In this paper the vision architecture, named ROVIS, of the robotic system FRIEND is presented. The main concept of the ROVIS is the inclusion of feedback structures between different components of the vision system as well as between the vision and other modules of the robotic system to achieve high robustness against external influences of the individual system units as well as of the system as whole. The novelty of this work lies in the inclusion of feedback control at different levels of the 2D object recognition system to provide reliable inputs to the 3D object reconstruction and object manipulation modules of the robotic system FRIEND. The idea behind this approach is to change the processing parameters in a closed-loop manner so that the current image processing result at a particular processing level is driven to a desired result. The effectiveness of the ROVIS system is demonstrated through the presentation of experimental results on 3D reconstruction of different objects from FRIEND environment.