FAMOS: Feed-Forward 3D Articulation Modeling from Sparse Observations

TL;DR

FAMOS is a feed-forward deep learning model that infers 3D articulated object structure from sparse, unordered point clouds by jointly analyzing multiple views. It uses a Multi-state Articulation Transformer with state-wise and global attention to aggregate motion and geometry cues across observations, along with a novel objective to supervise motion ranges. The method improves upon existing approaches on benchmark datasets and uses a procedural data generator to enhance training data diversity.

Why It Matters

Accurate 3D articulation modeling from limited observations is crucial for robotics and augmented reality applications. FAMOS enables robust inference without iterative optimization, making it suitable for real-time systems. Its ability to handle variable input counts and synthetic training data improves scalability and generalization.

Key Findings

• FAMOS achieves state-of-the-art performance on PartNet-Mobility, ACD, and ArtiCraft-10K by jointly reasoning over multiple partial views.
• The Multi-state Articulation Transformer effectively captures articulation dynamics through alternating state-wise and global attention mechanisms.
• The observed articulation span objective improves motion range estimation by encouraging full utilization of input observations.