to predict arm movement to right, as well.