In this thesis, a robot assembly wrist, which is able to assemble chamferless parts, has been developed. The RCC(Remote Center Compliance) structure is used as a basic structure. 5 position sensors and 4 pneumatic actuators are installed additionally to measure the deformation of RCC structure and correct the errors actively. Due to the restricted direction of actuation, a decision rule which selects the suitable actuator according to the position sensor signals is needed. For this purpose, neural network is used and it is experimentally shown that neural network overcomes system's nonlinearity. This paper presents fundamental experiment results for the insertion of parts with several clearance.