市場前線｜整合熱影像指向鎖定與輕量化語意映射之邊緣 服務機器人控制架構

隨著具身智能技術發展，服務機器人被期待能在真實環境中與人類進行互動。然而，傳統家電仰 賴硬編碼缺乏適應力，雲端運算亦有延遲及隱私疑慮。本文提出部署於地端設備之多模態功能型服務機 器人控制架構，以智慧語音風扇為載體進行驗證。本機器人具備三項能力：(1) 語音互動能力：整合大型 語言模型及零樣本語意映射；(2) 無光環境感知能力：以紅外線熱像儀結合雙層人體偵測，於無可見光條 件實現人體追蹤；(3) 指向鎖定能力：針對非步進馬達缺乏角度回授之瓶頸，設計遲滯控制迴路與漂移校正。 實驗結果顯示，語意映射正確率逾88%、推理延遲在500 ms 以內、移動追蹤誤差顯著降低，為地端物理 AI 應用提供實務參考。

As Embodied AI advances, service robots are expected to interact naturally and precisely with humans in real physical environments. However, conventional smart appliances still rely on hard-coded command sets, lacking adaptability to colloquial user intents, while cloud-based solutions introduce communication latency and privacy risks. This article presents a control architecture for edge service robots that integrates thermal imaging orientation locking and lightweight semantic mapping, validated through an intelligent voice-controlled fan platform. The robot features three core capabilities: (1) Voice interaction: integrating a lightweight LLM for zero-shot semantic mapping, enabling the robot to comprehend colloquial commands and provide spoken feedback in a closed-loop interaction; (2) Zero-visible-light perception: employing an infrared thermal imager with a dual-layer human detection algorithm for human localization and dynamic tracking in complete darkness; (3) Orientation locking: addressing the control bottleneck of non-stepper motors lacking angular feedback through a dynamic hysteresis control loop and drift correction mechanism. Experimental results demonstrate 88% semantic mapping accuracy, inference latency improved to 450-500 ms, and improved IR tracking efficiency with 23% error reduction in fast-movement scenarios, providing a practical system architecture reference for local Physical AI applications.