Group Travel Award for 2017 Workshop on Learning Perception and Control for Autonomous Flight: Safety, Memory, and Efficiency

空中机器人，通常称为无人机，在多种研究、教育、国防和商业应用中提供了前景。 一些例子包括精准农业、航空摄影、敏捷检查和监控以及包裹递送。 在空中机器人开始走出研究实验室并进入现实世界的大多数应用中，机器人操作通常是半自主的。 半自主操作通常假设有用于定位的 GPS 信号，并且至少有一些有关工作环境的先验信息。在未知环境中基于感官的完全自主操作仍主要处于研究阶段。 然而，赋予空中机器人完全自主权可以增强它们对国家教育、经济和国防的影响。 为此，在未知环境中无缝融合机器人自主飞行的感知、规划和控制非常重要。 这可以通过将机器学习工具集成到空中机器人感知和控制中来实现。 深度学习最近成为一种提取自主语义的有前景的方法。 学习自主飞行的感知和控制可以通过用原始传感器观察代替手工设计的地图表示并学习适当的响应来实现。 然而，这并不是一项简单的任务，仍然存在一些挑战。 本次研讨会重点讨论如何 i）最好地整合内存和 ii）为基于学习的系统提供安全保证。 这两个方面对于提高空中机器人在未知环境中自主操作的能力以及推动当前机器人飞行的最先进水平是必要的。 除了机器人飞行领域之外，本次研讨会的成果还涉及赋予能够学习的通用机器人系统自主性，从而帮助自主机器人变得无处不在。本次研讨会的目的是解决所面临的理论和技术挑战，以便赋予基于学习的系统在未知环境中自主操作的能力。 这项工作的关键一步是了解记忆增强型自主学习者如何在可证明的安全保证下进行操作。 因此，研讨会探讨了两个高度相关的问题。 i) 如何从理论上分析基于学习的系统的数据和结构，为安全和任务成功提供保证？ ii) 长期记忆的作用是什么，特别是循环连接或动态外部记忆能否取代全局地图信息？ 该研讨会汇集了来自机器人规划和控制、强化学习和深度学习以及形式化方法的专家，寻求这些问题的答案。该研讨会还邀请相关领域的贡献作者参与。 这些包括但不限于应用深度强化学习来对欠驱动机器人进行基于视觉的控制；学习视觉运动策略并基于神经网络得出学习的正式保证；并开发涉及时间重现和记忆的神经网络架构。 上述问题是在高速空中机器人自主导航的背景下提出的。 然而，它们的范围可以推广到机器人技术的其他领域，需要学习在未知环境中自主操作的感知和控制；例子包括操纵和腿部运动。

Aerial robots, commonly referred to as drones, offer promise in several research, educational, defense and commercial applications. Some examples include precise agriculture, aerial photography, agile inspection and monitoring, and package delivery. In most of those applications that aerial robots have started venturing outside the research lab and into the real world, robot operation is often semi-autonomous. Semi-autonomous operation typically assumes availability of GPS signal for localization, and at least some prior information about the working environment. Sensory-based, fully autonomous operation in unknown environments remains mostly at the research stage. Yet, endowing full autonomy to aerial robots can enhance their impact on the nation's education, economy, and defense. To this end, it is important to seamlessly merge perception, planning, and control for autonomous robotic flight in unknown environments. This can be achieved by integrating machine learning tools into aerial robot perception and control. Deep learning has recently emerged as a promising way to extract semantic meaning for autonomy. Learning perception and control for autonomous flight can be approached by replacing hand-engineered map representations with raw sensor observations, and learning appropriate responses. However, this is not a straightforward task, and several challenges remain. This workshop critically addresses how to i) best incorporate memory and ii) derive safety guarantees for the learning-based system. These two aspects are necessary to improve the capacity of aerial robots to operate autonomously in unknown environments, and to push forward the current state-of-the-art in robotic flight. In addition to the domain of robotic flight, the outcomes of this workshop are relevant to endowing autonomy in general robotic systems that are able to learn, thus helping make autonomous robots ubiquitous.The objective of this workshop is to address the theoretical and technical challenges faced in order to endow learning-based systems with the capacity to operate autonomously in unknown environments. A critical step in this effort is to understand how memory-augmented autonomous learners can operate with provable safety guarantees. The workshop thus examines two highly-relevant questions. i) How to theoretically analyze the data and structure of learning-based systems to provide guarantees on safety and task success? ii) What is the effect of long-term memory and, in particular, can recurrent connections or dynamic external memory replace global map information? The workshop seeks answers to these questions by bringing together experts from robot planning and control, reinforcement learning and deep learning, and formal methods. The workshop also solicits participation of contributed authors working in relevant areas. These include but are not limited to applying deep reinforcement learning for vision-based control of underactuated robots; learning visuomotor policies and deriving formal guarantees for learning based on neural networks; and developing neural network architectures that involve temporal recurrence and memory. The above questions are asked here in the context of high-speed aerial robot autonomous navigation. However, their scope can be generalized to other areas of robotics that learning perception and control for autonomous operation in unknown environments is desirable; examples include manipulation and legged locomotion.