In my talk, I will present neural control for locomotion of robot systems in particular walking machines. It is developed based on modular structures and utilizes discrete-time neurodynamics. As a result, it could generate a variety of walking patterns including reactive behaviors in a sensori-motor loop with respect to appropriate sensor inputs, to which a neural preprocessing is applied. Beside the generation of reactive walking behaviors, I will also show how learning mechanisms (i.e., supervised (error-based) and unsupervised (correlation-based) learning) can be integrated into the controller leading to adaptive behaviors. Furthermore, I will point out how reinforcement (reward-based) learning can be combined with unsupervised learning for optimal online learning and policy improvement of robot control. Ref.:Silke Steingrube, Marc Timme, Florentin Wörgötter & Poramate Manoonpong Self-organized adaptation of a simple neural circuit enables complex robot behaviour Nature Physics 6, 224 - 230 (2010)