CAREER: Investigating the Ultimate Mechanisms of Embodied Cognition

迄今为止，在实现在非结构化环境中持续执行简单但自适应行为的机器方面取得的成功相对较少（与工厂等结构化环境相比）。制造这种机器的流行方法是复制在动物身上观察到的生理和神经系统，并将它们构建成机器人。然而，这就提出了一个问题，即在无限的现有生物结构中，应该复制什么。该奖项下的研究正在寻求一种替代方法：而不是复制现有的生物系统，进化动力学在虚拟空间中复制和连接。由此产生的进化算法优化了虚拟机器人的神经结构，控制行为和身体计划。重要的是，这些研究中的进化是特定于任务和行为的，这项研究旨在为机器人和生物学做出重要贡献。对于机器人专家来说，这项工作将使计算机能够自动设计机器人的身体计划和神经控制器，这些机器人比手动设计的机器人更具适应性和鲁棒性。然后，自动设计的虚拟机器人可以作为物理设备构建并部署到现实世界的环境中，包括那些对人类有危险的机器人。对于生物学家来说，我们的研究将深入了解为什么以及如何在自然界中进化出特定的结构。例如，如果最初进化用于运动的腿式机器人然后被选择用于移动和抓取物体，则计算进化可以将机器人的前腿重新用于手臂和抓取器;或者，它可以将操纵附件添加到现有的身体计划上。最后，实验被安置在在线工具中，允许研究生，本科生和K-12学生在自己的机器上被动地运行进化模拟，以及积极参与这个过程：他们可以设计新的虚拟环境，机器人必须在其中进化。这种积极参与的目的是激励学生了解物理学，生物学，工程和计算过程的基础进化。

To date, relatively little success has been achieved in realizing machines that continually perform simple yet adaptive behaviors in unstructured environments (compared to a structured environment such as a factory). The prevailing approach to create such machines is to copy physiological and neurological systems observed in animals, and build them into robots. This raises the issue however of what from among the infinitude of existing biological structures should be copied. Research under this award is pursuing an alternative approach: rather than copy existing biological systems, evolutionary dynamics are copied and connected in a virtual space. The resulting evolutionary algorithm optimize virtual robots' neurological structures that control behavior and their body plans. Importantly, evolution in these studies is task and behavior specific.The research is intended to make important contributions to robotics and biology. For roboticists, this work will enable computers to automatically design the body plans and neural controllers for robots that are more adaptive and robust than robots designed manually. Automatically-designed virtual robots can then be built as physical devices and deployed into real-world environments, to include those that are dangerous to humans. For biologists, our studies will provide insight into why and how particular structures evolved in nature. For example, if legged robots originally evolved for locomotion are then selected to locomote and grasp objects, computational evolution may re-purpose the robot's front legs into arms and grippers; or, it may add manipulatory appendages onto the existing body plan. Either outcome would be of great interest to evolutionary biologists.Finally, experiments are being housed in online tools that will allow graduate, undergraduate and K-12 students to run evolutionary simulations passively on their own machines, as well as actively participate in the process: they may design novel virtual environments in which the robots must evolve. This active participation is intended to motivate students to understand the physics, biology, engineering and computational processes underlying evolution.