本研究提出一個創新的框架CoreBot，結合大型語言模型(Large Language Model, LLM)與機器人的高擬真模擬，在NVIDIA的Omniverse ISAAC Sim中透過直接的自然語言控制虛擬Dobot Magician E6機械手臂。透過引入系統運動學校正(Systematic Kinematic Calibration)和6路徑點關節空間軌跡差值系統(6-Waypoint Joint-Space Trajectory Inter-polation)，克服人機協作中一直以來的複雜性問題，並以全面且多層的驗證流程進行管控。CoreBot成功整合Google的Gemini LLM、ROS2 和 ISAAC Sim，構建出一套完整且即時的控制系統。關鍵技術貢獻在於提出一套專為模擬環境設計的校正流程，用於修正理論URDF參數和模擬環境間的差異。該方法顯著提高定位精度，將執行誤差降低至適合穩定執行任務的水準。為了提升使用者體驗，動作生成系統會將一連串獨立的指令串接為連續動作，相較於傳統的點對點控制更受使用者青睞。實驗結果表明，本系統具備穩健且快速的回應能力，能精確解析並執行指令。其中，多層驗證流程效果顯著，能有效阻止不安全或不可行的動作，例如超出工作區域限制與出現運動學錯誤。研究驗證了在事先模擬的模式下，開發精密且穩定的機器人控制系統的可行性。本研究提供一種可複現的校正方法，並將動作品質確立為以使用者為導向自然語言機器人領域的重要指標，為打造更易於操作且直覺的人機協作系統奠定基礎。;This research presents CoreBot, an innovative framework integrating a Large Language Model (LLM) with a high-fidelity robotics simulation to enable intuitive natural language control of a virtual Dobot Magician E6 robotic arm within NVIDIA′s Omniverse Isaac Sim. The system addresses the persistent complexity barrier in human-robot interaction with a virtual robot by introducing a Systematic Kinematic Calibration methodology and a 6-Waypoint Joint-Space Trajectory Interpolation system, all governed by a comprehensive multi-layer validation pipeline. The CoreBot architecture successfully integrates Google′s Gemini LLM, ROS2, and Isaac Sim into a cohesive, real-time control system. A key tech-nical contribution is a simulation-specific calibration process that corrects for discrepancies between theoretical URDF parameters and the simulation environment. This methodology yielded a substantial improvement in positioning accuracy, reducing execution errors to a level suitable for reliable task performance. To enhance user experience, the motion genera-tion system transforms discrete commands into continuous motions: a feature that was met with a strong user preference over traditional point-to-point control. Experimental results demonstrate a robust and responsive system capable of high-fidelity command interpretation and execution. The multi-layer validation pipeline proved highly effective, successfully pre-venting the execution of unsafe or infeasible motions, including workspace violations and kinematic failures. This research validates the viability of developing sophisticated and reli-able robotic control systems entirely within a simulation-first paradigm. It contributes a re-producible calibration methodology and establishes motion quality as a critical component of user-centric design in natural language robotics, laying the groundwork for more accessi-ble and intuitive human-robot collaborative systems.
