Improving Human Robot Collaboration Using Intention Signalling

【“提案”文档中完整引用的提案】物理人机协作是指人类和机器人为实现共同目标而共同努力的行为。人与机器人之间的合作通常与制造业有关，但在其他领域工作的人也很感兴趣。其中包括（但不限于）那些希望提高救援任务、太空旅行和医疗保健效率的人。例如，蛇形机器人被设计用来帮助快速找到被困在废墟下的人，机器人手臂被开发出来在精细的手术中提供帮助。尽管机器人的功能有了很大的发展，但研究表明，这些协作系统存在缺陷。一个突出的问题是，尽管机器可以对周围的物体和事件进行分类，但潜在的过程可能是不正确的。例如：一台机器可能能够正确地对狗和狼的图像进行分类——但它可能是基于背景中的雪进行分类的，直到出现错误我们才会发现。在机器对狗和狼的图像进行分类的情况下，错误的成本很低，但在人机协作中，这种分类错误的成本可能会导致资源（时间、精力和金钱）的损失，甚至对最终用户造成伤害。为了提高人机合作的效率，避免代价高昂的错误，有人建议，设计和建造系统的人与具有社会科学专业知识的人之间必须进行合作。具体来说，由于数据的复杂性和人类行为的复杂性，有人建议应该在认知科学、心理学和机器人技术方面的专家之间进行合作。这一提议源于这样一种组合——1)社会大脑行动实验室（一个专门研究心理学和认知神经科学的英国实验室）和2)反思互动、协作和参与实验室（一个专门研究计算机科学和机器人设计的加拿大实验室）之间的合作。尽管努力推导出能够清楚解释其决策基本原理的系统，但事实证明，这是极具挑战性和耗时的（Gilpin, 2018; Rudin, 2019）。我们提出了另一种解决方案，尽管有可能出现不正确的分类，但可以阻止错误的发生。具体来说，我们认为是机器人的动作导致了错误，而不是决策本身。因此，我们建议在行动之前，机器人应该向用户表明它将要做什么。一旦看到这个提示，用户可以允许运动继续，或停止机器，如果一个错误即将发生。除了提高协作的效率，我们预计对机器人的信任和互动的舒适度也将通过增加“意图信号”而得到改善。心理科学和机器人学的大量文献都表明，当所有成员的意图都被了解时，团队的工作效率会更高，所以我们有信心添加这个功能将提高互动的效率。为了研究机器人的“意图信号”是否提高了任务效率，提高了用户对机器人的感知（如喜欢、信任……），并提高了用户对协作的感知（如享受、舒适），我们打算进行一项实验。在实验中，我们建议进行一个桌面实验，参与者将与机械臂一起工作（要么发出信号，要么默默地行动）来完成一个谜题。多伦多大学的这项技术有能力处理多种信号，因此研究人员打算汇集他们的知识，决定在这个实验中应该使用哪种信号——例如，凝视提示、手势，或者是数字桌面上的可视化。

[Full referenced proposal in 'Proposal' document]Physical human-robot collaboration refers to the act of a human and a robot working together to achieve a common goal. Collaboration between humans and robots is generally associated with the manufacturing industry, however there is also much interest from those working in other sectors too. These include (but are not limited to) those hoping to improve the efficiency of rescue missions, space travel, and healthcare. For example, snake-like robots have been designed to help quickly find people trapped under rubble, and robotic arms have been developed to assist during delicate surgeries. Despite large developments in the functionality of robots, studies have indicated flaws in these collaborative systems. One prominent issue is that although machines can classify objects and events in its surroundings, the underlying processes can be incorrect. For example: a machine may be able to correctly classify images of dogs vs wolves - but it could be the case that it is classifying based on the snow in the background, and we wouldn't find out until a mistake was made. In the case of the machine sorting images of dogs vs wolves the cost of a mistake is low, but in human-robot collaboration the cost of such a classification error could lead to lost resources (time, effort, and money), or even injury to the end-user. To improve the efficiency of human-robot collaborations, and avoid costly errors, it has been suggested that there must be collaboration between those working to design and build the systems, and those with expertise in social sciences. Specifically, due to the complexity of the data and the intricacies of human behaviour, it has been suggested that there should be collaboration between experts in Cognitive Science, Psychology, and Robotics. This proposal stems from such a combination - a collaboration between 1) the Social Brain in Action Laboratory (a UK laboratory specialising in Psychology and Cognitive Neuroscience), and 2) the Rethinking Interaction, Collaboration and Engagement Laboratory (a Canadian Laboratory specialising in Computer Science and Robot Design).Despite much effort to derive systems which clearly explain the rationale for their decisions, it is proving extremely challenging and time-consuming (Gilpin, 2018; Rudin, 2019). We propose an alternative solution which, despite the possibility of incorrect classifications, could stop errors from errors occuring. Specifically, we argue that it is the action of the robot which leads to the error, not the decision itself. As a result, we propose that before acting, the robot should indicate to the user what it is going to do. Upon seeing this cue the user can allow the movement to continue, or halt the machine if a mistake is about to be made. In addition to increasing the efficiency of the collaboration, we anticipate that trust towards the robot, and comfortability in the interaction, will also be improved by adding 'intention signalling'. There is a wealth of literature from both psychological sciences and robotics which suggests that teams work more effectively when intentions of all members are known, so we are confident that adding this feature will increase the efficiency of the interaction.To investigate whether robot 'intention signalling' increases task efficiency, improves user perceptions of the robot (e.g. liking, trust...), and improves user perceptions of the collaboration (e.g. enjoyment, comfortability), we intend to conduct an experiment. In the experiment we propose to conduct a table-top experiment in which participants will work with a robotic arm (either signalling or acting silently) to complete a puzzle. The technology at the University of Toronto has the capability to do numerous signals, so the research intend to pool their knowledge and decide which signal should be used in this experiment - e.g. a gaze cue, a gesture, or perhaps a visualisation on a digital tabletop.