The performance of a predictive controller is typically poor when the true plant evolution deviates significantly from that predicted by the model. A robust control approach that considers model uncertainty explicitly is then needed. However, it is often difficult to find a single input profile that works well for the range of uncertainty considered. Thus, multiple input profiles, i.e. one for each realization of the uncertainty, need to be determined. Unfortunately, this is computationally extremely expensive. This paper proposes an alternative approach that is based on neighboring extremals, where the multiple input profiles are computed using a simple feedback law, thereby reducing considerably the computational burden. The idea is illustrated via the simulation of a continuous stirred tank reactor and an inverted pendulum on a cart. (c) 2005 Published by Elsevier Ltd.