A neural network approach to online learning control and real-time implementation for a flexible space robot manipulator is presented. Motivation for and system development of the Self-Mobile Space Manipulator (SM/sup 2/) are discussed. The neural network learns control by updating feedforward dynamics based on feedback control input. Implementation issues associated with online training strategies are addressed, and a simple stochastic training scheme is presented. A recurrent neural network architecture with improved performance is proposed. By using the proposed learning scheme, the manipulator tracking error is reduced by 85% compared to conventional PID control. The approach possesses a high degree of generality and adaptability in various applications and will be a valuable method in learning control for robots working in unstructured environments.< <ETX xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">></ETX>