近年來，人型機器人與機械手掌逐漸應用於遠端操控及精細操作等領域。然而，現有手部動作捕捉系統多以關節角度作為映射依據，較少考慮操作者與接收端機械手掌之手指長度差異。為使機械手掌完成捏合或穩定抓取，部分映射方法會進一步調整接收端之關節角度，造成機械手掌與操作者手指姿態不一致。若未來導入力回饋系統，關節角度差異亦可能改變手指接觸物體時的受力方向，使接收端之作用力難以準確對應操作者實際姿態，進而影響遠端操作之真實性。 為此，本研究提出一套手部動作捕捉及模型長度自動校正系統。系統以慣性量測單元（IMU）作為手部動作擷取核心，經感測器校正後，利用Madgwick姿態融合演算法取得各手指指節之空間姿態。為估測操作者手指長度，本研究利用食指與拇指捏合形成之幾何閉迴路，建立閉迴路約束方程式，並透過最小平方法求解未知連桿長度，在無需額外距離感測器的情況下完成手指長度估測，並作為虛擬模型各指節長度調整之依據。 為改善多顆IMU實際穿戴與串聯量測時因相對姿態不一致所造成之誤差，本研究提出IMU二次校正方法，利用校正治具結合李群與李代數理論建立共同參考姿態，計算各感測器之校正矩陣，以修正感測器間的相對姿態偏差。實驗結果顯示，經二次校正後閉迴路末端誤差明顯降低，上迴路長度估測準確率由 84.83% 提升至 86.67%，下迴路則由 69.80% 提升至 77.70%，驗證所提出方法可提升手指長度估測之準確性。 最後將手部姿態與長度估測結果整合至Isaac Sim虛擬環境，並以Shadow Hand作為接收端模型進行操作驗證。模型經長度校正後，可在保留操作者原始關節角度映射的情況下，降低因手指長度差異所造成之接觸位置偏差，使虛擬模型與操作者具有更一致的幾何關係。本研究建立一套結合手部動作捕捉、手指長度估測、IMU二次校正及模型長度自動校正之完整系統，可作為未來導入力回饋機制與強化學習控制之研究基礎。;In recent years， humanoid robots and robotic hands have been increasingly applied to teleoperation and dexterous manipulation. However， existing hand motion capture systems primarily focus on joint-angle mapping and rarely consider differences in finger length between the operator and the receiving robotic hand. To achieve stable pinching or grasping， some mapping methods modify the joint angles of the robotic hand， resulting in inconsistency with the operator’s original finger posture. Such angular differences may also alter contact force directions， potentially affecting force feedback in future teleoperation systems. This study proposes a hand motion capture and automatic model length calibration system. Inertial measurement units (IMUs) are used to capture finger motions， while the Madgwick sensor fusion algorithm estimates the spatial orientation of each finger segment. To estimate the operator’s finger lengths， a geometric closed-loop constraint formed by the index finger and thumb is established， and unknown link lengths are solved using the least-squares method without additional distance sensors. To reduce relative orientation errors among multiple IMUs， a secondary IMU calibration method combining a calibration fixture with Lie group and Lie algebra theory is developed. Experimental results show that the estimation accuracy increased from 84.83% to 86.67% for the upper loop and from 69.80% to 77.70% for the lower loop. Finally， the estimated hand poses and finger lengths are integrated into Isaac Sim using Shadow Hand for validation. The calibrated model reduces contact-position errors while preserving the operator’s original joint-angle mapping， providing a foundation for future force-feedback and reinforcement-learning-based teleoperation systems.
