Real-time Machine Intelligence for Assistive Technology

在日常生活中，人类利用一系列细致的感觉和冲动来解释和操纵周围的世界。然而，当身体的某个部分丢失或损坏时，例如，由于事故、疾病或出生并发症，个体与世界互动的能力会显著降低，从而降低生活质量。新的计算方法和辅助技术对于增强和恢复受伤者的感觉和运动能力至关重要。我提出的研究计划侧重于这一主题的一个具体例子：使用机器智能来创建更丰富、更强大的界面，以弥合人类与固定在其身体上的机器人假肢之间的差距。这类辅助机器人的设计者面临着许多挑战，其中最重要的可能是开发自然，灵活的控制和通信方法，将用户连接到他们的辅助设备。很明显，传统的控制和通信方法在面对越来越多的多样化、数据密集的传感器和致动技术时将变得不够。文献还表明，假肢需要自适应控制系统，即适应每个人的独特因素以及用户及其环境中发生的日常变化的界面。考虑到这一点，我提出的研究计划将研究，开发和评估机器智能方法，以加速下一代假肢技术的进展。除了研究机器智能和人机交互的核心原则外，我的工作还将把新算法的进步转化为实际应用，并评估它们在实际部署中的优势和劣势。我将特别关注预测学习、控制学习和元学习的计算效率方法，这些方法适合在可穿戴机器人中持续使用。 因此，我将采用强化学习领域的技术;我的工作将以通过广义值函数进行的时间扩展预测和行动者-批评者策略梯度方法为基础。该研究计划的最终目标是创建一个完整的表示，预测，控制和元学习系统，使人类和他们的辅助技术之间能够无缝协作。作为我提出的研究计划的积极组成部分，将显着提高我目前的学生和那些在程序的过程中加入的学术发展和职业潜力。此外，该计划产生的平台技术将直接促进强化学习理论和实践的进步，并将很容易转移到医疗和商业领域。

During every-day life, humans leverage a detailed stream of sensation and actuation to interpret and manipulate the world around them. However, when a body part is lost or damaged-for example, through accident, illness, or birth complications-an individual can suffer a significant reduction in their ability to interact with the world and, as a consequence, reduced quality of life. New computing methods and assistive technologies are essential to augmenting and restoring an injured individual's sensory and motor abilities. My proposed research program focuses on a concrete example of this theme: using machine intelligence to create richer and more powerful interfaces to bridge the gap between a human and a robotic artificial limb that is affixed to their body. Designers of assistive robots of this kind face a number of challenges, perhaps the most significant of which is the development of natural, flexible control and communication methods that connect users to their assistive devices. It is apparent that traditional approaches to control and communication will become insufficient when faced with an increasing repertoire of diverse, data-dense sensor and actuation technologies. The literature also indicates that artificial limbs are in need of adaptive control systems-interfaces that adapt to both the unique factors in each individual and the day-to-day changes that occur in the user and their environment.***With this in mind, my proposed program of research will study, develop, and evaluate machine intelligence methods that accelerate progress on next-generation artificial limb technologies. In addition to studying core principles of machine intelligence and human-machine interaction, my proposed work will translate new algorithmic advances to practical use and assess their strengths and weaknesses during real-world deployment. I will focus specifically on computationally efficient approaches to prediction learning, control learning, and meta-learning that are suitable for ongoing use within a wearable robot. As such, I will employ techniques from the area of reinforcement learning; my work will have its foundation in temporally extended predictions made via generalized value functions, and in actor-critic policy gradient methods. The ultimate goal of this program of research is to create a complete system of representation, prediction, control, and meta-learning that enables seamless collaboration between a human and their assistive technology. Being an active part of my proposed program of research will significantly enhance the academic development and career potential for my current students and those that join over the course of the program. Furthermore, platform technologies generated by this program will contribute directly to progress in the theory and practice of reinforcement learning, and will be readily transferrable to medical and commercial domains.**