CAREER: Safe and Influencing Interactions for Human-Robot Systems

职业：人机系统的安全且有影响力的交互

• 批准号: 1941722
• 负责人: Dorsa Sadigh
• 金额: $ 55万
• 依托单位: Stanford University
• 依托单位国家: 美国
• 项目类别: Continuing Grant
• 财政年份: 2020
• 资助国家: 美国
• 起止时间: 2020-02-01 至 2025-01-31
• 项目状态: 未结题
• 来源: https://www.nsf.gov/awardsearch/showAward?AWD_ID=1941722&HistoricalAwards=false
• 关键词: CAREER; Safe; Influencing; Interactions; Human

One of the key challenges when studying robotics systems that interact with people is the lack of access to a general model of how humans behave. Humans usually do not follow a fixed stationary model. They change and adapt to each other and to the robots over time. Humans gain experience - people's driving behavior when interacting with an autonomous car will be significantly different after many interactions. In assistive robotics, human responses will change as robots adapt. Routing decisions of autonomous cars influence other human drivers' routing choices and can result in undesirable global properties such as congestion. This introduces a new set of challenges including how robots should plan for safe and reliable strategies that are aware of their effects on people and the society as a whole. This project lays the foundations of analyzing and planning for repeated interactions between humans and robots. Our work will directly impact humans' comfort, safety, and public life by increasing robot understanding in interactions with humans in environments such as homes, hospitals, warehouses, and smart cities. The goal of this project is to focus on one of the key components of safe and interactive robotics -- formalizing influencing interactions, i.e., robot actions that influence human responses. This requires developing computational models of human behaviors, and leads to better understanding and formalisms for safe and reliable interactions with robots. This project investigates three main challenges: 1) human modeling: the investigator will develop data efficient methods to learn computational models of human behaviors while interacting with autonomous systems. One of the challenges in human-robot interaction is the lack of data from humans. This work develops active learning techniques that intelligently query and integrate different types of data collected from human feedback. 2) influencing interactions: it is clear that humans can be influenced by simple interactions with each other, e.g., people plan to arrive late if they are meeting a friend who is always late. Similarly, people's behavior changes when interacting with robots. If they observe an autonomous car being stuck at an intersection multiple times, they decide to navigate around it. This project plans to design robotics algorithms that are mindful of their effects on humans, how they can change the human behavior, and how that can help the overall system. 3) safe interactions: when planning for interactions that influence people, the robot will rely on learned human models. However, having access to a truly correct and reliable human model can be challenging due to lack of data or insufficient model parameters. This project designs verified robot policies that will be robust to inaccuracies present in human models, and takes the overall system to desirable states over long-term interactions.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.

研究与人类交互的机器人系统时面临的关键挑战之一是缺乏人类行为的通用模型。人类通常不遵循固定的静止模型。随着时间的推移，他们会改变并适应彼此以及机器人。 人类获得经验——人们在与自动驾驶汽车交互时的驾驶行为在多次交互后会显着不同。在辅助机器人技术中，人类的反应会随着机器人的适应而改变。自动驾驶汽车的路线决策会影响其他人类驾驶员的路线选择，并可能导致不良的全局特性，例如拥堵。 这带来了一系列新的挑战，包括机器人应如何规划安全可靠的策略，以了解其对人类和整个社会的影响。该项目为分析和规划人类与机器人之间的重复交互奠定了基础。我们的工作将通过增强机器人在家庭、医院、仓库和智能城市等环境中与人类互动的理解，直接影响人类的舒适、安全和公共生活。该项目的目标是专注于安全和交互式机器人技术的关键组成部分之一——形式化影响交互，即影响人类反应的机器人动作。这需要开发人类行为的计算模型，并导致更好的理解和形式主义，以实现与机器人的安全可靠的交互。该项目研究三个主要挑战：1）人类建模：研究人员将开发数据有效的方法来学习人类行为的计算模型，同时与自主系统交互。人机交互的挑战之一是缺乏人类数据。这项工作开发了主动学习技术，可以智能地查询和集成从人类反馈中收集的不同类型的数据。 2）影响互动：很明显，人类可以受到简单的相互互动的影响，例如，如果人们遇到一个总是迟到的朋友，他们就会计划迟到。同样，人们在与机器人互动时的行为也会发生变化。如果他们观察到一辆自动驾驶汽车多次卡在十字路口，他们就会决定绕过它。该项目计划设计机器人算法，考虑它们对人类的影响、它们如何改变人类行为以及如何帮助整个系统。 3）安全交互：在规划影响人类的交互时，机器人将依赖学习的人类模型。然而，由于缺乏数据或模型参数不足，获得真正正确且可靠的人体模型可能具有挑战性。该项目设计了经过验证的机器人策略，这些策略对人类模型中存在的不准确性具有鲁棒性，并使整个系统在长期交互​​中达到理想的状态。该奖项反映了 NSF 的法定使命，并通过使用基金会的智力评估进行评估，被认为值得支持。优点和更广泛的影响审查标准。

项目成果

Active Preference-Based Gaussian Process Regression for Reward Learning and Optimization

用于奖励学习和优化的基于主动偏好的高斯过程回归

• 发表时间: 2023-09
• 期刊: The International Journal of Robotics Research (IJRR
• 作者: Biyik, Erdem;Huynh, Nicolas;Kochenderfer, Mykel;Sadigh, Dorsa
• 通讯作者: Sadigh, Dorsa

Learning Bimanual Scooping Policies for Food Acquisition

学习双手舀取食物的策略

• 发表时间: 2022-01
• 期刊: Proceedings of the 6th Conference on Robot Learning (CoRL
• 作者: Jennifer Grannen; Yilin Wu
• 通讯作者: Yilin Wu

Learning Visuo-Haptic Skewering Strategies for Robot-Assisted Feeding

学习机器人辅助喂养的视觉触觉串接策略

• DOI: 10.48550/arxiv.2211.14648
• 发表时间: 2022-11-26
• 期刊: ArXiv
• 作者: Priya Sundaresan;Suneel Belkhale;Dorsa Sadigh
• 通讯作者: Dorsa Sadigh

Eliciting Compatible Demonstrations for Multi-Human Imitation Learning

引出多人模仿学习的兼容演示

• DOI: 10.48550/arxiv.2210.08073
• 发表时间: 2022-10-14
• 作者: Kanishk G;hi;hi;Siddharth Karamcheti;Madeline Liao;Dorsa Sadigh
• 通讯作者: Dorsa Sadigh

Inverse Preference Learning: Preference-based RL without a Reward Function

逆偏好学习：没有奖励函数的基于偏好的强化学习

• DOI: 10.48550/arxiv.2305.15363
• 发表时间: 2023-05-24
• 期刊: ArXiv
• 作者: Joey Hejna;Dorsa Sadigh
• 通讯作者: Dorsa Sadigh