This paper proposes a new system integration approach for dextrous manipulation of multifingered robot hands. The proposed approach uses a modular and distributed architecture to integrate sensing, task teaching, planning, grasp force optimization, and control. The function modules are assembled in a distributed maimer and running independently so that this system is flexible in adding new function modules and eliminating existing ones. This developed system achieves simultaneously (1) automatic motion generation from human demonstrations, (2) real-time grasp force optimization for stable manipulation and regrasping, and (3) decentralized grasping and manipulation control. This architecture has been implemented in the CUHK hand system using DSP C40's and workstations. The system performance has been validated by experiments.