Learning Representations for Autonomous Mobile Robotics to Enable Complex Tasks

近年来，我们看到了对移动的机器人技术的巨大投资，这被广泛认为是一个具有巨大经济和社会前景的领域。一般人工智能（AI），特别是机器学习（ML）的进步已经产生了影响日常生活的重大实际成果，例如自动人脸识别，语言翻译和有针对性的营销等等。然而，赋予机器人这种能力并将其部署在真实的世界中被证明更具挑战性。但产生社会影响的机会是巨大的。移动的机器人技术将在未来20年内改变社会，就像工业自动化在过去20年里已经改变的那样。我们已经看到移动的机器人技术渗透到消费市场的有限例子，最引人注目的应用是吸尘器和割草机，但大规模的市场仍然开放，如个人交通，货物运输，家庭援助和老年人护理等等。 任何自主移动的机器人的核心能力之一是能够感知世界。这需要将它收集的传感器数据流合并成一个连贯一致的表示。这种表示应该足够丰富，以支持机器人试图完成的任务类型。传统上，这些表示只包括关于世界的几何信息，因此机器人能够实现的任务的复杂性非常有限。这项研究计划的总体目标是开发新的表示，使机器人能够完成更复杂的任务。AI和ML的最新进展显示出令人难以置信的前景。然而，机器人用例具有独特的要求，例如：实时操作，对错误数据和故障的鲁棒性，可扩展性，以及理解机器人出于安全原因采取特定行动的能力。具体来说，在这项工作中，我将研究新的表示，包含更高层次的语义信息和统计的时间变化的世界。我还将特别关注这些算法的可扩展性和实时要求，因为它们与可用的板载资源有关。这将通过利用深度学习以对象检测和语义分割的形式提供数据预处理来实现。此外，计算、内存和带宽等资源的分配必须以最大限度地提高成功实现既定任务的概率的方式进行。要实现这一点，机器人必须安全、坚固和可靠。实现这些目标需要机器人对其周围环境有更深刻的理解，并且它们的算法能够很好地扩展。

We have seen dramatic investments in mobile robotics in recent years, and this is widely considered to be a field of great economic and social promise. The advances of general artificial intelligence (AI), and specifically machine learning (ML) have already yielded significant practical results that are impacting daily life, such as automated face recognition, language translation, and targeted marketing to name a few. However, endowing robots with such capabilities and deploying them in the real world is proving more challenging. But the opportunities for societal impact are significant. Mobile robotics is poised to transform society in the next 20 years in a similar way that industry automation has already in the last 20 years. We have seen limited examples of mobile robotics penetrating the consumer market, most notably applications such as vacuum cleaners and lawn mowers, but massive markets remain open, such as personal transportation, transportation of goods, home assistance and elder care to name a few. One of the core capabilities for any autonomous mobile robot is to be able to perceive the world. This requires amalgamating the stream of sensor data that it is collecting into one coherent and consistent representation. This representation should be rich enough to support the types of tasks that the robot is trying to accomplish. Traditionally, these representations have included little more than geometrical information about the world, and consequently the complexity of the tasks that a robot is capable of achieving have been quite limited. The overall objective of this research program is to develop novel representations that will enable robots to achieve more complex tasks. Recent advances in AI and ML have shown incredible promise. However, the robotics use-case has unique requirements, such as: real-time operation, robustness to incorrect data and failures, scalability, and the ability to understand why a robot took a certain action for safety reasons. Specifically, in this work I will investigate new representations that contain higher level semantic information and statistics about the temporal variability of the world. I will also pay special attention the scalability and real-time requirements of these algorithms as they relate to the available onboard resources. This will be achieved by leveraging deep learning to provide the data-preprocessing in the form of object detections and semantic segmentations. Additionally, the allocation of resources such as computation, memory, and bandwidth, must be done in a way that maximizes the probability of successfully achieving the stated task Robotics has the potential to impact society in a profound way. For this to become a reality requires that robots are safe, robust, and reliable. Achieving these goals requires that robots have a more profound understanding of their surroundings, and that their algorithms are able to scale well.