Towards Physics Aware Embodied Control with Graph based Object-centric Learning

Understanding physics is crucial for intelligent robots and embodied agents to sense the world, move and manipulate objects, interact safely with their environment, and optimize motions and processes. In this paper, we propose GraphSlot , a slot-based object-centric learning framework that leverages graph neural networks to model object interactions. GraphSlot dynamically constructs graphs based on the spatial proximity of objects and external influence from gravity. Embedding information from neighbors is propagated between connected nodes. Through comprehensive experiments on simulation datasets, we demonstrate that GraphSlot achieves state-of-the-art performance with a remarkable enhancement of 8.5% in foreground Adjusted Rand Index (fg-ARI) comparing to the baseline SAVi-L model. As part of our evaluation, we design a real-world ball-catching game environment to test the physical intuition of our proposed model. GraphSlot shows promise for slot-based methods with physical understanding common sense.