ArrangeNet: A Voting-Based Network for Predicting Goal Positions in Robotic Object Rearrangement

Robotic object rearrangement is an important task that involves pick-and-place operations to move objects according to the desired goal layouts. In this paper, we propose a novel method to solve the object rearrangement task: ArrangeNet. Given RGB images of the initial and target scenes and queried objects, ArrangeNet implicitly solves object segmentation and matching problems through the attention mechanism in transformer. The goal positions of objects are predicted by leveraging the contextual information among them. The final positions of objects are obtained through a voting mechanism to eliminate the uncertainties due to partial views or occlusions. Experimental results demonstrate the effectiveness of the proposed ArrangeNet in robotic object rearrangement tasks, especially in challenging scenes with partial occlusions.