CableSplat: Optimized 3D Gaussian Splatting for 1D Deformable Pose Estimation

Abstract

A key challenge in the robotic manipulation of deformable objects is the lack of accurate and efficient systems for estimating their pose in real-time, especially in the presence of occlusion. In this thesis we propose CableSplat, a novel non-parametric method leveraging 3D Gaussian Splatting to estimate the pose of a linear deformable object given RGB images of the object from multiple viewpoints. To facilitate the evaluation of the performance of this method, we develop both simulated and real-world pipelines to collect calibrated and segmented recordings of cables undergoing various manipulations and transformations. We find that our method is consistently able to estimate cable pose to within an average error of ∼2.5mm across simulated tasks. Furthermore, performance on a scene reconstruction metric drops only slightly between simulated and real-world data, suggesting high-fidelity state estimation even in the real world. CableSplat is therefore a promising candidate for the extension of existing manipulation systems to deformables.