Generative 3D State Estimation for DLOs From Partial Observations

Abstract

Accurate 3D shape estimation of deformable linear objects (DLOs) from sensory data is a critical prerequisite for downstream manipulation tasks like shape control. However, real-world scenarios often contain unstructured environments and yield noisy, partial observations, posing significant challenges on robust DLO perception. Moreover, the near-infinite degrees of freedom inherent to DLOs give rise to a vast range of possible deformations, further complicating reliable state estimation. To tackle these challenges, we introduce a novel generative approach for 3D DLO state estimation from point cloud observations. Our method utilizes a denoising diffusion model conditioned on partial point cloud inputs to reconstruct the complete DLO shape. Furthermore, the framework supports the integration of flexible guidance objectives during inference, enhancing the stability and reliability of shape estimation. Extensive experiments demonstrate that the proposed method achieves robust performance in both simulation and real-world scenarios.