We report on an experiment in which a population of rat cortical neurons, cultured on a micro-electrode array, was connected bi-directionally to a mobile robot. Bi-directional communication between a neural population and an external device requires to translate time-varying signals into spatio-temporal patterns of neural activity, and back. Here, we describe the experimental setup and the computational modules of the neural interface, and work on characterization of the ‘transfer function’ of the neural preparation, as it emerges from closed-loop experiments, and selection of stimulation and recording sites which are best compatible with a desired behavior.