모델 불확실성 상황에서 강건한 횡방향 제어를 위한 데이터 기반 하이브리드 손실 프레임워크 = Data-Driven Hybrid Loss Framework for Robust Lateral Control in the Presence of Model Uncertainty

다국어 초록 (Multilingual Abstract)

Trajectory tracking control is a crucial task for ensuring the safety of autonomous driving systems, and its control performance relies on the accuracy of high-fidelity dynamic models. However, in real driving environments, various uncertainties exist, such as changes in road surface conditions, variations in vehicle parameters, and disturbances like crosswinds. These uncertainties affect the model's accuracy, making the task challenging. To address these challenges, a novel Data-driven robust dynamic model estimation framework is proposed. This framework features three core components designed to infer robust vehicle dynamics under these compounding uncertainties. First, the dynamic parameter inference network (DPIN) is introduced. The DPIN employs a dedicated network architecture for sequence data processing and direct parameter mapping. This enables the effective extraction of dynamic characteristics from the current driving sequence, allowing for the continuous dynamic inference of core model parameters. Second, a model inference network (MIN) is configured. Parameterized in a complex diagonal form and recursively structured, the MIN draws inspiration from the linear recurrent unit (LRU) to effectively incorporate historical trajectory temporal features. Furthermore, physical constraints are directly embedded into the state space model layer to prevent model divergence, ensuring a theoretically consistent dynamic model structure. Third, training utilizes a multi-objective hybrid loss, which combines physics-informed and data-driven loss components. This hybrid approach maintains robustness by incorporating model uncertainties. Co-simulation experiments were conducted using Matlab/Simulink and CarMaker. The results demonstrate that the proposed framework exhibits improved performance and efficacy in dynamics scenarios with high uncertainty compared to the benchmark models.

목차 (Table of Contents)

• 제 1 장 서론 1
• 1.1 연구 배경 및 필요성 1
• 1.2 선행 연구 3
• 1.3 연구 목표 6
• 제 2 장 배경 이론 7
• 2.1 모델 예측 제어 7
• 2.2 Linear Recurrent Unit 11
• 제 3 장 프레임 워크 설계 14
• 3.1 시스템 모델 14
• 3.2 신경망 기반 차량 모델 학습 20
• 3.2.1 동적 파라미터 추론 네트워크 21
• 3.2.2 선형 상태 공간 모델 25
• 3.2.3 제어 적용형 모델 학습과 MPC 제어 설계 30
• 제 4 장 실험 및 결과 36
• 4.1 실험 방법 36
• 4.1.1 데이터 취득 36
• 4.1.2 실험 개요 및 실험 환경 37
• 4.1.3 복합 불확실성 환경 구성 39
• 4.2 실험 결과 42
• 4.2.1 기준 조건 실험 결과 분석 42
• 4.2.2 대표 복합 불확실성 환경에서의 실험 결과 44
• 4.2.3 전체 복합 불확실성 환경에서의 종합 비교 47
• 제 5 장 결론 50
• Appendix A 52
• A.1 Schur 안정과 Lipschitz 잔차 하의 안정성 증명 52
• 참고문헌 55