First off - stop cross-posting - it won't help you get your answer any faster, and it doesn't make you look good to others on the forum, either. That said...
Your statement of "we are familiar with neural networks but have no clue on how to implement it in a program" indicates to me that you -aren't- familiar with neural networks, except maybe in some vague and general sense. If you were familiar, you would at least have some concept of how to implement such a network (even if it were only a feed-forward perceptron model).
If you want to have any hope of implementing this robot successfully, you have a number of challenges ahead of you. You can tackle this in such a manner by using beacons or other external methods to help the robot navigate, or you can go the more difficult way by using a SLAM (Simultaneous Localization and Mapping) algorithm.
You can't just say to yourself "we'll use a neural network" and expect that to solve your problem. NNs can help within a SLAM algorithm, but there are probably better methods that could be utilized.
Actually, a neural network would be best for the voice recognition part, but that is actually likely to be the simplest (to a relative degree) part of the project. Tackling the mapping and localization aspect is the thing that will have you tearing your hair out; this is far from a generally solved problem - you could probably spend a few years on it as a researcher doing nothing else!
Unless you want to go the simpler route of providing external aids to the robot - beacons, line following, RFID tags, QR or other barcodes, etc - along with a pre-built map and such; then add on a machine learn algorithm (ANN or otherwise) for the voice recognition. But I would first get it to work with typed commands or such, before trying to get it working by talking to it.