Recent progress in embodied artificial intelligence and physiological sensing enables mobile robots to interpret heterogeneous biosignals directly on board with minimal latency. However, existing multimodal fusion methods are computationally demanding, limiting their deployment on embedded platforms with constrained resources. To address this, we propose a lightweight Depthwise-Separable Transformer-based multimodal fusion framework integrated into a ROS2-native pipeline for autonomous health monitoring. Our approach efficiently fuses ECG, respiration, IMU, and acoustic signals in real time on edge devices without relying on cloud computation. Experimental validation on a TurtleBot3 robot demonstrates a 31% reduction in inference time and a 26% decrease in energy consumption compared to a standard transformer, while maintaining comparable diagnostic accuracy. These results highlight the feasibility and effectiveness of fully on-board multimodal fusion for scalable, autonomous, and privacy-preserving healthcare robotics.