NRI: INT: COLLAB: Robust, Scalable, Distributed Semantic Mapping for Search-and-Rescue and Manufacturing Co-Robots

该项目的目标是使多个协作机器人能够绘制和理解它们所处的环境，以便在教育、医疗援助、农业和制造应用中有效地相互协作，并与人类操作员进行协作。这个项目的第一个显著特征是环境将在语义上建模，也就是说，除了几何数据之外，它将包含人类可解释的标签（例如，对象类别名称）。这将通过计算机视觉和机器人技术的新颖、强大的集成方法来实现，从而使机器人和人类在现场之间的通信更容易。这个项目的第二个显著特征是，由于添加了人类可解释的信息而增加的计算负载将通过明智地在整个网络中近似和分散计算来处理。新开发的方法将通过模拟制造业和搜救行动的真实场景进行评估，从而为社会的大部分领域带来潜在的好处。该项目将为高中、本科和研究生提供培训机会，促进市场所需技能的发展。该项目将通过以下方式推进多机器人鲁棒语义映射的最新技术：1)开发一种新的优化框架，可以在重大测量误差下处理大型、动态、不确定的环境；2)明确允许和研究与人类的交互和信息交换，并通过优化框架的混合离散-连续扩展。3)通过实现计算的近似和平衡，允许智能地使用和共享协作机器人网络作为一个整体所拥有的有限计算资源。这些发展将受到两个特定案例研究的推动：一个是车间（小工厂）场景，其中机器人和固定摄像机用于在生产和零件组装过程中跟踪和协助人类工人；在一个经典的搜救场景中，操作员使用一个不同类型的机器人团队来快速评估损失并发现幸存者。当这两个应用程序一起考虑时，突出了当前流行的几何映射解决方案的所有局限性，并将用作项目结果的基准。

The goal of this project is to enable multiple co-robots to map and understand the environment they are in to efficiently collaborate among themselves and with human operators in education, medical assistance, agriculture, and manufacturing applications. The first distinctive characteristic of this project is that the environment will be modeled semantically, that is, it will contain human-interpretable labels (e.g., object category names) in addition to geometric data. This will be achieved through a novel, robust integration of methods from both computer vision and robotics, allowing easier communications between robots and humans in the field. The second distinctive characteristic of this project is that the increased computation load due to the addition of human-interpretable information will be handled by judiciously approximating and spreading the computations across the entire network. The novel developed methods will be evaluated by emulating real-world scenarios in manufacturing and for search-and-rescue operations, leading to potential benefits for large segments of the society. The project will include opportunities for training students at the high-school, undergraduate, and graduate levels by promoting the development of marketable skills.The project will advance the state of the art in robust semantic mapping from multiple robots by 1) developing a new optimization framework that can handle large, dynamic, uncertain environments under significant measurement errors, 2) explicitly allowing and studying interactions and information exchanges with humans with an hybrid discrete-continuous extension of the optimization framework, and 3) allowing an intelligent use and sharing of the limited computational resources possessed by the network of co-robots as a whole by enabling approximations and balancing of the computations. These developments will be driven by two particular case studies: a job-shop (small factory) scenario, where robots and fixed cameras are used to track and assist human workers during production and assembly of parts; and a classic search-and-rescue scenario, where operators use an heterogeneous team of robots to quickly assess damages and to discover survivors. These two applications, when considered together, highlight all the limitations of the currently prevalent geometric mapping solutions, and will be used as benchmarks for the project's results.