This paper proposes a reinforcement learning based radial-basis function network control system (RL-RBFNCS) to solve non-training data based learning of radial-basis function network controllers (RBFNC). In learning process, a major contribution is by using the critic signal and the stochastic exploration method to estimate the “desired output”, reinforcement learning is considered and solved from the point of view of training data based learning. Computer simulations of robot obstacle avoidance in unknown environment are conducted to show the performance of the proposed method.