Rehabilitation robot efficacy for restoring upper extremity function post-stroke could potentially be improved if robot control algorithms accounted for patient-specific neural control deficiencies. As a first step toward the development of such control algorithms using model-based methods, this study provides general guidelines for creating and simulating closed chain arm-robot models in the OpenSim environment, along with a specific example involving a three-dimensional arm moving within a two degree-of-freedom upper extremity rehabilitation robot. The closed chain arm-robot model developed in OpenSim was evaluated using experimental robot motion and torque data collected from a single healthy subject under four conditions: 1) active robot alone, 2) active robot with passive arm, 3) passive robot with active arm, and 4) active robot with active arm. Computational verification of the combined model was performed for all four conditions, whereas experimental validation was performed for only the first two conditions since torque measurements were not available for the arm. For the four verification problems, forward dynamic simulations reproduced experimentally measured robot joint angles with average root-mean-square (RMS) errors of less than 0.3 degrees and correlation coefficients of 1.00. For the two validation problems, inverse dynamic simulations reproduced experimentally measured robot motor torques with average RMS errors less than or equal to 0.5 Nm and correlation coefficients between 0.92 and 0.99. If patient-specific muscle–tendon and neural control models can be successfully added in the future, the coupled arm-robot OpenSim model may provide a useful testbed for designing patient-specific robot control algorithms that facilitate recovery of upper extremity function post-stroke.