manipulation experiment with NN in virtual environment

[dugu]

Hi all.
My name is Dugu, I am pretty new around here.
I joined this forum because I have a big problem. I need to build an experiment in an virtual environment, using a robot manipulator (which I already designed in CAD, and exported it into OGRE) and ANN's (artificial neural networks). I am trying to build an experiment similar to the one you can see in this video:

www.youtube.com/watch?v=xRq76EATroM&feature=related

I have several questions which I kindly post to you:
1. What task should be handled by the ANN? Is it avoind the obstacle, or grasping, or maybe the path planning of the arm? I'm not sure I can follow this.
2. What type of neural network should I use (feedforward, backprop, etc.)? What kind of learning? How can I input
the obstacles to the ANN?
3.What software should I use? As I mentioned, I already have designed the arm in an CAD environment. Could I use only Matlab's neural network plugin and then export the results in a C/C++ compliant way that could be integrated in my app? how could I accomplish this?