The input and output torque testing is crucial for improving the quality of Industry robot reducers. In this study, the TMMISR and TMMOSR for a vertical-type robot reducer detector were designed. The length of measurement chain between the torque transducer and the tested reducer was shorten. The overall stiffness of this instrument has been improved through structural optimization. The characteristics of the two main parts of the torque-measurement errors were also analyzed. A high precision torque calibrator with a standard torque output is used to handle the torque calibration process. An error compensation method based on a backpropagation neural network was adopted for the error compensation process in this study. After error compensation, torque-measurement precision of 0.1% can be achieved by the reducer detector over the full torque measurement scale, and the instrument can be used for both static and dynamic measurements.