具資訊引導之非線性模型預測控制於非完整約束輪椅操控與線上參數估測

Abstract

在醫院與照護設施等受限室內環境中，若使用移動式機器手臂操控被動輪椅必須在確保安全的前提下，執行精準且平順的轉向與移動動作。然而，傳統輪椅操控方法多依賴已知或離線辨識之動態參數，並採用分離式的路徑規劃與追蹤控制架構，當系統慣性參數存在不確定性時，往往導致追蹤誤差累積，甚至違反空間與動態限制。此外，僅以追蹤誤差為目標的控制策略通常無法提供足夠的動態激發，使關鍵慣性參數在實際操作中難以被可靠辨識，進而限制自適應控制效能。本論文提出一套具資訊引導之自適應非線性模型預測控制（NMPC）架構，用於處理具不確定慣性參數之非完整約束輪椅系統。首先，本文以受限拉格朗日（constrained Lagrangian）方法推導一個以力為輸入的非線性動態模型。透過顯式消除與非完整約束相關的拉格朗日乘子，進一步得到一個 NMPC 預測的投影動態模型。接著，利用延伸卡爾曼濾波器（EKF），對未知慣性參數進行線上估測。為克服傳統追蹤式控制在激發不足下所面臨的基本限制，本文於 NMPC 成本函數中引入一項資訊引導之成本項。該成本項透過鼓勵角加速度的產生，主動促進具資訊性的運動行為，進而提升參數可識別性並加速 EKF 的收斂效果。EKF 與 NMPC 被整合於一個統一的遞迴預測回授迴路中，使控制器的預測模型能夠隨著參數估測結果持續更新。所提出的方法透過多種具代表性的模擬情境進行驗證，包括直線追蹤、走廊轉彎以及受限空間操控等案例。與基於回授線性化之 LQR、自適應倒退步進控制（adaptive backstepping control），以及未引入資訊引導激發之傳統 NMPC 相比，模擬結果顯示，本文方法在終點位置追蹤精度、參數收斂速度以及限制條件滿足性方面均展現出明顯優勢，即使在初始參數不確定性顯著的情況下亦能維持穩健表現。

In hospital and care-facility environments with confined indoor spaces, mobile manipulator–assisted control of passive wheelchairs must achieve precise and smooth maneuvering while strictly ensuring operational safety.However, conventional wheelchair control approaches often rely on known or offline-identified dynamic parameters and adopt decoupled planning-and-tracking architectures.When significant uncertainty exists in the inertial parameters, such methods tend to accumulate tracking errors and may even violate spatial and dynamic constraints.Moreover, control strategies that focus solely on tracking performance typically provide insufficient dynamic excitation, making key inertial parameters difficult to identify reliably during operation and thereby limiting the effectiveness of adaptive control.This thesis proposes an information-guided adaptive nonlinear model predictive control (NMPC) framework for nonholonomic wheelchair systems with uncertain inertial parameters.First, a force-based nonlinear dynamic model is derived using a constrained Lagrangian formulation.By explicitly eliminating the Lagrange multipliers associated with the nonholonomic constraints, a projected dynamic model suitable for NMPC prediction is obtained.An extended Kalman filter (EKF) is then employed to estimate the unknown inertial parameters online.To overcome the fundamental limitation of insufficient excitation in conventional tracking-based control, an information-aware cost term is incorporated into the NMPC formulation.This cost explicitly promotes informative motions by encouraging angular acceleration, thereby enhancing parameter identifiability and accelerating EKF convergence.The EKF and NMPC are integrated within a unified receding-horizon feedback loop, allowing the prediction model to be continuously updated using the estimated parameters.The proposed approach is evaluated through multiple representative simulation scenarios, including straight-line tracking, corridor turning, and confined-space maneuvering.Comparisons with feedback-linearization-based LQR, adaptive backstepping control, and conventional NMPC without information-aware excitation demonstrate that the proposed method achieves superior terminal position tracking accuracy, faster parameter convergence, and improved constraint satisfaction, even under significant initial parameter uncertainty.The results confirm the effectiveness of integrating excitation-aware estimation objectives directly into the NMPC framework for safe and reliable wheelchair motion in constrained environments.