3D Assembly Completion

Automatic assembly is a promising research topic in 3D computer vision and robotics. Existing works focus on generating (e.g., IKEA furniture) from scratch with set of parts, namely part assembly. In practice, there are higher demands for the robot to take over finish an incomplete half-assembled off-the-shelf toolkit, especially human-robot multi-agent collaborations. Compared assembly, it more complicated nature remains unexplored yet. The must understand structure, infer what parts missing, single out correct toolkit finally, assemble them appropriate poses Geometrically similar can interfere, this problem will be exacerbated missing parts. To tackle issue, we propose novel task called completion. Given aims find its predict 6-DoF make complete. end, FiT, framework Finishing Transformer. We employ encoder model into memories. Candidate interact memories memory-query paradigm final candidate classification pose prediction. Bipartite matching symmetric transformation consistency embedded refine For reasonable evaluation further reference, design two standard toolkits different difficulty, containing compositions conduct extensive comparisons several baseline methods ablation studies, demonstrating effectiveness proposed method.