CAREER: Learning Symbolic Representations for Robot Manipulation

近年来，通用机器人硬件的质量和成本都有了戏剧性的提高。然而，对硬件进行编程以解决任何重要任务是极其困难的。如果机器人能够计划自己达到用户指定的目标，而不需要非常详细的编程，那将是更可取的。这里的一个关键挑战是处理感知和感知的低级别细节，同时还在高级别上对要完成的任务进行推理。该项目旨在开发一个框架，允许机器人学习如何操纵对象，如何有效地抽象地表示这些对象，以及如何对看起来不同但具有相同功能的对象进行泛化(例如，不同的微波)。该项目将开发新的算法，使机器人能够在现实世界中的复杂场景中进行推理和规划；因此，它可以大大加快复杂机器人在家庭、医院、轻工制造设施和空间等半结构化环境中的部署。该项目旨在使机器人能够自主学习可重复使用的以对象为中心的运动技能以及支持使用这些技能进行规划的便携式符号表示法。学习运动技能来操纵世界上的物体，并学习抽象的表征来推理世界上的物体-同时对类似功能的物体进行泛化-将使机器人能够产生智能的、以目标为导向的移动操纵行为。该项目将1)设计实用的算法，通过与物体的互动来发现操纵物体的运动技能，2)设计在外观不同但功能相似的物体上推广这些技能的算法，3)开发一个理论上合理的框架，用于学习以对象为中心的抽象表征，支持使用这些技能进行目标导向的规划，并展示其在移动操纵机器人上的应用。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。

Recent years have seen a dramatic improvement in the quality and cost of general-purpose robot hardware. However, programming that hardware to solve any non-trivial task is extremely hard. It would be far preferable if robots could plan to reach user-specified goals on their own, without requiring highly detailed programming. A key challenge here is dealing with the low-level details of sensing and perception, while also reasoning at a high-level about the task to be completed. This project aims to develop a framework that allows robots to learn how to manipulate objects, how to usefully represent those objects abstractly, and how to generalize across objects that appear different but have the same functionality (e.g., different microwaves). This project will develop new algorithms that will enable robots to reason and plan in complex scenarios in the real world; it could therefore substantially accelerate the deployment of complex robots in semi-structured environments like the home, hospitals, light manufacturing facilities, and space. This project aims to enable robots to autonomously learn reusable object-centric motor skills and the portable symbolic representations that support planning with those skills. Learning motor skills to manipulate, and abstract representations to reason about, objects in the world---while generalizing across objects of similar functionality---will enable robots to generate intelligent, goal-directed mobile manipulation behavior. The project will 1) design practical algorithms that discover motor skills for manipulating objects by interacting with them, 2) design algorithms for generalizing those skills across objects with different appearances but similar functionality, and 3) develop a theoretically sound framework for learning object-centric abstract representations that support goal-directed planning using those skills, and demonstrate its use on a mobile manipulation robot.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.