Large scale automatic scene generation to support deep-reinforcement-learning based navigation in autonomous mobile robots

• 主要作者: Bay, Natania Yining
• 其他作者: Andy Khong W H
• 格式: Final Year Project
• 語言: English
• 出版: Nanyang Technological University 2024
• 主題: Engineering; Medicine, Health and Life Sciences; Robotics; Simulation training; Dataset; Procedural generation; Reinforcement learning; Autonomous mobile robot
• 在線閱讀: https://hdl.handle.net/10356/177284
• 機構: Nanyang Technological University

With the ageing global population, already understaffed healthcare institutions face rising rates of hospitalization and chronic illnesses. This increases nurse burnout, decreasing quality-of-care and nurse retention. It also increases rates of patient oversight, causing prolonged hospital stays and higher mortality. Use of reinforcement-learning trained autonomous mobile robots to support intra-hospital prescription delivery and vital-sign monitoring at patient residences, can potentially alleviate understaffing and ensure continuity-of-care for patients with chronic illnesses, providing physicians with more comprehensive knowledge of patient health conditions. Focusing on navigation along corridors, shows distinct lack of relevant training data required for training robust navigation policies in this context. Furthermore, the high-dimensionality required of effective training data for robotic-AI applications, increases the difficulty and complexity of curating or constructing such datasets. This work develops an algorithm for bulk generation of logical yet diverse virtual corridor environments for such applications. Associated JSON information files, allow interfacing with the StableBaselines3 reinforcement-learning framework, facilitating policy training with multiple environments and target locations. Testing has shown efficacy of generated environments for training a navigation policy. Furthermore, the algorithm is designed for extensibility, allowing easy inclusion of more variations and new features, which stand to further increase the algorithm’s diversity, robustness, and functionality.