Abstract: The control strategy of rehabilitation robots should not only adapt to patients with different levels of motor function but also encourage patients to participate voluntarily in rehabilitation training. However, existing control strategies usually consider only one of these aspects. This study proposes a voluntary and adaptive control strategy that solves both questions. Firstly, the controller switched to the corresponding working modes (including challenge, free, assistant, and robot-dominant modes) based on the trajectory tracking error of human-robot cooperative movement. To encourage patient participation, a musculoskeletal model was used to estimate the patient’s active torque. The robot’s output torque was designed as the product of the active torque and a coefficient, with the coefficient adaptively changing according to the working mode. Experiments were conducted on two healthy subjects and four hemiplegic patients using an ankle rehabilitation robot. The results showed that this controller not only provided adaptive the robot’s output torque based on the movement performance of patients but also encouraged patients to complete movement tasks themselves. Therefore, the control strategy has high application value in the field of rehabilitation.