NRI: INT: SCHooL: Scalable Collaborative Human-Robot Learning

为了在仓库、家庭和从学校到零售店的其他环境中发挥作用，机器人需要学习如何稳健地操纵各种各样的物体。 例如，为了提高人类工人的生产力，服务和工厂机器人可以通过识别、抓取和将物体重新定位到适当的位置来保持指定表面的清洁。 对机器人进行预编程以执行如此复杂的操作任务是不可行的;相反，该项目将研究可扩展的机器人操作，其中多个机器人协同向多个人类学习。 该项目将贡献新的模型、算法、软件和实验数据，以推进深度学习、人机交互和云机器人技术的发展。 为了向学生和公众广泛传达这项研究的结果，该项目将与劳伦斯科学馆和非洲机器人网络合作制作一本书和视频。目前对合作机器人演示学习（LfD）的理解存在两个主要差距：1）缺乏一个包含人类和机器人的理论框架，以产生合作学习行为作为最佳解决方案; 2）缺乏将LfD与深度学习，分层规划和人机交互联系起来的研究。该项目通过基于逆向强化学习和人类与机器人之间通信的博弈论模型的统一理论框架来解决这些差距，将LfD视为一个可扩展的协作机器人过程，其中多个人类和多个联网机器人在一组分布式环境中工作，以最大化集体奖励函数，人类学习如何成为机器人更有效的演示者。这项研究几乎可以应用于任何机器人可以从人类演示中学习的环境，并将在项目过程中以越来越复杂的“表面清理”基准进行评估。

To be useful in warehouses, homes, and other environments from schools to retail stores, robots will need to learn how to robustly manipulate a wide variety of objects. For instance, to enhance the productivity of human workers, service and factory robots could keep specified surfaces clear by identifying, grasping, and relocating objects to appropriate locations. Pre-programming robots to perform such complex manipulation tasks is not feasible; instead this project will investigate scalable robot manipulation, where multiple robots collaboratively learn from multiple humans. The project will contribute new models, algorithms, software, and experimental data to advance the state-of-the-art in deep learning, human-robot interaction, and cloud robotics. To broadly convey the results of this research to students and the public, the project will create a book and video with the Lawrence Hall of Science and the African Robotics Network. Two primary gaps in current understanding of co-robotic Learning from Demonstration (LfD) are: 1) the absence of a theoretical framework that encompasses humans and robots to produce cooperative learning behaviors as optimal solutions; and 2) the lack of research linking LfD with deep learning, hierarchical planning, and human-robot interaction. The project addresses those gaps with a unified theoretical framework based on Inverse Reinforcement Learning and game-theoretic models of communication between humans and robots, treating LfD as a scalable co-robotic process in which multiple humans and multiple networked robots work in a distributed set of environments to maximize a collective set of reward functions and humans learn how to become more effective demonstrators for robots. The research can be applied to almost any context where robots can learn from human demonstrations and will be evaluated in "surface decluttering" benchmarks of increasing complexity over the course of the project.