A new AI system called FAMOS can figure out how a chair, drawer, or robot arm moves after seeing just a few partial 3D scans of it.
Researchers built FAMOS, a feed-forward model that predicts which parts of an object are movable and estimates their joint parameters - the hinges, slides, and rotations that define how something like a laptop hinge or cabinet door actually works. Unlike most prior systems, which infer motion from a single snapshot and lean heavily on learned shape assumptions, FAMOS can take in any number of sparse, unordered point-cloud observations, including just one, and reason across all of them at once. It does this with what the team calls a Multi-state Articulation Transformer, which alternates attention within each observation and across the full set. The researchers also trained it with a new objective that tracks the full range of motion a part shows across observations, plus a procedural data generator that creates its own labeled training assets to work around the shortage of existing articulated-object datasets.
Robots and AR systems increasingly need to predict not just what an object looks like, but how its parts move, and doing that from a handful of quick scans - rather than a full 3D scan or CAD model - is the practical bottleneck. FAMOS reportedly beat both other feed-forward models and slower optimization-based methods on three benchmark datasets, PartNet-Mobility, ACD, and ArtiCraft-10K.
It's a narrow, technical improvement, not a flashy demo - but stitching together partial views instead of guessing from a single glance is exactly the kind of unglamorous fix that tends to compound in robotics pipelines.