Purpose - The purpose of this paper is to introduce a development and error modeling of a serial redundant manipulator system applied in nuclear fusion environment. Detailed mechanical design of vacuum-compatible arms and actuators are presented, and manipulator flexibility is studied through experiments and model identification algorithm to evaluate the deformation. Design/methodology/approach - First, the manipulator is designed to be several modular segments to obtain enough and flexible workspace inside the fusion device with narrow and complex geometries. Joint actuators with "rotation-linear-rotation" chains are developed to provide both huge reduction ratios and vacuum sealing. The redundant manipulator system has 11 degree of freedoms in total with a storage cask system to dock with the device vacuum vessel. In addition, to improve the position accuracy, an error prediction model is built based on the experimental study and back-propagation neural network (BPNN) algorithm. Findings - Currently, the implementation of the manipulator system has been successfully carried out in both atmosphere and vacuum condition. Excellent performance indicates that the mechanical design is suitable. The validation of BPNN model shows an acceptable prediction accuracy (94 similar to 98 per cent) compared with the real measurement. Originality/value - This is a special robot system which is practically used in a nuclear fusion device in China. It will allow remote inspection and maintenance of plasma facing components in the vacuum vessel of fusion device without breaking the ultra-high vacuum condition during physical experiments. Its design, mechanism and error prediction strategy have great reference values to the similar robots in vacuum and temperature applications.