Research Experiences for Undergraduates (REU), NSF NHERI 2025: Novel View Synthesis for Autonomous UAV Path Planning Simulation in Hazardous Environments

This project creates a training pipeline for autonomous UAVs to perform damage assessment in hazardous post-disaster environments. The approach utilizes photorealistic 3D reconstruction from a real-world tornado damage dataset to create complex training environments through the Gaussian splatting algorithm. We apply reinforcement learning to train a PPO agent that autonomously detects and navigates to damage markers while maintaining operational safety boundaries. The pipeline integrates a minimal 2-value observation space and 3-action control paradigm, with strict training constraints and episode termination to develop precise learning behaviors. The datasets and methodology can be reused for testing autonomous navigation algorithms, computer vision benchmarks, and disaster assessment applications. This project is unique in attempting to develop an integrated approach that combines the 3D Gaussian Splatting algorithm with reinforcement learning methods specifically for disaster response, working toward a simplified yet effective solution to bridge the simulation-to-reality gap through photorealistic training environments. The primary audience includes disaster response researchers, autonomous systems engineers, emergency management agencies, and academic researchers in computer vision, reinforcement learning, and robotics working on hazardous environment applications.

本项目搭建了一套面向自主无人机（Unmanned Aerial Vehicle，简称UAV）的训练流程，用于在灾后危险环境中开展损毁评估工作。该方法依托真实龙卷风损毁数据集生成照片级真实感三维重建结果，通过高斯溅射（Gaussian Splatting）算法构建复杂的虚拟训练场景。我们采用强化学习技术训练近端策略优化（Proximal Policy Optimization，简称PPO）智能体，使其能够自主检测并导航至损毁标记点位，同时严格维持作业安全边界。该训练流程集成了极简二元观测空间与三动作控制范式，并设置了严格的训练约束与回合终止规则，以培育精准的学习行为模式。本数据集与研究方法可复用至自主导航算法测试、计算机视觉基准评测以及灾情评估应用场景中。本项目的独特性在于，尝试构建一套集成化方案，将三维高斯溅射算法与强化学习方法相结合，专门面向灾害响应场景，旨在通过照片级真实感训练场景打造简化且高效的解决方案，弥合仿真-现实鸿沟。本项目的核心受众包括灾害响应研究人员、自主系统工程师、应急管理机构，以及从事危险环境应用研究的计算机视觉、强化学习与机器人学领域的学术研究者。