EAGER: Human-Aware Navigation in Populated Indoor Environments

目前的自主移动机器人能够在人口稀少的地区准确导航，而不会撞到东西。然而，在公共建筑中常见的情况下，他们会遇到更多麻烦，比如在狭窄的走廊上通过人，在开阔、人口稠密的空间移动，或者在离开礼堂的人群中穿过人群。因此，自主机器人要想充分发挥其潜力，在对社会产生积极影响方面，它们需要提高自己的导航能力，使其更具“人类意识”。也就是说，他们将明确需要考虑需要在公共场所与之互动的人的特征。考虑到这一动机，这项研究的目标是了解如何最好地使移动机器人在人类密集的室内环境中平稳、稳健和安全地导航，以追求更高水平的目标，并由人类操作员以完全人类意识的方式进行不同级别的指导。本项目专注于两个互补的、高风险的、潜在的基础性研究，这两个研究对最终开发一个强大的、人类感知的导航系统至关重要。首先，它的目标是利用概率时序逻辑为安全的机器人-操作员-行人交互开发形式规范。其次，它的目标是开发生成操作员偏好的学习模型的方法，这些模型可以影响机器人关于例如轨迹平滑、子目标完成顺序、任务完成时间、旅行速度以及轨迹与行人和固定物体的接近程度的路径选择，学习用户偏好并确定如何将其与完成任务的奖励功能相结合。

Current autonomous mobile robots are able to navigate accurately through sparsely populated areas without bumping into things. However, they have more trouble in situations that commonly arise in public buildings, such as when passing people in narrow hallways, when moving through open, populated spaces, or when crossing a crowd of people exiting an auditorium. Thus, for autonomous robots to reach their full potential, in terms of positive impact on society, they will need to improve their navigational abilities to be more "human-aware." That is, they will explicitly need to take account the characteristics of the people with whom they need to interact in public spaces. With this motivation in mind, the goal of this research is to understand how best to enable mobile robots to navigate smoothly, robustly, and safely through human-populated indoor environments in pursuit of high-level goals, with varying levels of guidance from a human operator in a fully human-aware manner.This project focuses on two complementary, high-risk, and potentially foundational research thrusts as being crucial to laying the groundwork for eventual development of a robust, human-aware navigation system. First, it aims to develop formal specifications for safe robot-operator-pedestrian interactions, using probabilistic temporal logics. Second, it aims to develop methods for generating learned models of operator preferences that can influence the robot's choice of paths with regards to, for example, trajectory smoothness, order of subgoal achievement, task completion time, travel speed, and proximity of trajectory to pedestrians and fixed objects, learning user preferences and determining how to combine them with task-achieving reward functions.