EAGER: Reconciling Model Discrepancies in Human-Robot Teams

机器人在许多领域极大地补充了人类的能力，但让它们作为队友发挥作用往往被证明是具有挑战性的。这部分是由于缺乏对机器人的了解，这经常导致人们发现他们对机器人队友的期望在现实中无法实现。提出的研究通过让机器人推理人们对它们的模型，并使用这些模型以更接近期望的方式行事，或者解释感知到的差异，来解决这个问题。这项工作可以对涉及人在环人工智能系统的关键领域产生重大影响，例如决策支持、自动化制造、老年人护理和医疗机器人。为了成为可靠的队友，机器人必须理解人类伙伴对机器人任务和能力的期望，并能够在期望和实际模型之间寻求协调。在这个项目中，调和将通过以下方式实现：1)对机器人的行为进行偏置，以隐式地适应模型差异；或者2)通过沟通明确地减少差异。第一种方法，称为模型协调规划，将被表述为一个优化问题，该问题生成一个机器人执行的计划，同时最小化其与人类设想的预期计划的距离。将开发启发式搜索方法来适应这种方法。第二种方法将生成解释，以更新人类对机器人的模型，使机器人的计划更接近于更新模型中人类的期望。此外，机器学习将用于在没有明确提供的情况下近似人类期望模型，并且将开发与这些学习和不完整模型相协调的模型方法。该奖项反映了美国国家科学基金会的法定使命，并通过使用基金会的知识价值和更广泛的影响审查标准进行评估，被认为值得支持。

Robots greatly complement human capabilities in many domains, but having them function as teammates has often proven to be challenging. This is due in part to a lack of understanding of robots, which often leads humans to find that their expectations of robotic teammates are not met in reality. The proposed research addresses this by having robots reason about the models people have of them and using those models to either behave in ways that meet the expectations more closely or explain the perceived differences. The work can have a significant impact on critical domains that involve human-in-the-loop AI systems, such as decision support, automated manufacturing, eldercare, and medical robotics. To become reliable teammates, robots must understand the expectations of their human partners regarding the robots' tasks and abilities, and be able to seek reconciliation between the expected and actual models. In this project, reconciliation will be achieved either by 1) biasing the robot's behavior to implicitly accommodate model differences; or 2) communicating to explicitly reduce the differences. The first approach, termed model reconciliation planning, will be formulated as an optimization problem that generates a plan for the robot to execute while minimizing its distance to the expected plan that the human envisions. Heuristic search methods will be developed to accommodate this approach. The second approach will generate explanations to update the human's model of the robot in such a way that the robot's plan more closely matches that of the human's expectation in the updated model. In addition, machine learning will be used to approximate the model of human expectation when not provided explicitly, and methods for model reconciliation with these learned and incomplete models will be developed.This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.