機器人無治具組裝技術研究

無治具組裝 (Fixtureless Assembly) 為機器人邁向高度彈性化的核心指標，亦為製造業面對少量 多樣生產之關鍵挑戰。傳統組裝技術仰賴精密治具與預先編程，難以因應頻繁換線之成本與時間壓 力。本文整合三大核心技術：(1) 以 Google Gemini API 為基礎之大型語言模型 (LLM) 任務規劃，可 將自然語言指令轉換為可執行動作序列並支援即時錯誤恢復；(2) 視覺、力覺與多模態大型語言模型 (MLLM) 之三重感測融合，依序執行夾取、組裝與判定；(3) 雙臂協作控制演算法，搭配 YOLO 物件檢 測與 Gemma 3 4B 多模態模型。於木製椅子組裝實驗中，達成物件辨識率 95%、組裝成功與失敗判定正 確率分別為 95% 與 94%，並維持力控誤差小於 1 N，驗證本架構於無治具情境下兼具彈性與穩健性。

Fixtureless assembly is a core indicator of robotic flexibility and a critical challenge for high-mix, low-volume manufacturing. Traditional robotic assembly relies on precisely designed hardware fixtures and pre-programmed motions, incurring high cost and long lead time when product lines change frequently. This article integrates three core technologies: (1) large language model (LLM)–based task planning using the Google Gemini API, which translates natural-language instructions into executable action sequences and supports real- time error recovery; (2) triple-modality sensor fusion combining vision, force/torque sensing, and a multimodal large language model (MLLM) for the grasping, assembly, and success-detection phases; and (3) a dual-arm collaborative control algorithm coupled with YOLO object detection and the Gemma 3 4B MLLM. In a wooden-chair assembly experiment, the proposed system achieves a 95% object recognition rate, 95% / 94% success/failure detection accuracy, and force-control error below 1 N, demonstrating both flexibility and robustness in fixtureless scenarios.