| 摘要: | 本論文旨在設計一套僅需安裝輪速感測器、而無須額外力量感測器的智慧型輪椅動力輔助系統，以提升手動輪椅使用者的操作便利性與自主性。隨著高齡社會來臨，智慧輔助裝置於復健與日常生活中的需求顯著提升，然而現有動力輔助輪椅多依賴力量感測器量測推力或扭矩，安裝與維護成本高、系統整合亦較複雜。基於此，本研究發展外力觀測器估測駕駛者操作意圖，並結合強化學習演算法使輔助力輸出得以自動調整，提出一套兼具低感測器需求與即時適應能力之新型智慧型輪椅輔助控制架構。
本研究首先建立完整的輪椅動態模型與駕駛者模型，其中動態模型以拉格朗日方法推導其運動方程以描述輪椅在不同驅動條件下的行為；同時根據文獻與實際經驗建構具「一推一放」操作特性的駕駛者模型，使其能根據輪椅當前與目標狀態調整推力輸出與接觸時間。接續發展以多項式逼近法擴展的外力暨狀態估測器，將外力視為待估狀態並結合自適應卡爾曼濾波器，以在無外力感測器的條件下達成即時、穩健的狀態估測。此外，本研究亦建立向量自迴歸(Vector Autoregression, VAR)意圖預測模型，以推估駕駛者的速度與偏擺率操作趨勢，並於強化學習訓練中搭配不同模型設定進行系統性分析。訓練階段共使用兩種強化學習演算法，並結合兩類意圖預測模型定義；每種定義均採用2至6階的VAR結構，共形成二十組參數與模型的訓練組合，且每組參數與訓練組合接使用75種測試情境，用以探討不同預測準確度與模型階數對策略學習結果之影響。最終將強化學習策略與意圖預測模型整合至整體控制流程，並以多種典型軌跡進行模擬，以驗證架構的可行性與適應能力。
模擬驗證結果顯示，本研究提出之基於意圖預測的強化學習輔助策略，可有效降低使用者操作輪椅時的推動力需求，其省力效果相較於無輔助情形最高可達32%。若與傳統比例輔助策略相比，比例輔助於整體推動力降低方面具有較佳表現，其省力效果可達56%，顯示比例輔助在減輕使用者操作負擔上具有效率優勢；然而，比例輔助係依據推力大小進行放大，較難反映駕駛者即時操作意圖之變化。相對地，本研究所提出之強化學習輔助策略，雖在整體省力效果上不及比例輔助，但其輔助扭矩輸出能隨駕駛者操作狀態與意圖預測結果進行調整，在動態控制過程中更能貼近駕駛者實際操作意圖。就偏擺率控制精度而言，強化學習輔助策略之整體均方根誤差為0.08 rad/s，低於傳統比例輔助策略的0.12 rad/s；此外，當駕駛者目標速度提升至1.1 m/s以上時，由於傳統比例輔助的偏擺率誤差增加，使兩者的控制差異更加明顯。綜合上述成果，本研究建構之外力觀測、意圖預測與強化學習整合式架構具備實質效益，為智慧型輪椅於意圖貼合與動態輔助控制領域奠定了可行的技術基礎，並具有後續系統優化與應用拓展的發展潛力。
未來研究將進一步強化系統對不同駕駛者與操作情境之適應能力，並發展具備即時自適應特性的意圖預測與智慧輔助控制策略。同時，透過長期使用者體驗與人因評估，驗證其於實際應用中對操作負擔與使用品質之改善效果。
;This thesis aims to develop an intelligent power-assist system for manual wheelchairs that requires only wheel-speed sensors without relying on additional force sensors, thereby improving user convenience and operational autonomy. With the increasing demand for assistive technologies in an aging society, conventional power-assisted wheelchairs often depend on force sensors to measure user-applied push forces or torques, resulting in higher installation and maintenance costs as well as increased system complexity. To address these limitations, this study proposes a novel assistive control framework that integrates an external force observer for estimating user intent and reinforcement learning (RL) algorithms for adaptively adjusting assistive torques, achieving a system architecture characterized by low sensor requirements and real-time adaptability.
A complete dynamic model of the wheelchair and a driver behavior model were first established. The wheelchair dynamics were derived using the Lagrangian method to describe system behavior under various driving conditions, while the driver model was constructed based on literature and empirical observations to capture the “push-and-release” propulsion characteristics and adjust push force and contact duration according to the wheelchair’s current and target states. An external force and state estimator extended through polynomial approximation and integrated with an adaptive Kalman filter was then developed to achieve robust real-time estimation without requiring force sensors. In addition, a Vector Autoregression (VAR)–based intent prediction model was designed to estimate the user’s speed and yaw-rate intentions. During RL training, two reinforcement learning algorithms were combined with two definitions of the intent prediction model, each employing VAR structures of orders two through six. A total of twenty training configurations were constructed to analyze the influence of prediction accuracy and model order on learning performance. The learned strategies were subsequently integrated into the overall control framework and evaluated through simulations on multiple representative paths to verify system feasibility and adaptability.
Simulation results show that the proposed RL-based assistive strategy with intent prediction effectively reduces user effort, achieving up to 32% reduction in push force compared with the no-assistance condition. When compared with traditional proportional assist control, the two approaches demonstrate different application characteristics: the proposed strategy exhibits clear advantages in aligning with the user’s operational intent. In terms of yaw-rate control accuracy, the RL-based strategy achieves an overall root-mean-square error (RMSE) of 0.08 rad/s, lower than the 0.12 rad/s achieved by proportional assistance. Furthermore, when the user’s target speed exceeds 1.1 m/s, the yaw-rate error of proportional assistance increases, further enlarging the performance gap between the two strategies. Overall, the integration of external force observation, intent prediction, and reinforcement learning provides practical benefits for intelligent assistive control in wheelchairs, and establishes a solid technical foundation for future system optimization and broader applications. |
