کنفرانس مهندسی برق ایران

صفحه اصلی / سی و چهارمین کنفرانس بین المللی مهندسی برق

Koopman-based Disturbance-Aware Trajectory Tracking via Neural Residual MPC for Autonomous Omnidirectional Wheeled Robots

نویسندگان :

Ashkan Jelodar¹ Kevin Babakhanloo² Mohammad Olyai³ Khalil Alipour⁴ Bahram Tarvirdizadeh⁵ Mohammad Ghamari⁶

1- دانشگاه تهران 2- دانشگاه تهران 3- دانشگاه تهران 4- دانشگاه تهران 5- دانشگاه تهران 6- Department of Electrical Eng, California Polytechnic State University

کلمات کلیدی :

omnidirectional mobile robots،Koopman model،neural residual learning،model predictive control،disturbance

چکیده :

This paper presents a simulation-based model predictive control (MPC) framework for omnidirectional wheeled mobile robots that integrates Koopman-based system modeling with neural residual learning. The proposed approach combines a model-based Koopman representation with a compact feed-forward neural network that compensates for unmodeled disturbances and approximation errors. The Koopman model is identified from disturbance-free simulation data using an Extended Dynamic Mode Decomposition (EDMD) procedure, ensuring a compact yet expressive disturbance-free simulation data structure. A neural residual module is subsequently trained to capture the discrepancies induced by constant biases, slow varying drifts, and stochastic process noise. The resulting hybrid predictor is incorporated into a discrete-time MPC formulation for robust trajectory tracking. Simulation studies under diverse disturbance conditions demonstrate that the hybrid Koopman– neural framework achieves significantly improved one-step and multi-step prediction accuracy compared with the analytic baseline, leading to enhanced closed-loop tracking performance and disturbance rejection. The proposed method provides a systematic and computationally efficient pathway for integrating model-based structure with data-driven refinement in nonlinear robotic control.