NRI: Peer-to-Peer Human-Robot Coalitions

这项研究旨在创建由人类和机器人同行组成的大规模团队，在相同的物理空间中并行操作，每个人类和机器人基于自己的技能和能力执行身体动作。其目的是产生一种交互风格，这种交互风格不是基于人类对机器人的直接命令和控制，而是基于这样一种想法，即机器人可以通过被动观察隐含地推断出人类队友的意图，然后在当前环境下采取适当的行动。在这种交互中，人类以非常自然的方式执行任务，就像他/她与人类队友一起工作时一样，从而绕过了当人类被要求明确监督几个机器人团队成员的行动时发生的认知过载的困难。这项研究可以彻底改变人类和机器人在搜救、消防、安全、国防、灯光建筑、制造、家庭辅助和医疗保健等应用中的合作方式。这项研究集中在两个关键挑战上：(1)机器人如何仅通过传感器观察来确定人类当前的目标、意图和活动；(2)机器人如何做出适当的响应，以帮助人类完成正在进行的任务，与推断的人类意图一致。机器人系统的输入是一组学习的模型，以及颜色和深度感知。模型的学习使用人类感知和表示的新特征，包括兴趣深度特征、4维局部时空特征、自适应以人为中心的特征和基于单纯形的方向描述符。学习技术利用新的最大时间确定性模型进行序列活动识别，并利用条件随机场进行环境监测。机器人活动选择通过一种新颖的风险感知认知模型来实现。这项研究的结果将是新的软件方法，使机器人的认知、学习、传感、感知和行动选择能够用于对等的人-机器人合作。

This research aims to create large-scale teams of human and robot peers that operate side-by-side in the same physical space, with each human and robot performing physical actions based upon their own skills and capabilities. The intent is to generate an interaction style that is not based on direct commands and controls from humans to robots, but rather on the idea that robots can implicitly infer the intent of human teammates through passive observation, and then take appropriate actions in the current context. In this interaction, humans perform tasks in a very natural manner, as he/she would when working with a human teammate, thus bypassing the difficulty of cognitive overload that occurs when humans are required to explicitly supervise the actions of several robot team members. This research can revolutionize how humans and robots work together in applications such as search and rescue, firefighting, security, defense, light construction, manufacturing, home assistance, and healthcare.This research focuses on two key challenges: (1) how robots can determine humans' current goals, intents, and activities via sensor observation only, and (2) how robots can respond appropriately to help humans with the ongoing task, consistent with the inferred human intent. Input to the robot system is a set of learned models, along with color and depth sensing. Models are learned using novel features for human perception and representation, including Depth of Interest features, 4-dimensional local spatio-temporal features, adaptive human-centered features, and simplex-based orientation descriptors. Learning techniques make use of novel maximum temporal certainty models for sequential activity recognition, and conditional random fields for environmental monitoring. Robot activity selection is achieved via a novel risk-aware cognitive model. The outcome of this research will be new software methodologies enabling robot cognition, learning, sensing, perception, and action selection for peer-to-peer human-robot teaming.