RI: Small: Probabilistic Goal-Based Imitation Learning

RI：小：基于概率目标的模仿学习

• 批准号: 1318733
• 负责人: Rajesh Rao
• 金额: $ 40万
• 依托单位: University of Washington
• 依托单位国家: 美国
• 项目类别: Standard Grant
• 财政年份: 2013
• 资助国家: 美国
• 起止时间: 2013-08-01 至 2018-07-31
• 项目状态: 已结题
• 来源: https://www.nsf.gov/awardsearch/showAward?AWD_ID=1318733&HistoricalAwards=false
• 关键词: RI; Small; Probabilistic; Goal; Based

Humans are extremely adept at learning new skills by watching and imitating others. Attempts to endow robots with a similar ability have failed to generalize beyond specific tasks, partly because the focus has been on following the trajectory of an action demonstrated by an expert. The current project investigates a new interdisciplinary approach to imitation learning that is inspired by how humans learn via goal-based imitation. The project's specific objectives include: (1) a new method for imitation based on inferring the underlying goals of human actions rather than following trajectories: actions are executed based on sequences of inferred goals and successfully executed action sequences are cached as higher level goals, leading to hierarchical goal-based imitation; (2) a new approach based on hierarchical Bayesian models (HBMs) is proposed for generalization across objects and tasks, and (3) developmental studies of goal-based imitation learning are proposed for testing predictions of the project?s models in imitation learning experiments with children. The project represents one of the first efforts to develop rigorous probabilistic models of goal-based imitation learning based on insights from human learning. The results are expected to pave the way for a new generation of machines that can interact fluently with humans, learn new skills from human teachers, and cooperatively solve problems with human partners. The project also provides graduate and undergraduate students with multidisciplinary training in computer science and cognitive science, with K-12 outreach activities aimed at encouraging students from underrepresented groups to pursue careers in science and engineering.

人类非常善于通过观察和模仿他人来学习新技能。赋予机器人类似能力的尝试未能推广到具体任务之外，部分原因是重点一直放在跟踪专家展示的动作轨迹上。目前的项目调查了一种新的跨学科的模仿学习方法，这种方法受到人类如何通过基于目标的模仿进行学习的启发。该项目的具体目标包括：(1)一种新的基于推断人类行为潜在目标的模仿方法，而不是跟踪轨迹：基于推断的目标序列执行动作，并将成功执行的动作序列缓存为高层目标，从而产生基于分层目标的模仿；(2)提出一种基于分层贝叶斯模型的跨对象和任务泛化的新方法；(3)提出基于目标的模仿学习的发展研究，以检验该项目的预测-S模型在儿童模仿学习实验中的应用。该项目是根据人类学习的洞察力开发基于目标的模仿学习的严格概率模型的首批努力之一。预计这一结果将为新一代机器铺平道路，这些机器可以流畅地与人类互动，从人类教师那里学习新技能，并与人类合作伙伴合作解决问题。该项目还为研究生和本科生提供计算机科学和认知科学方面的多学科培训，开展K-12外联活动，旨在鼓励来自代表性不足群体的学生从事科学和工程职业。