Learning Adaptive Sensorimotor Representations

任何智能物理系统的核心能力都是能够在广泛的环境中成功运行，并自主适应任务条件的变化。目前最先进的机器人感知和控制系统越来越能够克服具有挑战性的环境因素，通常是通过使用“大数据”和机器学习技术，但这些系统需要广泛的人类工程和数据注释，而且通常不够灵活，如果没有人类设计者的动手交互，就不能克服任务变化。 我的研究将集中在我认为是智能机器人的核心挑战--自动学习将机器人的传感器和执行器连接起来的表示，以实现感知和控制任务的自适应解决方案。我相信，这可以通过结合学习视觉表征(例如，现在用于识别对象的深度表征)和自适应控制(例如，现在用于学习运动技能的策略梯度方法)的最先进技术来实现。这一研究方向的关键假设是，自动化的中间表征将比人工设计者选择的更灵活地捕捉知觉和控制之间的关联。如果研究成功，关键的优势将是机器人需要极大地减少安装和培训才能成功执行新任务。 这项工作的一个鼓舞人心的例子是一个自主机器人厨师。今天，我们可以在一个特别设计的厨房里可行地制造机器人来烹饪小种类的食物。人们可以访问日本来寻找这种技术的例子。然而，要制造出一种甚至可以在以前从未见过的厨房里烹饪一顿饭的机器人，需要付出巨大的努力。主要的挑战将是机器人无法将其知识概括为物品放置及其属性的不同(例如，我的刀略小，放在不同的抽屉里)。 这项研究计划将提高机器人处理新环境的能力，减少人类工程师的工作量。除了烹饪，我们还将针对社会应用，如灾难救援，以及工业，如采矿和农业。

A core competency of any intelligent physical system is its ability to operate successfully in a wide range of environments and to adapt autonomously to changes in task conditions. Current state-of-the-art systems for robot perception and control are increasingly able to overcome challenging environmental factors, typically through the use of "big-data" and machine learning techniques, but these systems require extensive human engineering and data annotation and are typically not flexible enough to overcome task variations without hands-on interaction from their human designer. My research will focus on what I believe to be a core challenge of intelligent robotics - the automated learning of representations that link a robot's sensors and actuators to achieve adaptive solutions for perception and control tasks. I believe this can be accomplished by combining the state-of-the-art in learning visual representations (e.g., deep representations now used for recognizing objects) and in adaptive control (e.g., policy-gradient approaches now used to learn motor skills). The key assumption of this research direction is that an automated intermediate representation will more flexibly capture the correlations between perception and control than those chosen by a human designer. The key advantage, if the research is successful, will be robots that require drastically reduced setup and training to perform new tasks successfully. A motivational example for this work is an autonomous robotic chef. Today, we can feasibly produce robots to cook a small variety of meals in a specially-designed kitchen. One can visit Japan to find examples of this technology. However, it would require enormous effort to produce a robot that could cook even a single meal in a kitchen it had not seen before. The main challenge would be the robot's inability to generalize its knowledge to differences in the placement of items and their properties (e.g., my knife is slightly smaller and is kept in a different drawer). This research program will improve a robot's ability to handle new environments with a reduced amount of effort from human engineers. Beyond cooking, we will target applications to society, such as disaster rescue, and industry, such as mining and agriculture.