Abstract: The proprioceptive training provided by existing robot-assisted (RA) rehabilitation methods or physical therapists is inadequate, and the repetitive nature of therapic walking and strength exercises limits rehabilitation efficiency. To overcome these challenges, a lower-limb rehabilitation robot system is developed, emphasizing proprioceptive training through complex foot movements, including dorsiflexion, eversion/inversion, and abduction/adduction. This system incorporates surface electromyography (sEMG) sensors, pose sensors, and six-axis force sensors to measure the robot end-effector trajectory and human-robot interaction data. A proportional-derivative controller with a disturbance observer achieves closed-loop control of robot end-effector pose, while a responsive variable-admittance control strategy based on sEMG creates a controllable and unstable environment through various task scenarios, prompting patients to make autonomous pose adjustments for avoiding falls, and thus enhancing proprioceptive input. Two training modes are designed for proprioceptive training in standing and walking on dynamic terrain. The results indicate that the activation of major lower-limb muscles is significantly enhanced, with continuous variations in muscle load and contraction types, thereby improving proprioceptive function and neuromuscular responses. These findings validate the effectiveness and safety of the proposed rehabilitation strategy. The designed robot and control strategy demonstrate potential for practical application in rehabilitation training to enhance recovery efficiency.