REArtGS++: Generalizable Articulation Reconstruction with Temporal Geometry Constraint via Planar Gaussian Splatting

TL;DR

REArtGS++ tackles generalizable articulated object reconstruction from two arbitrary RGB states without depth supervision. It introduces part-aware planar Gaussian representations, a decoupled screw motion model, and a temporal geometric constraint via a Taylor-based regularization to enforce consistent normals and depth across time. Key contributions include (1) unsupervised part segmentation and joint parameter estimation without priors, (2) a time-continuous geometry constraint that improves optimization across motion, and (3) robust performance on synthetic and real-world datasets, including screw and multi-part objects. The method demonstrates superior part-level surface reconstruction and joint parameter estimation, offering practical benefits for robotics and embodied intelligence in dynamic environments. Practical impact lies in enabling accurate, generalizable articulated modeling from minimal supervision, with competitive efficiency relative to prior state-of-the-art approaches.

Abstract

Articulated objects are pervasive in daily environments, such as drawers and refrigerators. Towards their part-level surface reconstruction and joint parameter estimation, REArtGS introduces a category-agnostic approach using multi-view RGB images at two different states. However, we observe that REArtGS still struggles with screw-joint or multi-part objects and lacks geometric constraints for unseen states. In this paper, we propose REArtGS++, a novel method towards generalizable articulated object reconstruction with temporal geometry constraint and planar Gaussian splatting. We first model a decoupled screw motion for each joint without type prior, and jointly optimize part-aware Gaussians with joint parameters through part motion blending. To introduce time-continuous geometric constraint for articulated modeling, we encourage Gaussians to be planar and propose a temporally consistent regularization between planar normal and depth through Taylor first-order expansion. Extensive experiments on both synthetic and real-world articulated objects demonstrate our superiority in generalizable part-level surface reconstruction and joint parameter estimation, compared to existing approaches. Project Site: https://sites.google.com/view/reartgs2/home.

Figures (7)

• Figure 1: Given multi-view RGB images at arbitrary two different states of an unseen articulated objects, our method achieves high-quality part-level dynamic reconstruction and joint parameter estimation, without any external models.
• Figure 2: Framework of REArtGS++. Our method jointly optimizes part segmentation, joint parameters and planar Gaussians using multi-view RGB images from arbitrary two states, and achieves high-quality part-level mesh reconstruction of any states and accurate joint parameter estimation for an unseen articulated object. "Diff." denotes the difference approximation.
• Figure 3: The qualitative results of dynamic surface reconstruction at start state and end state on ArtGS-Multi dataset. We show both part segmentation and surface meshes for best comparison. The red arrows represent joints.
• Figure 4: The qualitative results of dynamic surface reconstruction at start state and end state on PARIS dataset. We show both articulated modeling and surface meshes for best comparison.
• Figure 5: The qualitative results of dynamic surface reconstruction at start state and end state on PARIS dataset. We show both articulated modeling and surface meshes for best comparison.