CAREER: Machine Learning Control of Underactuated Mechanical Systems

非线性欠驱动系统是机器人学中一类重要而普遍的问题，对于分析和数字控制设计范例来说，这类问题已被证明是最难解决的。机器学习方法用于欠驱动控制，它利用近似使数字最优控制技术易于掌握，将有广泛的应用，从行走的机器人到空中飞行器和流体系统的控制。在这里，我们继续对应用于线性时不变(LTI)系统的算法进行仔细的分析。这将有助于关于不同学习算法的收敛速度的基本结果，以及最大限度地提高收敛速度的机器人机构和输入-输出“特征”的设计。将理论结果与一个强非线性控制问题--两连杆两足机器人在崎岖地形上行走的实验相结合。为这个机器人获得一个接近最优的反馈策略将产生一个令人惊讶和令人信服的结果(因为一个简单的机器人将穿越比任何人形机器人所展示的更复杂的地形)，但也清楚和揭示(因为机器人的简单暴露了行走中的基本问题，仅此而已)。理论和实验都使我们能够对机器学习控制和现代控制中较为成熟的控制方法进行仔细的比较。

The nonlinear underactuated systems represent an important and generalclass of problems in robotics which have proven mostly intractable foranalytical and numerical control design paradigms. Machine learningapproaches to underactuated control, which employ approximations tomake numerical optimal control techniques tractable, will have broadapplications from walking robots to the control of aerial vehicles andfluid systems. Here we pursue a careful analysis of the algorithmsapplied to linear time-invariant (LTI) systems. This will contributefundamental results on the convergence rate of different learningalgorithms, and the design of robot mechanisms and input-output"features" which maximize the rate of convergence. Theoreticalresults are coupled with experiments on a strongly nonlinear controlproblem - a two-link bipedal robot walking over rough terrain.Acquiring a near optimal feedback policy for this robot would producea result that is surprising and compelling (because a simple robotwill be traversing more complicated terrain than has been demonstratedby any humanoid), but also clear and revealing (because the simplicityof the robot exposes the fundamental problems in walking and nothingmore). Both theory and experiment allow us to perform a carefulcomparison between machine learning control and more mature controlapproaches from modern control.