Humanlike physics understanding for autonomous robots

在凌乱的冰箱里，你如何拿起一瓶牛奶，它躺在一些酸奶罐后面？当人类能够使用视觉信息来计划和选择与外部物体的这种熟练动作时，非常容易和快速-这是物种历史上获得的一种能力，并且随着儿童的发展而发展- *机器人挣扎*。事实上，虽然人工智能在国际象棋和围棋等任务上取得了巨大的飞跃，但今天机器人技术的计划和执行能力却被普通的幼儿所超越。考虑到我们所处的复杂而不可预测的世界，这些看似微不足道的任务是高度复杂的神经计算的产物，这些神经计算可以概括和适应不断变化的情况：不断地在多个目标和行动选项之间进行选择。我们的目标是研究如何将这样的计算转移到机器人身上，使它们能够更有效地操作物体，以一种比目前更像人类的方式，并能够执行目前超越技术水平的操作。让我们回到冰箱的例子：你首先需要决定最好移开哪个酸奶罐，以便接触到牛奶瓶，然后做出适当的动作来安全地抓住罐子——这是“接触前”的抓握阶段。然后你需要决定对锅施加什么类型的力（将它向左或向右推，轻推它或可能将它抬起并将锅放在另一个架子上等），即“接触”阶段。虽然这些步骤以快速和自动化的方式实时发生，但我们将在实验室控制的情况下探索这些过程，系统地检查抓取的接触前和接触阶段，以确定哪些因素（空间位置，锅的大小，锅的质地等）使人类倾向于选择一个动作（或一系列动作）而不是其他可能性。我们假设我们可以提取一组高层次的规则，使用定性的时空形式来表达，可以捕捉到这种专业知识的本质，并结合更多定量的低级表征和推理。我们将开发一个计算模型，为测试影响行为的因素的假设提供正式的基础，并最终使用该模型来预测在这种情况下最可能发生的行为，以响应给定的感知（视觉）输入。我们认为，对人类如何执行这些动作的计算理解可以弥合机器人与人类的技能差距。最先进的机器人运动/操纵规划者使用概率方法（随机抽样，例如RRTs， PRMs，是当今该领域的主要运动规划方法）。因此，规划者无法解释他们的决定，类似于电话会议中提到的“黑匣子”机器学习方法，产生难以理解的模型。然而，如果机器人可以与世界产生类似人类的互动，如果它们可以利用人类行动选择的知识进行规划，那么这将允许机器人解释为什么它们以特定的方式执行操作，并且还可以促进“易读操作”-即人类可以预测的动作，因为它与人类的行为方式密切相关，这是机器人社区最近一些研究的目标。这项工作将阐明在行动控制中使用感知信息——这是一个非常有学术兴趣的话题，同时与机器人专家面临的许多实际问题直接相关，这些问题是为了控制在混乱环境中工作的机器人：从机器人在仓库里挑选物品，到需要区分健康组织和癌组织的新手术技术。

How do you grasp a bottle of milk, nestling behind some yoghurt pots, within a cluttered fridge? Whilst humans are able to use visual information to plan and select such skilled actions with external objects with great ease and rapidity - a facility acquired in the history of the species and as a child develops - *robots struggle*. Indeed, whilst artificial intelligence has made great leaps in beating the best of humanity in tasks such as chess and Go, the planning and execution abilities of today's robotic technology is trumped by the average toddler. Given the complex and unpredictable world within which we find ourselves situated, these apparently trivial tasks are the product of highly sophisticated neural computations that generalise and adapt to changing situations: continually engaging in a process of selecting between multiple goals and action options. Our aim is to investigate how such computations could be transferred to robots to enable them to manipulate objects more efficiently, in a more human-like way than is presently the case, and to be able to perform manipulation presently beyond the state of the art.Let us return to the fridge example: You need to first decide what yoghurt pot is best to remove to allow access to the milk bottle and then generate the appropriate movements to grasp the pot safely- the *pre-contact *phase of prehension. You then need to decide what type of forces to apply to the pot (push it to the left or the right, nudge it or possibly lift it up and place the pot on another shelf etc) i.e. the *contact* phase. Whilst these steps happen with speed and automaticity in real time, we will probe these processes in laboratory controlled situations to systematically examine the pre-contact and contact phases of prehension to determine what factors (spatial position, size of pot, texture of pot etc) bias humans to choose one action (or series of actions) over other possibilities. We hypothesise that we can extract a set of high level rules, expressed using qualitative spatio-temporal formalisms which can capture the essence of such expertise, in combination with more quantitative lower-level representations and reasoning. We will develop a computational model to provide a formal foundation for testing hypotheses about the factors biasing behaviour and ultimately use this model to predict the behaviour that will most probably occur in response to a given perceptual (visual) input in this context. We reason that a computational understanding of how humans perform these actions can bridge the robot-human skill gap. State-of-the-art robot motion/manipulation planners use probabilistic methods (random sampling e.g. RRTs, PRMs, is the dominant motion planning approach in the field today). Hence, planners are not able to explain their decisions, similar to the "black box" machine learning methods mentioned in the call which produce inscrutable models. However, if robots can generate human-like interactions with the world, and if they can use knowledge of human action selection for planning, then this would allow robots to explain why they perform manipulations in a particular way, and also facilitate "legible manipulation" - i.e. action which is predictable by humans since it closely corresponds to how humans would behave, a goal of some recent research in the robotics community. The work will shed light on the use of perceptual information in the control of action - a topic of great academic interest and simultaneously have direct relevance to a number of practical problems facing roboticists seeking to control robots working in cluttered environments: from a robot picking items in a warehouse, to novel surgical technologies requiring discrimination between healthy and cancerous tissue.