Robotic Skill Learning Using Riemannian Task-Parameterized Kernelized Movement Primitives

Abstract

Robotic skill learning has gained prominence as robots evolve from executing controlled, repetitive tasks to operating in complex and unpredictable environments. Traditional hand-crafted motion design methods fall short in such settings, prompting the need for data-driven approaches such as imitation learning and reinforcement learning. However, existing methods often encounter challenges in generalizing to unforeseen conditions. Recent studies have shown that accurately modeling manifold-valued data such as orientations and poses instead of confining analysis in Euclidean space can substantially improve the generalization of learned skills. Moreover, combining this intrinsic geometric modeling with task parametrization further enhances robustness in unfamiliar environments. This thesis addresses these limitations by integrating Riemannian geometry into Kernelized Movement Primitives with task parametrization. In particular, a novel method Riemannian Task-Parameterized Kernelized Movement Primitives (Riemannian TP-KMPs) is introduced and Single Tangent Space Task-Parameterized Kernelized Movement Primitives (STS TP-KMPs) is used for comparison. This approach leverages the intrinsic geometric structure of manifold-valued data to robustly generalize complex movements, thereby enhancing the adaptability and robustness of learned skills in various environments. The methods are evaluated through both simulation and real-robot experiments focused on a specific task, with their performance being analytically compared. The results indicate that explicitly modeling the intrinsic geometry of robotic data in conjunction with task parametrization offers a compelling alternative for advancing robotic skill learning and generalization.