Skill Learning for Humanoid Robots

目前，人形机器人并不像人类那样移动。它们不够熟练，不能执行需要与人类和环境互动的任务。通过刻意的练习，人类可以非常容易地学习新的技能。如果类人机器人能够获得像人类一样的技能，它们将能够帮助满足社会的许多需求。人类在运动技能方面表现出非凡的灵活性。我们认为，人类的灵活性源于人类在表现上不一致的事实。这种不一致让人们可以灵活地学习新技能。如果仿人机器人有能力获得像人类一样的技能，我们可以期待仿人机器人在各种任务中的表现得到显著提高。该方案从运动学的角度，由机器人学研究人员和运动学习与控制研究人员进行独特的合作，旨在捕捉人类运动学习和控制的这些特征，然后在仿人机器人的协调和控制算法中例示这些特征。此外，与日本国家先进工业科学技术研究所(AIST)的研究人员进行的国际合作将使我们能够在HRP-2类人机器人上进行实验研究。在这个项目中，我们将研究人类受试者在学习一些复杂任务时的运动。这些受试者要么作为一个整体练习技能，要么练习技能的个别部分。然后，我们将观察受试者如何很好地转移到一种新的运动技能。我们将使用网络模型对正在学习的人类进行建模，然后尝试将这些模型放入仿人机器人中。人形机器人会表现出和人类一样的学习迁移吗？人形机器人能像人类一样表现出同样类型的错误吗？这项研究的更广泛的影响包括：(I)系统地调查运动技能的学习和迁移，它们的局限性，以及它们的整合，以生产智能人形机器人；(Ii)工程和文科研究人员从不同角度从事信息技术研究和教育的研究合作，以及与日本AIST的国际研究合作；(Iii)开发一个网站，供学生使用OpenHRP软件模拟人形机器人；以及(Iv)研究成果将在专业会议和档案期刊上传播。此外，该项目将对本科生和研究生教育产生影响，特别是实验研究和高级设计项目。最后，为了扩大影响，人形机器人和模拟工具将成为教育高中生团队合作的极好工具，并鼓励他们为自己的高等教育和职业选择工程学。

Currently humanoid robots do not move like human beings. They are not skillful enough to perform tasks that require interactions with humans and the environment. Humans can learn new skills very easily with deliberate practice. If humanoid robots could acquire skills like humans, they would be able to help with many needs of society.Humans exhibit remarkable flexibility in motor skill. We believe that human flexibility is derived from the fact that humans are inconsistent in performance. This inconsistency allows people the flexibility to learn new skills. Should humanoid robots have the ability to acquire skills like humans, we can expect a dramatic improvement in the performance of humanoid robots on various tasks.This proposal, a unique collaboration between a robotics researcher and a motor learning and control researcher from a kinesiology perspective, aims at capturing these characteristics of human motor learning and control and then instantiating these characteristics in the coordination and control algorithms of a humanoid robot. In addition, an international collaboration with researchers at the National Institute of Advanced Industrial Science and Technology (AIST) in Japan will allow us to examine the goodness of our research experimentally on an HRP-2 humanoid robot.In this project, we will be studying the motion of human subjects as they learn a few complicated tasks. These subjects will either practice the skill as a whole or practice the individual parts of the skill. We then will observe how well subjects transfer to a novel motor skill. We will model the learning human with network models and then attempt to put these models in a humanoid robot. Will the humanoid robot show the same transfer of learning as the human? Can the humanoid robot show the same types of errors as the human? These are some of the questions that we seek to answer from this research.The broader impacts of this research include: (i) a systematic investigation of motor skill learning and transfer, their limitations, and their integration to produce intelligent humanoid robots; (ii) research collaboration between researchers from engineering and liberal arts to engage in information technology research and education from different perspectives, and international research collaboration with the AIST in Japan; (iii) development of a web site for students to simulate humanoid robots using OpenHRP software; and (iv) research results will be disseminated in professional conference and archival journal publications. In addition, the project will have an impact on undergraduate and graduate education, especially experimental research and senior design projects. Finally, for the outreach impact, humanoid robots and simulation tools will be an excellent vehicle to educate high-school students in team work and encourage them to select engineering for their higher education and career.