Learning from Crowdsourced Virtual Reality Demonstrations
Learning from Crowdsourced Virtual Reality Demonstrations
复制标题
从众包虚拟现实演示中学习
DOI：
--
复制
发表时间：
2018
期刊：
影响因子：
--
通讯作者：
Stefanie Tellex
中科院分区：
文献类型：
--
作者：
Eric Rosen;David Whitney;Stefanie Tellex
从演示中学习（LfD）已经成为一种广泛流行的教机器人如何执行操作任务的方法，因为它利用了人类的知识。然而，收集可用于学习机器人策略的高质量演示可能既耗时又困难。最近，一些研究人员已经开始使用消费级虚拟现实（VR）硬件作为更有效的远程操作机器人收集演示的手段。在这个领域以前的工作集中在任务和算法，需要相对较少的数据，由于演示生成的时间下沉。我们提出了一个新的众包框架，利用大型虚拟现实游戏社区。通过将这些经验丰富的VR用户视为公民科学家，我们将为机器人提供完成复杂操作任务所需的演示数据。
Learning from demonstration (LfD) has been a widely popular methodology for teaching robots how to perform manipulation tasks because it leverages human knowledge. However, collecting high quality demonstrations that can be used for learning robot policies can be time-consuming and difficult. Recently, some researchers have begun using consumer-grade virtual reality (VR) hardware as a more efficient means of teleoperating a robot for collecting demonstrations. Previous work in this space has focused on tasks and algorithms that require relatively little data due to the time-sink of demonstration generation. We propose a novel crowd-sourcing framework that takes advantage of the large virtual reality gaming community. By treating these experienced VR users as citizen scientists, we will empower our robot with the demonstration data needed to complete complex manipulation tasks.