Study results in the last decades show that amount and quality of physical exercises, then the active participation, and now the cognitive involvement of patient in rehabilitation training are crucial to enhance recovery outcome of motor dysfunction patients after stroke. Rehabilitation robots mainly have been developed along this direction to satisfy requirements of recovery therapy, or focused on one or more of the above three points. Therefore, rehabilitation robot based on neuro-machine interaction has been proposed for the paralyzed limb training of post-stroke patient, which utilizes motor related EEG, UCSDI (Ultrasound Current Source Density Imaging), EMG for the robot control and feeds back the multi-sensory interaction information such as visual, auditory, force, haptic sensation to the patient simultaneously. This neuro-controlled and perceptual rehabilitation robot will bring great benefits to post-stroke patients. In order to develop such a kind of rehabilitation robot, some key technologies, such as noninvasive precise measurement and decoding of neural signals, realistic sensation feedback, coordinated control for both the rehabilitation robot and the patient, need to be solved. In this paper, some fundamental problems in developing and optimizing such a kind of rehabilitation robot based on neuro-machine interaction are proposed and discussed.