S U M M A R YThe goal of this dissertation was to improve rehabilitation robots by developing new patient-friendly devices which can assist therapists in the rehabilitation of neurological movement disorders of the upper extremities, such as hemiparetic stroke. In all, three novel rehabilitation devices were developed. Given the changing demographics of developed nations, over the next two decades fewer therapists will be available to treat an increasing number of stroke patients. With patient-friendly robots assisting therapists, proven therapeutic exercises can be automated and new and better targeted interventions developed and tested. In addition to facilitating therapy sessions, robots can provide objective measurement of impairment. Overall, robots can make therapy more productive for patients and less labor-intensive for therapists, and provide physicians, therapists and the scientific community with more objective data.To handle the wide range of impairments found in hemiparetic stroke patients, multiple devices are needed. Mildly impaired patients may have a near-normal range of motion, but have problems with fine motor control or moving heavier objects. Severely affected patients may not be able to even lift the weight of their own arm. Despite these differences, certain general strategies appear to work best. First and foremost, rehabilitation therapy works best when the patient is actively and intensively involved. Exercises should use repetitive movements that closely resemble those used in daily living for best results in reshaping the recovering brain.Studying motor learning in healthy subjects supported the need for active patient involvement. Given the artificial motor relearning task of moving in a visuomotor-rotated field, the healthy subject fully adapted when he could freely make, and correct, errors in their movement execution. On the other hand, active delivery of a passive hand to the targets resulted in much less and much slower adaptation. In between lay the adaptation achieved with hard and soft guidance of the hand over virtual tracks. The conclusion is that both minimization of execution errors and control effort drive kinematical adaptation in a novel visuomotor task, but the latter occurs at a much slower rate.Should the patient not be assisted at all? Perhaps the type of assistance is important. For instance, weight support of the arm facilitates movement, but movement initialization and control are left unchanged. Most rehabilitation devices for upper extremities include some form of weight support. An analysis of these devices concluded that weight support is most easily realized through a cablesuspension system that supports the arm via slings. However, the best possible solution for weight support depends on the primary design of the device. Careful upfront consideration of various design options will lead to better choices.The first rehabilitation device was designed using this knowledge. The Freebal is a dedicated weight-support system that is less complex and has less movem...