Z-REG: Domain-Robust Zero-Shot Point Cloud Registration via Equivariance-Aware Stabilization

Abstract

Point cloud registration is crucial in real-world applications such as robotics and automation. While deep learning-based methods have made substantial progress in this domain, their generalization remains limited in heterogeneous environments. This limitation primarily arises from their strong reliance on domain-specific tuning or retraining for each target scenario. To address this, we propose Z-REG, a domain-robust zero-shot point cloud registration framework that stabilizes equivariant local representations and enables reliable pose estimation in diverse scenarios. Specifically, a Confidence-Guided Equivariance Alignment (CGEA) is introduced that employs the confidence-gated canonicalization mechanism to softly modulate equivariant alignment, suppressing erroneous rotations induced by unreliable local reference frames. To mitigate geometric domain shift under fixed spherical discretization, a Patch-Adaptive Spherical Angular Modulation (PASAM) is proposed that adaptively learns a patch-specific elevation-azimuth modulation mask to re-calibrate equivariant angular features, thereby suppressing distortions from sparse or noise-dominated regions and yielding more consistent local embeddings across domains. Subsequently, Z-REG performs Cross-Scale Pose Verification (CSPV) to refine pose candidates by robust inlier-guided estimation, from which the optimal transformation is selected based on consensus support and residual error. Comprehensive experiments on challenging cross-domain scenarios demonstrate that Z-REG consistently achieves superior accuracy and robustness over the state-of-the-art methods. The code will be released soon. Note to Practitioners—The motivation of our work is to address the challenge of robust point cloud registration for robotic and automation systems in heterogeneous environments. In practical applications, point clouds are often collected by different sensors under varying conditions, resulting in substantial variation in point density, noise patterns, and geometric structures. Most existing learning-based methods rely on domain-specific fine-tuning or retraining when applied to new environments, which leads to substantial data annotation costs and long system deployment cycles, thereby limiting their practicality in real-world settings. To address this issue, this paper proposes Z-REG, a domain-robust zero-shot point cloud registration framework that improves cross-domain reliability without requiring domain-specific tuning or retraining. The proposed method stabilizes equivariant local representations by mitigating the influence of unreliable local reference frames and employing the patch-adaptive angular modulation. In addition, a cross-scale pose verification strategy is performed to improve the robustness of pose estimation. Extensive experiments have been conducted on several challenging cross-domain scenarios with significantly different characteristics to evaluate the performance of Z-REG. The results demonstrate that our Z-REG achieves more robust and accurate registration than existing state-of-the-art methods.