Previously, we reported that the MIME robotic device for post-stroke neuro-rehabilitation has quantifiable clinical benefits. To identify which aspects of the robotic training were most effective, we examined the relative effectiveness of eight different shoulder-elbow reaching movements. Thirteen chronic stroke subjects trained in MIME for 24 one-hour sessions. In each session, all eight movement patterns were trained, with the robot constraining the movements to be toward the target. The movement velocity was proportional to the force produced along the trajectory. After training, subjects had significant gains in the work produced in all 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 tabletop patterns. We hypothesize that compensation with shoulder girdle movement limited the effectiveness of the tabletop movements in promoting neuro-rehabilitation. We conclude that effective robotic treatment depends critically upon using movement patterns and modes of assistance that limit the effectiveness of compensation and require activation of the target muscle groups.