FAMOSFeed-Forward 3D Articulation Modeling
from Sparse Observations
1 Stanford University2 ETH Zürich
Summary
FAMOS is a feed-forward method that predicts movable-part segmentation and joint parameters from a sparse set of monocular observations. By jointly reasoning over the whole input set, it grounds articulation prediction in observed motion rather than shape priors alone. The model can handle a variable number of inputs (including a single view) and is trained at scale by extending the training data with assets from our procedural data generator.
- +28.2 pts Movable-part segmentation F1 gain on ACD over the strongest feed-forward baseline
- +27.8 pts Motion estimation MAO F1 gain on ACD over the strongest feed-forward baseline
- 0.1 s Inference time nearly 3,000× faster than per-object optimization
Abstract
Modeling articulated objects from sparse monocular views is challenging because each observation reveals only partial geometry and motion evidence. Most feed-forward methods infer articulation from a single observation and therefore rely heavily on learned category-level shape priors. We present FAMOS, a feed-forward model that predicts movable-part segmentation and joint parameters from a sparse, unordered set of partial point clouds. Our model jointly reasons over multiple observations and naturally supports a variable number of inputs, including a single view. To aggregate articulation cues across observations, we introduce a Multi-state Articulation Transformer with alternating state-wise and global attention. We further propose an observed articulation span objective that supervises the motion range each part exhibits across the input observations, encouraging the model to leverage the full observation set. To overcome the limited scale and diversity of existing datasets, we introduce a procedural data generator that synthesizes self-annotated assets during training. Experiments on PartNet-Mobility, ACD, and ArtiCraft-10K demonstrate consistent improvements over both feed-forward and optimization-based baselines.
How It Works
Multi-State Articulation Transformer
-
01
Point encoding
Each observation is lifted to a partial point cloud, then tokenized with a frozen PartField backbone and geometric embeddings. No state-index embedding is added, preventing the model from learning state-index-specific identities and allows it to flexibly process a variable number of inputs during inference.
-
02
Alternating attention
Each transformer layer alternates between two attention stages. State-wise attention lets the point tokens of each input view aggregate context over the geometry of their own partial point cloud. Global attention then refines all point and query tokens jointly, allowing global information exchange across observations.
-
03
Prediction heads
Per-query heads decode joint type, axis and origin; a per-point head decodes the movable-part segmentation. Because one query segments a part across every observation, the segmentation is consistent between states.
Observed Articulation Span
We introduce an auxiliary objective that predicts, for each movable part, the range of joint motion observed across the input states. Since this quantity is defined over the full observation set, it encourages consistent part localization and aggregation of motion evidence across states, rather than relying on a single observation. The span is also invariant to the unknown canonical rest pose, making it a well-defined supervision signal for sparse, unordered inputs.
Procedural Training Data
To increase the scale and diversity of our training data, we design a procedural asset generation pipeline that assembles articulated objects from basic shape primitives (boxes, cylinders, hollow cylinders, half-tori, spheres, holed panels). The procedural generation introduces minimal computational overhead and can run on-the-fly during training in the dataloader.
– object families
Comparison to Baselines
Two-door cabinet
PartNet-MobilityVanity desk
ACDSideboard
ACDFAMOS visualization: detached point clusters and one unsupported axis omitted.
Stove cabinet
ArtiCraft-10KDrawer cabinet
ArtiCraft-10KStatic basePredicted movable partsRed axes: revolute jointsYellow axes: prismatic jointsReArt receives complete point clouds.† Particulate is re-trained on our data.For clarity, predictions on one input view are shown.
Real-World Objects
Despite being trained on synthetic data only, FAMOS generalizes to casual real-world captures.
Drawer unit





Laptop




Metal cabinet




Nightstand




Kitchen cabinet




Cane cabinet




Wall cabinet




Hardcover book




Static basePredicted movable partsRed axes: revolute jointsYellow axes: prismatic joints
Quantitative Results
FAMOS consistently outperforms the baselines in movable-part segmentation and articulation estimation, for both multi-state and single-state settings.
Particulate† is re-trained on our data.
Movable-part segmentation F1 at IoU ≥ 0.5. † Re-trained on our data. : gain over the strongest baseline.
Runtime per object
Inference Time
Four input states on one NVIDIA RTX 5090 · logarithmic scale.
Why Multiple States Matter
Articulation estimation from a single state is intrinsically ambiguous: several kinematic models explain the same visible geometry, causing the prediction to fall back on learned category-level shape priors. FAMOS instead grounds its prediction in the motion cues that additional observations reveal.
Citation
TBD — BibTeX citation coming soon.