Previously, we reported that chronic stroke subjects had significant improvements in isometric strength, free reaching extent, and clinical evaluations of function after training in the mirror-image movement enabler (MIME) robotic device. Our primary goal in this analysis was to investigate the hypothesis that the robotic training promoted improved muscle activation patterns. To this end, we examined the interaction forces, kinematics, and electromyograms recorded during training of eight different movement patterns in active-constrained mode. In this mode, the robot constrained the reaching movements to be toward the target, and the movement velocity was proportional to the force produced along the trajectory. Thirteen chronic stroke subjects trained in MIME for 24 1-h sessions over an eight-week period. Work output was significantly increased by week five in all eight movement patterns. Low-level subjects increased their extent of reach, while high-level subjects increased their speed. Directional errors in force production were reduced in six of eight movement patterns. Electromyographic data provided evidence for improved muscle activation patterns in the four movement patterns that started at tabletop level and ended at shoulder level. In contrast, there was no evidence of improved muscle activation patterns in any of the tabletop movements, with increased activation of antagonists in two movement patterns. This dichotomy may have been related to compensation at the shoulder girdle during movements that remained at tabletop level. A simple biomechanical model will be introduced to demonstrate the likelihood of this possibility.