Collaborative Research: Robots that Influence Human Behavior across Long-Term Interaction

每当机器人与人类并肩工作时，影响是不可避免的。然而，尽管目前的方法可以在短期内影响人类，但从长远来看，需要新的方法来理解和控制这种影响。这项工作的动机是观察到人与机器人的交互是双向的和不断发展的：随着人类适应机器人，一旦人类学习了机器人的模式、能力和目标，以前影响人类的机器人行为可能不再具有预期的效果。随着时间的推移，人类/机器人的交互可能会变得对抗性，人类可能会学习利用自主系统行为，突显出有影响力的行动和透明度之间的明显权衡。该奖项支持的研究将建立对人类认知和决策的新理解，不仅产生新的模型，捕捉人们如何随着时间的推移适应有影响力的机器人行为，而且还产生新的控制策略，在保持安全和激励的同时影响人类的期望行为。总体而言，随着时间的推移，在安全关键的情况下，该奖项有可能通过提高对人/机器人交互的理解来改善社会福祉。为了激励和培训K-12学生未来在工程领域的职业生涯，该团队将在大学开放日举办现场演示，学生们将直接与应用我们有影响力的算法的机器人互动。为了支持这些影响，本研究引入了一种多代理系统的建模和控制形式主义，使机器人能够长期影响人类。关键的见解是--而不是假设人类总是以同样的方式对机器人的行为做出反应--机器人应该将人类建模为动态的、适应性的主体，通过观察机器人的行为来学习。在这样做的过程中，研究人员明确地解决了透明行为和有影响力的行为之间的权衡问题，减少了人类利用自主系统的策略的机会，并提高了机器人保持影响代理行为的能力的能力。这一多学科调查团队的基础性工作将：(I)建立理论和数学模型，解释人类/机器人在长期对抗性环境中的适应；(Ii)控制机器人，即使人类与机器人一起共同适应，也保持影响力；以及(Iii)正式确定优化框架，以确保由此产生的行为在短期内是安全的，从长期来看不会产生意想不到的后果。这种人类建模和博弈论控制器的结合将为机器人如何在考虑到它们对人类的影响的同时做出决策提供新的、统一的见解。该奖项反映了NSF的法定使命，并通过使用基金会的智力优势和更广泛的影响审查标准进行评估，被认为值得支持。

Whenever robots work next to humans, influence is inevitable. Yet while current approaches can influence humans in the short-term, new approaches are needed to understand and control such influence over the long-term. This work is motivated by the observation that human-robot interaction is bi-directional and evolving: as humans adapt to robots, robot actions that previously influenced the human may no longer have the intended effect once the human learns the robot’s patterns, capabilities, and objectives. Over longer time frames, human/robot interactions may become adversarial, where humans may learn to take advantage of autonomous system behaviors, highlighting an explicit tradeoff between influential actions and transparency. This award supports research that will build new understanding of human cognition and decision-making, resulting not only in new models that capture how people adapt to influential robot behaviors over time, but also in novel control strategies that influence humans towards desired behaviors while maintaining safety and incentives. Overall, this award has the potential to improve societal well-being by improving understanding of human/robot interaction over time and in safety-critical situations. To inspire and train K-12 students for future careers in engineering, the team will host live demonstrations at university open houses where students directly interact with robots that apply our influencing algorithms. To support these impacts, this research introduces a modeling and controls formalism for multi-agent systems that enables robots to influence humans over the long-term. The key insight is that — instead of assuming humans always respond to robot behaviors in the same way — robots should model humans as dynamic, adaptive agents that learn by observing the robot’s actions. In doing so, the researchers explicitly address tradeoffs between transparent behavior and influential actions, mitigating the chance that humans will exploit the autonomous system’s policy and increasing the capability of the robots to maintain the ability to influence agent behavior. Foundational work by this multidisciplinary team of investigators will (i) build theory and mathematical models that explain human/robot adaptation in adversarial settings over the longer-term, (ii) control robots to maintain influence even as the human co-adapts alongside the robot and (iii) formalize an optimization framework to ensure the resulting behaviors are safe in the short-term, and do not have unintended consequences in the long-term. This combination of human modeling with game-theoretic controllers will provide novel, unified insights into how robots can make decisions while accounting for their effects on humans.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.