Dynamic intelligence through online optimization

这项研究的目标是开发用于控制复杂机器人设备的算法和软件。这种方法是在一个新的数学框架内结合在线重新规划和离线学习，产生比目前可能的更丰富和更适应的行为。智能在线轨迹优化是控制缓慢和平滑的动力学(如化学过程)的首选方法。然而，机器人的动力学要快得多，而且不平滑，这对现有的方法提出了巨大的挑战。所提出的工作将克服这些挑战，通过利用随机最优控制的新的数学框架，即使潜在的动力学是非线性的，问题也被简化为线性方程。这些算法将使用一种新的物理引擎，依赖于并行计算来模拟机器人动力学，速度比实时快几个数量级。更广泛的影响拟议中的工作将改变机器人和动画角色的移动方式。目前，许多机器人系统都是开环控制的，或者被设计成能很好地执行一项特定的任务，但不具有通用性。动画主要是手绘的或基于运动捕捉数据的回放。这项工作将使机器人和动画人物都能表达更自然和多才多艺的动作。另一个重要的贡献是神经科学和生物力学，许多研究人员认为大脑对身体的控制是最佳的，但很难预测复杂任务中的最佳动作是什么。这里开发的算法将产生这样的预测，并使定量的模型数据比较能够促进我们对感觉运动大脑功能的理解。

The objective of this research is to develop algorithms and software for control of complex robotic devices. The approach is to combine online re-planning and offline learning within a novel mathematical framework, yielding richer and more adaptive behavior than what is currently possible.Intellectual MeritOnline trajectory optimization is the method of choice for controlling slow and smooth dynamics such as chemical processes. However robot dynamics are much faster and non-smooth, presenting formidable challenges to existing methods. The proposed work will overcome these challenges, by leveraging a new mathematical framework for stochastic optimal control where the problem is reduced to a linear equation even though the underlying dynamics are nonlinear. These algorithms will use a new physics engine that relies on parallel computing to simulate robot dynamics orders-of-magnitude faster than real-time. Broader ImpactThe proposed work will change how robots and animated characters move. Currently many robotic systems are controlled in open loop, or are designed to execute one specific task well but cannot be versatile. Animation is mostly hand-drawn or based on playback of motion capture data. This work will enable both robots and animated figures to express more natural and versatile movements. Another important contribution is to neuroscience and biomechanics, where many researchers believe that the brain controls the body optimally, yet it is difficult to predict what the optimal movements are in complex tasks. The algorithms developed here will generate such predictions, and enable quantitative model-data comparisons advancing our understanding of sensorimotor brain function.