S&AS: FND: Reflective Learning of Stochastic Physical Models for Robust Manipulation

为了让机器人在日常生活环境中发挥作用，从灵活的制造和仓库领域到家庭，它们需要自主地掌握和操纵各种各样的潜在未知物体。目前，自主机器人实际上无法在高度控制的环境中工作，在这种环境中，每个物体都提供了精确的模型。这种限制部分是由于缺乏强大的算法来抓取和操纵具有未知几何或机械特性的物体。拟议中的项目将对强大的机器人操作进行基础研究，以使自主机器人能够长时间有效地与各种日常物理对象进行交互。目标是让自主机器人能够有效地从经验中学习物体之间以及与机械臂之间的物理相互作用。下一步是利用这种经验来执行健壮的操作任务。有许多典型的机器人任务可以从提出的改进中受益，这将构成项目实验过程的基础。它们包括将物体推到所需的姿势，重新配置物体以简化它们的拾取和工具的处理。该项目开发了三个关键部分：(1)通过观察未知物体在机器人操作时的运动方式来学习未知物体的惯性、弹性和摩擦特性的算法。该项目将研究用于黑盒系统识别的新颖贝叶斯优化技术，以学习对象的概率模型。(2)一个物理逼真的模拟器，在给定第一个组件学习到的物理参数的情况下，可以提供物体运动的随机模型。这将通过利用在线非参数学习方法来加速不确定性下的物理逼真模拟来实现。(3)一种鲁棒规划算法，该算法利用模拟器在给定学习到的随机模型上寻找最优动作。目标是随着计算时间的增加和机器人对环境中物体的经验的增加，收敛到越来越鲁棒的解决方案。为了加强项目的广泛影响，pi将以开源软件包的形式向研究社区提供其解决方案的实现。这将与教育材料的生成相结合，旨在吸引本科学生在学习早期学习STEM。pi还将组织学术会议，将基础领域的研究人员、机器人专家和行业代表聚集在一起。

In order for robots to function in everyday life environments, from flexible manufacturing and warehouse domains to households, they need to autonomously grasp and manipulate a wide variety of potentially unknown objects. Currently, autonomous robots are practically unable to work outside highly-controlled environments wherein an accurate model of every object is provided. This limitation is partially due to the lack of robust algorithms for grasping and manipulating objects with unknown geometric or mechanical properties. The proposed project will perform fundamental research into robust robotic manipulation in a way that will enable autonomous robots to interact efficiently with a large variety of everyday physical objects for extended periods of time. The objective is for autonomous robotic manipulators to effectively learn from experience how objects may physically interact with each other and with the robotic arm. The next step is to utilize this experience so as to perform robust manipulation tasks. There are many exemplary robotic tasks that can be benefited from the proposed improvements and which will form the basis of the project's experimentation process. They include the pushing of objects to desired poses, reconfiguration of objects to simplify their picking and the handling of tools. The project develops three key components: (1) An algorithm for learning inertial, elastic, and friction properties of an unknown object by observing how the object moves when manipulated by a robot. The project will research novel Bayesian optimization techniques for black-box system identification in order to learn probabilistic models of objects. (2) A physically realistic simulator that can provide a stochastic model of an object's motion given the physical parameters learned by the first component. This will be achieved by utilizing online non-parametric learning methods for speeding up physically realistic simulations under uncertainty. (3), A robust planning algorithm that utilizes the simulator for finding optimal actions to apply on the object given the learned stochastic model. The objective is to converge to increasingly robust solutions as computation time increases and the robot acquires increased experience with objects in an environment. To strengthen the project's broader impact, the PIs will provide implementations of their solutions to the research community as open-source software packages. This will be coupled with the generation of educational material, which will aim to attract undergraduate students early in their studies to STEM. The PIs will also aim to organize academic meetings that will bring together researchers from foundational domains, robotics experts and industry representatives.