NRI: Collaborative Goal and Policy Learning from Human Operators of Construction Co¬-Robots

这项研究的总体目标是调查并显著推进人类操作员和协作机器人之间的协作交互的科学。这包括开发算法，可以用来从熟练的人类操作员那里训练合作机器人，以便在面对现实世界的不确定性时高效地执行复杂的任务，并指导新手操作员执行此类任务。主要的目标应用是建筑和农业设备行业，其中包括复杂的协作机器人，如挖掘机、轮式装载机、拖拉机、饲料收割机，这些领域非常需要了解和改进人-机器人协作学习。为协作机器人开发有效的算法存在重大的科学挑战，这些算法可以通过观察和提出适当的问题来关闭协作机器人和人类操作员之间的反馈回路，从而主动向熟练的人类操作员学习。该项目通过系统地制定和研究重点问题来解决这些挑战，以创建能够增强协作人类-机器人学习的高效算法。为了实现这一目标，设计了从模糊定义的复杂任务中协作学习潜在子目标结构的算法，开发了用于协作机器人实现所学习的子目标的实时路径规划和控制算法，并开发了为操作员提供特定技能的任务分解和运动执行指导的技术。此外，所开发的算法还得到了仿真、实验室和野外协作机器人的硬件实验以及理论分析的验证。

The overall goal of this research is to investigate and significantly advance the science of collaborative interaction between human operators and co-robots. This includes the development of algorithms that can be used to train co-robots from skilled human operators to efficiently perform complex tasks in the face of real-world uncertainty, and to guide novice operators in performing such tasks. The primary targeted application is the construction and farming equipment industry that includes complex co-robots such as excavators, wheel loaders, tractors, forage harvesters where there is a significant need to understand and improve human-robot collaborative learning.There are significant scientific challenges in developing efficient algorithms for co-robots that can actively learn from skilled human operators by observing and posing appropriate queries to close the feedback loop between the co-robot and the human operator. This project addresses these challenges by systematically formulating and investigating focused problems to create efficient algorithms that can enhance collaborative human-robot learning. To achieve the goal, algorithms are designed to collaboratively learn latent subgoal structures from ill-defined complex tasks, real-time path planning and control algorithms are developed for co-robots to achieve the learned subgoals, and techniques are developed to provide operator skill specific task decomposition and motion execution guidance. In addition, the developed algorithms are corroborated by simulators, hardware experimentation on laboratory and field co-robots, and theoretical analysis.