Task-Oriented Functional Part Extraction Via Hierarchical Semantic 3D Gaussian Splatting

Abstract

Task-oriented grasping requires precise functional part extraction for post-grasping manipulation, yet existing 3D reconstruction methods suffer from semantic ambiguity due to single-label assignments. To address this challenge, we propose an LLM-driven hierarchical semantic 3D Gaussian Splatting framework. First, by integrating the Segment Anything Model and CLIP, we construct a multi-scale semantic field covering the background, complete objects, and subparts to effectively mitigate point-level conflicts. Within this field, an LLM-guided reasoning module translates natural language task descriptions into specific functional part queries without manual annotations. Furthermore, a two-stage semantic correlation mechanism progressively isolates the target object from the robotic end-effector to extract fine-grained functional parts. Experimental results demonstrate that our method achieves a localization accuracy of 81.2% and a mean intersection over union of 89.7%, outperforming the baseline approaches. This framework provides reliable 3D semantic guidance for complex manipulation tasks in unstructured environments.