The Physical Exploration Challenge: Robots Learning to Discover, Actuate, and Explore Degreesof Freedom in the World

该项目将为现实世界的机器人系统配备智能最有趣的方面之一：内部学习驱动力，即，表现出最大化学习进度的行为的能力。具体地说，我们将设计机器人系统，通过推拉和扭曲看似有趣的部件来物理地探索它们的环境，目的是学习如何发现，驱动和探索世界上的自由度（DoF）。我们的方法受到最先进的机器学习探索和主动学习研究的启发，特别是优化策略以获得预期信息的概念。 然而，我们的真实探索场景提出了一些在现有研究中尚未解决的基本问题。三个开放的问题定义了这个建议的主要研究领域：（1）~我们如何能够制定一个听话的信念在运动学结构，表示相关方面的相关不确定性，如DoF的存在，他们的属性，和他们的关系？什么方法可以得到信息增益驱动的探索strategies为这些表示，鉴于感知和行动的代理是受不确定性和不断改进的经验？ (2)~什么感知方法对识别和确定自由度有用，对形成关于有希望的相互作用点的假设有用，以及对将场景分割成刚体假设有用？(3)~我们如何在探索的背景下，为成功的交互和自由度操作的目标，参数化和优化运动连续性？拟议的研究解决了机器学习和机器人技术交叉领域的一个根本挑战。如果成功的话，我们相信这将在机器人系统的行为方式上，在它们的自主学习上，在它们获得知识的方式上，以及最终在它们与人类互动的方式上，在工业和私营部门的广泛应用中发挥作用。

This project will equip real-world robotic systems with one of themost interesting aspects of intelligence: an internal drive to learn,i.e., the ability to exhibit behavior that maximizes learning progresstowards an objective. Concretely, we will design robotic systems thatphysically explore their environment by pushing, pulling, and twistingseemingly interesting parts, with the goal of learning how todiscover, actuate, and explore degrees of freedom (DoF) in the world.Our approach is informed by state-of-the-art Machine Learning researchon exploration and active learning, especially the concept ofoptimizing policies for expected information gain. Our real-worldexploration scenario, however, raises fundamental issues that have notpreviously been addressed in existing research. Three open questions define the major research areas of this proposal:(1)~How can we formulate a tractable belief over kinematic structuresthat represents the relevant uncertainties over relevant aspects, suchas the existence of DoF, their properties, and their relationships?What methods can derive information-gain driven exploration strategiesfor these representations, given that perception and actions of theagent are subject to uncertainty and are continuously improved fromexperience? (2)~What perception methods are useful for discoveringand identifying DoF, for forming hypotheses about promisinginteraction points, and for segmentating the scene into rigid bodyhypotheses? (3)~How can we parameterize and optimize motion primitivesfor the objective of successful interaction and manipulation ofdegrees of freedom in the context of exploration?The proposed research addresses a fundamental challenge in theintersection of machine learning and robotics. If successful, webelieve that this will make a transitional change in the way roboticsystems behave, in their autonomy of learning, in the way they groundacquired knowledge, and eventually also in the way they will interactwith humans and play their part in a wide range of applications inindustry and the private sector.