The automatic flatness control system applied to cold multi-roll mills for the production of metal strips has been subject of research efforts in many directions since the last 30 years. An innovative multivariable control approach for this application context is presented by means of which potential stability problems of controllers based on least mean squares are addressed. The proposed methodology decomposes the array of strip-elongation measurements produced by a conventional shape-meter into orthogonal components corresponding to the main actuator directions in order to reduce the dimensionality of the problem. Then, the control actions are computed by solving a prioritized constrained quadratic optimization problem. In this regard, the control problem is reformulated within a task-space control formalism - originally conceived in the robotics context - for which very efficient solution procedures do exist. Furthermore, in order to account for the model uncertainties that typically affects this kind of systems, the model of the process is adapted on-line by a numerically robust technique. The adaptive task-space flatness control scheme dealt with in the paper has been already commissioned in several installations having different actuator configurations and the experimental results shows its effectiveness and its easy configurability. (C) 2012 Elsevier Ltd. All rights reserved.