Learning-Augmented MPC for Autonomous Vehicle Path Tracking via Ensemble Residual Dynamics Learning.

Area of Science:

• Control Engineering
• Machine Learning
• Automotive Dynamics

Background:

• Accurate vehicle dynamics modeling is crucial for path tracking control.
• Nonlinear and time-varying behaviors degrade traditional Model Predictive Control (MPC) performance.

Purpose of the Study:

• To propose a learning-augmented MPC framework using a Data-Driven Dynamics Refinement (DDR) model.
• To enhance predictive accuracy and control robustness in vehicle path tracking.

Main Methods:

• An ensemble learning-based DDR Model complements nominal vehicle dynamics.
• An ensemble of neural predictors improves generalization and stability.
• A feature-driven activation mechanism selectively applies refinement to reduce computational load.

Main Results:

• The refined dynamics significantly improve tracking accuracy and robustness.
• Maximum lateral deviation reduced by ~6 cm (single-lane change) and ~4 cm (double-lane change).
• Maximum heading error reduced by 0.02 rad and 0.015 rad, respectively.

Conclusions:

• The proposed learning-augmented MPC framework effectively enhances vehicle path tracking.
• The DDR Model accurately captures complex dynamics, improving control performance under demanding conditions.