CAREER: Robust Bipedal Locomotion in Real-World Environments

本学院早期职业发展（Career）计划的目标是开发用于分析和优化人类和类人装置高维模型的准周期双足步态的工具。从观看奥运会运动员到沿着泥泞的小路徒步旅行，我们的经验表明，双足运动提供了一种高度敏捷和强大的移动方式。然而，今天的人形机器人远不如他们虚构的好莱坞同行。倒立摆动力学使直立行走在期望的控制输入下具有高度可操作性，但也使其极易受到不稳定的影响。不像经典的倒立摆问题（例如，一个由底部推进器推动的火箭在连续飞行中保持稳定），行走的过程因脉冲脚步的不连续性而变得复杂，这些脚步在宽度和高度上都有变化，导致“准周期性”步态。特别是，工作重点是在不平坦的地形上行走的挑战，包括间歇性的障碍。该方法利用了这样一种观察，即关节的位置和速度状态的逐步“快照”往往位于此类系统的高维状态空间中的二维流形上。这种降维可以使用机器学习技术进行控制策略评估和改进，以定量地估计两足行走者和随机地形的给定组合的坠落率、能量使用和速度。如果成功，这项工作将为分析和/或优化各种现实世界的系统提供重要的工具。从评估步态受损的中风幸存者的跌倒风险，到为受伤退伍军人设计智能下肢假肢，所开发的建模方法将为具有如此高维数和复杂性的系统提供量化可靠性的手段，而传统的稳定性保证是无法做到的。此外，机器人研究为K-12和大学STEM教育提供了一个天然的门户。该项目包括几个外展元素来鼓励这种兴趣，从为当地FIRST机器人参与者和主要是西班牙裔小学生举办校园实地考察，到赞助加州大学圣巴巴拉分校的新机器人俱乐部。

The objective of this Faculty Early Career Development (CAREER) Program grant is to develop tools for analyzing and optimizing quasi-periodic biped gaits for high-dimensional models of both humans and humanoid devices. From watching Olympic athletes to taking a hike along a dirt path, our experiences demonstrate that biped locomotion provides a highly agile and robust means of mobility. However, today's humanoid robots are far less capable than their fictional, Hollywood counterparts. The same inverted-pendulum dynamics that make upright walking highly maneuverable under desired control inputs also make it highly susceptible to destabilization. Unlike classic inverted pendulum problems (for example, a rocket propelled by thrusters at its base stabilized during continuous flight), the process of walking is complicated by the discontinuities of impulsive footsteps that vary in both width and height, resulting in a 'quasi-periodic' gait. In particular, work focuses on the challenges of walking on terrain that is not flat and includes intermittent obstacles. The approach exploits the observation that step-to-step 'snapshots' of the position and velocity states of the joints tend to lie on two-dimensional manifolds within a much higher dimensional state space for such systems. This dimensionality reduction enables the use of machine learning techniques for control policy evaluation and improvement to quantifiably estimate fall rates, energy use, and speed for a given combination of biped walker and stochastic terrain.If successful, this work will provide important tools to analyze and/or optimize a variety of real-world systems. From evaluation of the risk of falling for a stroke survivor who walks with an impaired gait, to the design of smart lower-limb prostheses for injured veterans, the modeling approach developed will provide a means of quantifying reliability for systems with such high dimensionality and complexity that traditional guarantees of stability cannot be made. Additionally, robotics research provides a natural gateway for both K-12 and university STEM education. This project includes several outreach elements to encourage this interest, from hosting on-campus field trips for local FIRST Robotics participants and predominantly Hispanic elementary school students to sponsorship of a new Robotics Club at UC Santa Barbara.