Hierarchical control for space robotic applications

为什么你可以走到一个想要的地方，伸手抓住一个物体，如此优雅和轻松，而我们的机器人却如此笨拙、缓慢和容易出错？这是一个让工程师关注的问题，他们探索自然世界是如何运作的，以使我们创造的人工制品更好（被称为生物灵感或仿生学）。我特别关心的是让机器人去探索人类无法探索的遥远的太空和行星。为了做到这一点，我们需要让我们的机器人更有能力、更聪明。毫无疑问，如果我们能让机器人变得更聪明，我们就能让它们做更多的事情，从而使它们成为更有效的替代品。我们可以在更大的背景下更快、更便宜、更有效地了解我们的环境。特别是，天体生物学对我们的机器人探险者在火星上寻找生命提出了越来越高的要求。我的方法是研究生物有机体学会了做什么以及如何做。一个重要的教训是，机器人和其他应用中普遍采用的传统反馈控制本身不足以赋予鲁棒性和适应性行为。生物进化已经进化出了额外的策略来提高生物体的性能，以补偿神经“湿件”中信息传输的缓慢。这些策略是基于前馈预测的，要么是遗传的（就像刚出生的小马驹一出生就会走路），要么是通过经验学习的（就像人类婴儿需要一年或更长时间才能学会走路）。通过将这种前馈方法与机器人的传统反馈方法相结合，我希望实现一种更自然的行为，这种行为快速，可以适应可变的环境。这将使我们能够创造出更高效的探索其他星球的机器人。但这只是第一步——前馈控制只是智能控制器的一个方面，但它是一个重要的方面。

How is it that you can walk to a desired location and reach out and grasp an object with such grace and ease yet our robots are so clunky, slow and error-prone? That is a question that concerns engineers who explore how the natural world works in order to make our created artifacts better (known as bio-inspiration or biomimetics). My particular concern is to make robots to explore space and planets which are too far away for humans to explore. In order to do this we need to make our robots more capable and smarter. There is no doubt that if we can make robots smarter we can get them to do more things and therefore make them more effective surrogates. We could learn more about our environment in its larger context faster, cheaper and more efficiently. In particular, the science of astrobiology is making ever greater demands on our robot explorers to search for life on Mars. My approach is to examine what biological organisms have learned to do and how. One important lesson is that traditional feedback control commonly adopted in robotics and other applications is by itself insufficient to impart robust and adaptable behaviour. Biological evolution has evolved additional strategies to enhance an organism's performance to compensate for the slow transmission of information in their neural "wetware". These strategies are based on feedforward prediction either inherited genetically (like a newborn foal walking immediately after being born) or learned through experience (like a human baby which takes a year or more to learn to walk). By incorporating such feedforward approaches to the traditional feedback approaches in robots, I hope to achieve a much more natural behaviour which is rapid and can adjust to a variable environment. This will enable us to create more efficient robot explorers to other planets. But this will only be the first step - feedforward control is merely one aspect of intelligent controllers but it is an important one.