This paper presents a real-time fuzzy modeling approach based on on-line clustering for a family of complex systems with severe nonlinearity such as robotic manipulators. The fuzzy model (Takagi-Sugeno fuzzy system) is identified real-time by online clustering and recursive least square estimation (RLSE). Using this method, the fuzzy rules can be added, modified and deleted automatically when the new data comes, and the consequence parameters of the T-S model can be recursively updated. Simulation results for a two-degree-of-freedom robot demonstrate the effectiveness and advantages of this approach.