ArtiSG: Functional 3D Scene Graphs for Robotic Manipulation

This paper addresses a critical limitation in robotic scene understanding: the lack of functional information about articulated objects. Existing methods struggle with visual ambiguity and often miss fine-grained functional elements. ArtiSG offers a novel solution by incorporating human demonstrations to build functional 3D scene graphs, enabling robots to perform language-directed manipulation tasks. The use of a portable setup for data collection and the integration of kinematic priors are key strengths.

• •Proposes ArtiSG, a framework for constructing functional 3D scene graphs.
• •Utilizes human demonstrations to overcome limitations of existing methods.
• •Employs a portable setup for robust data collection.
• •Integrates kinematic priors and interaction data.
• •Demonstrates superior performance in real-world experiments.