In this article, we present neural control of a three-legged reconfigurable robot with omnidirectional wheels. It is systematically synthesized based on a modular structure such that the neuromodules are small and their structurefunction relationship can be understood. The resulting network consists of four main modules. A so-called minimal recurrent control (MRC) module is for sensory signal processing and state memorization. It directly drives the motion of two front wheels while a rear wheel is indirectly controlled through a velocity regulating network (VRN) module. In parallel, a simple neural oscillator network module serves as a central pattern generator (CPG) producing basic rhythmic signals for sidestepping where stepping directions are controlled by a phase switching network (PSN) module. The combination of these neuromodules generates various locomotion patterns. Applying sensor inputs to the neural controller enables the robot to avoid obstacles as well as a corner. The presented neuromodules are developed and firstly tested using a physical simulation environment, and then finally transferred to the real robot.