Causal Scene Deconfounding for 6-DoF Grasp Detection

Abstract

Robotic grasp detection aims to learn reliable grasp poses from 3D point clouds. Existing methods primarily train neural networks to estimate grasp poses but often overlook the dataset biases stemming from imbalanced scene distributions. Specifically, due to the underrepresentation of densely cluttered and severely occluded scenes, models are prone to learning spurious correlations between 3D point cloud features and grasp poses. To address this issue, we introduce causal inference into grasp detection and propose a Scene Causal Intervention for Grasping (SCIG) network. We first build a tailored causal graph to formalize the confounding effect introduced by scene context. Building on this analysis, we construct a scene confounder dictionary that captures diverse scene prototypes through clustering of pre-trained scene features. Leveraging this dictionary, SCIG then implements a back-door intervention to mitigate such confounding, effectively reducing scene-induced biases. Importantly, our method is designed as a plug-and-play, model-agnostic plugin that can be seamlessly integrated into existing grasping frameworks. Extensive experiments on the GraspNet-1Billion dataset demonstrate that SCIG substantially improves generalization, particularly in minority and complex scenes. Furthermore, cross-dataset evaluations on the GraspClutter6D dataset validate the effectiveness of our causal approach for zero-shot 3D grasp detection.