This paper presents the work currently done at LAAS/CNRS about scene interpretation required for manipulation tasks by a mobile arm. This task is composed of two steps: the approach of the mobile platform along the manipulation site and the grasping itself. The paper focuses on the object recognition and localization: the approach step is performed by a simple laser-based navigation procedure. For the grasping step, we use a CAD model of the object and discuss of the problems linked with such a representation: visibility informations must be added so that recognition and grasping strategies could be selected in a formal way. For the recognition, first matchings concerning discriminant patterns allow to generate a first prediction about the object situation; an optimal verification viewpoint can be computed. From this new camera position, we search for maximal sets of matched image features and model primitives; the best recognition hypothesis is determined by the best score. If no prediction can be determined, the system may switch to other discriminant patterns or move the camera respectfull to the arm and robot constraints.