There has been a rapid increase in the last ten years in the number of research groups and companies that are developing robotic devices to partially automate motor rehabilitation following neurologic injuries like stroke and spinal cord injury. A common approach is to design the robotic therapy device to physically assist the patient in making desired movements, mimicking the approach of "active assistance" sometimes used by rehabilitation therapists. In this talk, I will show that the motor system responds to such assistance differently for different movement tasks. For training of reaching movements, most clinical studies of robot therapy to date suggest that assistance neither hinders nor improves motor recovery. Rather, its primary benefit may be to provide a more motivating training environment. These findings bring into question the Hebbian-like conceptual framework of motor plasticity that is sometimes used to provide a scientific rationale for robot-assisted therapy. For training of walking, robot assistance allows a patient to practice faster, more natural looking movements more safely, but there is evidence it sometimes hinders recovery. This may be because the human locomotor system systematically "slacks" when given the opportunity. In a third application we are currently developing - teaching children with a severe disability to drive a powered wheelchair - we recently found that robotic assistance enhanced motor learning in a pilot study with non-disabled children. Understanding the neural mechanisms that elicit these differing responses - indifference, slacking, and enhancement - is essential for determining whether and how robots can actually be useful in rehabilitation and movement training practice.