Decomposing the Configuration of an Articulated Object via Graph Neural Network

Seunghyeon Lim, Kisung Shin, Nakul Gopalan, Jun Ki Lee, Byoung-Tak Zhang

Keywords: Articulated Objects Manipulation, Message Passing Neural Network, Language Embeddings

Abstract: Configuring links, joints, and their combinations is a critical skill for robots to manipulate complex articulated objects. In this paper, we employ a graph-based representation to model an articulated object, where the parts (i.e., links) are treated as nodes and the joints connecting them as edges. We train a Graph Neural Network (GNN) to learn the language embedding of the individual parts, determine the connectivity between different links, and predict the joint state values and types. Our model demonstrates superior performance across four key metrics, highlighting its applicability to robotic manipulation tasks. Then, we conducted initial experiments on the effectiveness of using joint information that our model can provide in learning the manipulation skills for articulated objects and presented its results. This emphasizes the potential of our model to offer significant advancements in reinforcement learning for robotic manipulation in the near future.

Submission Number: 25