This paper proposes a new l_infty observer design for fuzzy descriptor systems with unknown inputs. The descriptor form is treated using a singular redundancy system representation. To keep the consistency of the resulting fuzzy observer structure, we make use of a virtual variable playing the role of the one-step ahead state estimate. As a result, the observer gain can be constructed with free-structure decision variables to reduce the design conservatism. Using a fuzzy-basis-dependent Lyapunov function, the observer design is reformulated as a convex optimization problem with a single line search parameter. In particular, the error bounds of both the state and the unknown input estimations can be minimized through the guaranteed l_infty performance level. The effectiveness of the our result is demonstrated with a challenging application on robot manipulators.