S&AS: FND: Learning-Enabled Autonomous 3D Exploration for Underwater Robots

人类在水下工作通常会带来高昂的成本和安全风险，因为人们经常需要检查海底基础设施和环境的健康状况。这激发了对智能和可执行任务的自主机器人的需求，这些机器人可以监测和检查海底环境，并在无法获得准确的先验模型时探索周围环境。理想的任务机器人在没有精确指示的情况下，应该能够绘制出全面、准确的周围地图，反复决定下一步要去哪里，并确保在此过程中避免碰撞。该项目将利用机器学习技术来产生和部署新的算法，这些算法有可能提高水下机器人探索未知环境的速度和效率，并随着探索机器人获得更多经验，进一步提高性能。具体来说，该项目将引入机器学习技术，以(1)从水下区域的稀疏和噪声声纳数据中构建更具描述性的占用图；(2)节省了对许多候选感知动作进行穷尽评估的计算量；(3)学习探索真正需要三维空间推理的复杂、非结构化环境的有效行为。实时的3D勘探任务将与其他相关目标一起管理，例如最小化定位和地图的不确定性，以及与旅行相关的时间和精力支出。一个相关的目标是开发机器人系统，其性能随着经验的提高而提高，动态地在其组合中选择最有效的决策工具，并为手头的任务自参数化最合适的地图表示。

There are often high costs and safety risks associated with humans performing work underwater, which is frequently required to inspect the health of our subsea infrastructure and environment. This motivates a need for smart and taskable autonomous robots that can monitor and inspect the subsea environment, as well as explore their surroundings when they do not have access to an accurate prior model. Without precise instructions on where and how to explore, the ideal taskable robot should be able to produce comprehensive, accurate maps of its surroundings, make repeated decisions about where to travel next, and ensure that it avoids collisions in the process of doing so. This project will leverage machine learning techniques to produce and deploy new algorithms with the potential to enhance both the speed and efficiency with which underwater robots explore unknown environments, and to enable further gains in performance as the exploring robots gain more experience.Specifically, this project will introduce machine learning techniques to (1) build more descriptive occupancy maps from the sparse and noisy sonar data that typifies the subsea domain; (2) to save computational effort in the potentially exhaustive evaluation of many candidate sensing actions; and (3) to learn effective behaviors for exploring complex, unstructured environments that truly require three-dimensional spatial reasoning. The real-time, 3D exploration task will be managed in concert with other relevant objectives, such as minimizing localization and map uncertainty, and the time and energy expenditures associated with travel. A related goal is to develop robot systems whose performance improves with experience, dynamically choosing the most effective decision-making tools in its portfolio and self-parameterizing the most appropriate map representations for the task at hand.