There is no "magic bullet" in rehabilitation. In the absence of direct neural transplants, neurological rehabilitation is an arduous process. We have pioneered the clinical application of robotics in stroke rehabilitation and have shown evidence of the positive impact of targeted exercise on stroke recovery. In this article, we will review results obtained in the initial clinical trials with 96 stroke patients at the Burke Rehabilitation Hospital. We will provide evidence that robot-aided training enhances recovery, that this enhanced recovery is sustained in the long term, and that this recovery is not due to a general physiological improvement--in fact, it appears to be limb and muscle group specific. An evidence-based approach must now segue into a more scientific approach to stroke rehabilitation. Given the length of the required protocols and patients' variability and limited census, the practical limitations of the evidence-based approach are self-evident and extend trials for years. Each patient and lesion is unique in stroke rehabilitation, so there is no reason to believe that a "one-size-fits-all" optimal treatment exists. To optimize therapy for individual patients, we need science-based models. In this article, we will summarize the scientific tools and models that we are investigating and present some of the results to date.