This paper outlines an experimentally based design method for a compatible 3-DOF shoulder exoskeleton with an adaptive center of rotation (CoR) by matching the mechanical CoR with the anatomical CoR to reduce human–machine interaction forces and improve comfort during dynamic humeral motion. The spatial–temporal description for anatomical CoR motion is obtained via a specific experimental task conducted on six healthy subjects. The task is comprised of a static section and a dynamic section, both of which are recorded with an infrared motion capture system using body-attached markers. To reduce the influence of human soft tissues, a custom-made four-marker group block was placed on the upper arm instead of using discrete markers. In the static section, the position of anatomical CoR is kept stationary and calculated using a well-known functional method. Based on the static results, the dynamic section determines the statistical relationship between the dynamic CoR position and the humeral orientation using an optimization method when subjects move their upper arm freely in the sagittal and coronal planes. Based on the resolved anatomical CoR motion, a new mechanical CoR model derived from a traditional ball-and-socket joint is applied to match the experimental results as closely as possible. In this mechanical model, the CoR motion in three-dimensional space is adjusted by translating two of the three intersecting joint axes, including the shoulder abduction/adduction and flexion/extension. A set of optimal translation parameters is obtained through proper matching criterion for the two CoRs. Based on the translation parameters, a compatible shoulder exoskeleton was manufactured and compared with a traditional shoulder exoskeleton with a fixed CoR. An experimental test was conducted to validate the CoR motion adaptation ability by measuring the human–machine interaction force during passive shoulder joint motion. The results provide a promising direction for future anthropomorphic shoulder exoskeleton design.