Using stochastic optimal feedback control and computational motor control to design personalized and adaptive human robot interfaces

从智能手机到智能汽车，从外骨骼到肌电接口，人机接口和人机接口无处不在。随着这些接口变得更加复杂，工程师更难将接口的可调参数与人类的复杂目标、奖励和动态相匹配--特别是对于随机(受噪声破坏的)接口。人体运动理论的最新进展已经在输入参数(如运动和力)和人们关心的事物(如准确性、努力和响应性)之间产生了简单的因果映射。现有的人机界面设计通常只关注一个方面(如最大限度地提高准确性)，并没有利用我们对这些动态映射的理解来最大化最终用户所重视的回报。我的长期目标是颠覆人机界面设计的范式，这样工程师就可以使用这些动态映射在最终用户关心的内容和工程师可以调整的参数之间建立一座桥梁。我的短期目标是：1.开发因果的、高效的、类似人类的模型，并包含由信号损坏引起的人机界面不确定性。我们将使用高效的数值方法将计算电机控制领域的各个分支的先进技术融合在一起。2.完善计算机电机控制实验。我们的初步工作表明，一些传统的方法由于用户潜意识的适应而导致偏差，可以通过改进实验技术和使用随机过程分析技术来缓解。3.开发人机界面个性化调谐平台。我们将开发数字高效的技术，为给定的奖励找到最佳映射，允许用户通过阐明他们的相对奖励偏好来安全地实时调整参数。4.将我们的方法扩展到具有长期目标的任务，如教练。类似于国际象棋算法如何看得足够远才能赢得一场比赛，我们将使用最近的进展，在选择训练动作时隐含地将潜在的学习收益考虑在内。这种方法将使教练能够选择针对个人的长期动态和隐含目标量身定做的训练行动。我们的每个目标都将通过实验来验证，将我们的解决方案与传统技术进行比较，使用适当的统计设计并考虑EDI因素。这一广泛的计划利用人体运动科学来为与人类互动的设备的最佳设计提供信息。它代表了我们对设计思维方式的高度原创性范式转变，将有助于在智能手机、人机接口和外骨骼等辅助/增强设备等多个领域取得突破性进展，并产生具体技术，满足关键的社会经济需求，使设备的复杂性与个人的复杂目标相协调。

Human-computer interfaces and human-robot interfaces are everywhere, ranging from smart-phones to intelligent cars to exoskeletons to myoelectric interfaces. As these interfaces become more sophisticated, it becomes harder for engineers to match the tunable parameters of the interface to the complex goals, rewards, and dynamics of humans-particularly for stochastic (noise-corrupted) interfaces. Recent advances in human-movement theory have produced simple causal mappings between input parameters (like motion and force) and the things people care about (like accuracy, effort, and responsiveness). Existing human-machine interface designs typically only focus on one aspect (such as maximizing accuracy) and do not leverage our understanding of these dynamical mappings to maximize the rewards that end-users value. My long-term objective is to flip the paradigm of human-machine interface design, such that engineers use these dynamical mappings to form a bridge between what end-users care about and the parameters that engineers can tune. My short-term objectives are: 1.Develop models that are causal, efficient, human-like, and incorporate human-machine interface uncertainty caused by signal corruption. We will merge advances across branches of the field of computational motor control using numerically efficient methods. 2.Refine computational motor control experiments. Our preliminary work suggests that some conventional methods result in bias due to subconscious user adaptation, which can be mitigated by refining experimental techniques and using random-process analytical techniques. 3.Develop a platform that enables personalized tuning of human-machine interfaces. We will develop numerically efficient techniques that find optimal mappings for given rewards, allowing the user to safely adjust parameters in real-time by articulating their relative reward preferences. 4.Extend our approach to tasks with long-term goals, such as coaching. Similar to how chess algorithms look far enough ahead to win a game, we will use recent advances that implicitly factor in potential learning gains when choosing training actions. This approach will enable coaches to choose training actions that are tailored to the long-term dynamics and implicit goals of individuals. Each of our aims will be validated across experiments comparing our solutions to conventional techniques, using appropriate statistical design and considering EDI factors. This broad program leverages human movement science to inform the optimal design of devices that interact with humans. It represents a highly original paradigm shift in how we think about design that will contribute to groundbreaking advances in several fields including smartphones, human-robot interfaces, and assistive/augmenting devices such as exoskeletons, and lead to concrete technologies that address the critical socio-economic need to harmonize the sophistication of devices with the complex goals of individuals.