This paper presents a visual feedback available robotic gait training system for motor recovery of hemiplegic stroke survivors. The system is composed of a treadmill consisting of two split belts, a pelvic support manipulator assisting patient's pelvic movement, and a visual interface feedbacking patient's gait phase. The split-belt treadmill allow patient to walk in different velocities between sound side and affected side legs, and detect patient's gait phase by current value of DC motor. The pelvic support manipulator provides three active actuations to assist patient's leg swinging during walk training. The virtual walking scenario gives visual feedback for patient while providing patient's gait phase calculated from motor current of the treadmill. One subject with simulated stroke participated in this study. Experiment results indicate gait phase of the virtual model can well track that of patient's walking, verify the feasibility of the proposed system to improve gait recovery during rehabilitation.