CAREER:Designing Robots that Learn: Closing the Gap Between Machine Learning and Engineering

当机器人与其环境的相互作用可以被精确定义时，机器人学就获得了巨大的成功。然而，如果机器人用来预测、计划和控制其行为的模型不准确，就可能导致次优甚至危险的行为。不幸的是，随着机器人及其环境变得越来越复杂，准确地指定机器人的行为变得越来越困难。另一种方法是从经验中学习模型，但大多数这种方法在各种情况下需要大量数据，而收集这些数据实际上是不可行的。这个项目试图结合手工制作的、基于物理的模型和机器学习算法的优势来解决这些问题。这种结合的方法将使工程师能够更好地设计能够在结构较少的现实世界环境中运行的机器人。该项目还通过提供结合机器学习和机器人学的跨学科体验，实现了与这一愿景相关的教育目标。这个项目的目标是开发新的理论和算法，以弥合机器学习和机器人学的工程方法之间的差距，这些方法传统上是单独研究的。工程学使用物理知识来提供可解释性、透明度，并保证工程系统的可靠性和健壮性，而机器学习研究数据和信息，并提供侧重于模型的可表现性、计算成本和学习算法的样本效率的保证。目前的项目包括三项研究倡议，以使这些学科更紧密地联系在一起。第一个目标是为机器人开发结合参数物理模型和非参数统计模型的新的半参数模型。第二种方法将解决在给定关于动力学和状态的不完全信息时，从数据中学习状态空间模型的问题。第三个奖项将研究如何使用工程学中的结构知识来约束非参数学习算法和深度神经网络模型的假设空间。这一奖项反映了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。

Robotics has enjoyed great success when a robot's interaction with its environment can be precisely defined. However, if the models that robots use to predict, plan, and control their behaviors are inaccurate, it can lead to suboptimal or even dangerous behaviors. Unfortunately, as robots and their environments become more complex, it is increasingly difficult to accurately specify robot behavior. An alternate approach is to learn the models from experience, but most such approaches need large amounts of data, in a variety of situations, which is practically infeasible to collect. This project attempts to combine the strengths of hand-crafted, physics-based models and machine learning algorithms to tackle such problems. Such a combined approach will better position engineers to design robots that can operate in less-structured real-world environments. The project also addresses educational goals related to this vision by providing cross-disciplinary experiences that combine machine learning and robotics. The goal of this project is to develop new theory and algorithms that close the gap between machine learning and engineering approaches to robotics, which have traditionally been studied separately. While engineering uses physics knowledge to provide interpretability, transparency, and guarantees about the reliability and robustness of engineered systems, machine learning studies data and information, and provides guarantees that focus on the expressivity of models, computational cost, and sample efficiency of learning algorithms. The current project consists of three research initiatives to bring these disciplines closer together. The first aims to develop new semi-parametric models for robotics that combine parametric physical models and non-parametric statistical models. The second will tackle the problem of learning state space models from data when given incomplete information about dynamics and state. The third will investigate how structural knowledge from engineering can be used to constrain the hypothesis space of nonparametric learning algorithms and deep neural network models.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.