作者：Yunhai Han Mandy Xie Ye Zhao Harish Ravichandar
基于学习的方法的最新进展已经产生了令人印象深刻的右旋操纵能力。然而，我们还没有看到这些能力在实验室之外被广泛采用。这可能是由于实际限制，如巨大的计算负担、难以理解的策略架构、对参数初始化的敏感性，以及实施所需的可考虑的技术专业知识。在这项工作中，我们研究了库普曼算子理论在缓解这些限制方面的效用。库普曼算子是一种简单而强大的控制理论结构，有助于将复杂的非线性动力学表示为高维空间中的线性系统。由于复杂的非线性动力学是灵巧操作的基础，我们开发了一种模仿学习框架，利用库普曼算子同时学习机器人和物体状态的预期行为。我们证明了一个基于Koopmanoperator的框架非常有效
Recent advances in learning-based approaches have led to impressive dexterousmanipulation capabilities. Yet, we haven’t witnessed widespread adoption ofthese capabilities beyond the laboratory. This is likely due to practicallimitations, such as significant computational burden, inscrutable policyarchitectures, sensitivity to parameter initializations, and the considerabletechnical expertise required for implementation. In this work, we investigatethe utility of Koopman operator theory in alleviating these limitations.Koopman operators are simple yet powerful control-theoretic structures thathelp represent complex nonlinear dynamics as linear systems inhigher-dimensional spaces. Motivated by the fact that complex nonlineardynamics underlie dexterous manipulation, we develop an imitation learningframework that leverages Koopman operators to simultaneously learn the desiredbehavior of both robot and object states. We demonstrate that a Koopmanoperator-based framework is surprisingly effective for dexterous manipulationand offers a number of unique benefits. First, the learning process isanalytical, eliminating the sensitivity to parameter initializations andpainstaking hyperparameter optimization. Second, the learned reference dynamicscan be combined with a task-agnostic tracking controller such that task changesand variations can be handled with ease. Third, a Koopman operator-basedapproach can perform comparably to state-of-the-art imitation learningalgorithms in terms of task success rate and imitation error, while being anorder of magnitude more computationally efficient. In addition, we discuss anumber of avenues for future research made available by this work.
论文链接：http://arxiv.org/pdf/2303.13446v1
更多计算机论文：http://cspaper.cn/