Learning From Demonstration

该研究的目标是开发新的机器人学习算法，用于比迄今为止尝试的更复杂的机器人和任务，并大大提高实现机器人学习的自动化水平。研究将集中在：1：具有多自由度的任务和机器人；2：让一个机器人学习执行多个任务，并在任务之间进行适当的推广；3：长时间跨度的学习，其中机器人，任务和环境都在变化；4：将多种方法结合起来进行机器人学习。该研究将有助于实现任务结构、表征、可调参数和函数的自动选择。将强调两种类型的学习：从演示中学习，机器人从如何执行任务的演示中学习，以及强化学习，机器人通过优化奖励函数来学习。强调知识表示的灵活方法，包括局部加权学习和其他非参数学习技术。在实现从演示中学习的过程中开发的技术将成为解决更一般学习问题（如强化学习）方法的基础。这项研究将与日本先进电信研究人类信息处理实验室的Mitsuo Kawato博士合作进行。这项研究的预期意义在于，它将使机器人和基于机器的系统的编程变得更容易和更便宜。从工程的角度来看，目标是减少机器人编程中昂贵的专家人力投入。从心理学的角度来看，目标是了解人们是如何学习的，这种工作导致了学习行为如何完成的模型。

The goals of the research are to develop new robot learning algorithms for qualitatively more complex robots and tasks than have been attempted so far, and to greatly increase the level of automation of implementing robot learning. The research will focus on: 1: tasks and robots with many degrees of freedom, 2: having one robot learn to perform multiple tasks, and generalize appropriately between tasks, 3: learning over a long time span, in which the robot, the task, and the environment change, and 4: combining multiple approaches to robot learning. The research will contribute to a more automatic process for selecting task structure, representations, and adjustable parameters and functions. Two types of learning will be emphasized: learning from demonstration, where the robot learns from a demonstration of how to perform a task, and reinforcement learning, where the robot learns by optimizing a reward function. Flexible methods to represent knowledge will be emphasized, including locally weighted learning and other nonparametric learning techniques. The techniques developed in implementing learning from demonstration will form the basis of the approach to more general learning problems, such as reinforcement learning. This research will be conducted in collaboration with Dr. Mitsuo Kawato at the Advanced Telecommunications Research Human Information Processing Laboratory in Japan. The expected significance of this research is that it will make it easier and less expensive to program robots and machine-based systems in general. From an engineering point of view the goal is to reduce the amount of expensive expert human input into robot programming. From a psychological point of view the goal is to understand how people learn, and this kind of work leads to models of how learning behavior might be accomplished.