[en] This paper describes an original and robust method to optimize the design of closed-loop mechanisms, especially parallel manipulators. These mechanisms involve non linear assembling constraints. During optimization, the Newton-Raphson algorithm we use to solve these constraints may fail when the Jacobian matrix of the constraints is ill-conditioned and stops the redesign process. To circumvent the difficulty, the technique we propose takes advantage of numerical conditioning to penalize the objective function. Applications to an academic example and parallel robots demonstrate the capabilities of the methodology.