To address the challenges of insufficient dynamic acupoint recognition accuracy, inadequate operational compliance, and safety concerns in physiotherapy robots during treatment, this paper proposes a novel multimodal interactive physiotherapy robot system. Initially, we constructed an acupoint dataset and optimized a keypoint detection model through training, achieving significant improvement in dynamic acupoint identification precision. Subsequently, a pneumatic end-effector capable of simulating various clinical manipulation techniques was integrated. By incorporating physiotherapy specific motion characteristics and implementing a force-position hybrid control strategy, the system enables precise reproduction of therapeutic maneuvers and compliant operation. Experimental results demonstrate that this integrated system, combining high-accuracy dynamic acupoint recognition with multimodal therapeutic techniques, effectively delivers efficient, safe, and personalized physiotherapy services across diverse clinical scenarios.