생성형 모델을 활용한 노지 이동로봇의 멀티모달 전복방지 제어에 대한 연구 = A Study on Multimodal Rollover Avoidance Control of Field Mobile Robots Using Generative Model

다국어 초록 (Multilingual Abstract)

This study proposes a multimodal cGAN-based slope image estimation model that enables mobile robots to detect and avoid rollover hazards in unstructured and uneven outdoor environments. Conventional approaches relying solely on single-sensor data (e.g. camera, LiDAR data) are limited in estimating the absolute ground slope within the gravity reference frame. To address this limitation, a novel multimodal dataset composed of depth images and quaternion-based robot attitude information was constructed to reflect the robot’s inclination during operation. A FiLM-modulated cGAN generator architecture was then designed to incorporate robot state information as conditional features for image-based slope estimation.
The proposed model was trained using a dataset collected from real outdoor terrain, and its performance was validated using PSNR and SSIM metrics, showing improvements over previous studies. Real-time inference was achieved on an NVIDIA Jetson AGX Orin embedded platform mounted on the robot. The generated slope images were integrated into the ROS navigation stack to conduct real-time rollover-risk detection and avoidance experiments. Experimental results demonstrated that the proposed approach offers more than an eightfold improvement in computational efficiency compared to traditional PCA-based slope estimation and effectively avoids hazardous high-slope regions during autonomous navigation.
This thesis demonstrates the feasibility of integrating robot attitude information as multimodal conditions for real-time slope estimation and rollover avoidance control. Future work includes expanding the dataset to encompass various terrain and weather conditions, incorporating additional sensors such as LiDAR and GNSS, and adopting lightweight generative models to further enhance outdoor autonomous navigation capabilities.

목차 (Table of Contents)

• 1. 서론 1
• 1.1. 연구의 배경 및 목적 1
• 1.2. 관련 연구 6
• 1.2.1. 이동로봇의 비정형 노면 주행 연구 6
• 1.2.2. 이미지 생성 모델 10
• 1.2.3. 조건 정보 기반 특징 변조 기법 13
• 1.3. 논문의 기여 14
• 1.4. 논문의 구성 14
• 2. 노지 주행을 위한 이동로봇 시스템 구성 15
• 2.1. 전체 시스템 구성 15
• 2.2. 연구에서 사용한 이동로봇의 H/W 구성 17
• 3. 멀티모달 cGAN(conditional Generative Adversarial Network) 모델 19
• 3.1. 데이터 셋 수집 19
• 3.2. 지형 분석을 위한 이미지 데이터 전처리 21
• 3.3. 멀티모달 cGAN 모델 아키텍처 29
• 3.4. 모델학습 32
• 4. 이동로봇의 전복 회피 제어 35
• 4.1. 사전 지도 생성 35
• 4.2 이동로봇의 전복 방지를 위한 노면 경사 임계각 분석 39
• 4.3. 멀티모달 cGAN 추론 결과의 costmap 변환 43
• 5. 실험 및 결과 45
• 5.1. 데이터 구성 45
• 5.2. 실험환경 46
• 5.3. 평가 방법 47
• 5.4. 실험 결과 50
• 5.4.1. 멀티모달 cGAN 모델의 정량적 성능 평가 50
• 5.4.2. 전복 위험 지형 회피 제어 실험 결과 55
• 6. 결론 58
• 참고문헌 59