Fusion 360 Gallery Dataset Overview

Updated 26 January 2026

Fusion 360 Gallery Dataset is a comprehensive collection of CAD models featuring detailed B-rep segmentation, human-authored design sequences, and multi-part assembly annotations.
It supports a variety of tasks including geometric deep learning, programmatic CAD analysis, and assembly reasoning via annotated operation sequences and graph-based representations.
The dataset serves as a benchmark for evaluating methods in B-rep segmentation, CAD program synthesis, and joint prediction, driving reproducible research in structured CAD model understanding.

The Fusion 360 Gallery Dataset is a comprehensive suite of datasets and environments derived from real CAD design workflows in Autodesk Fusion 360. It supports research across geometric deep learning, programmatic CAD analysis, and assembly reasoning, by providing annotated boundary-representation (B-rep) models, human-authored design sequences, and richly labeled mechanical assemblies. The datasets expose both the geometric and topological structure of CAD parts and assemblies, with detailed segmentations, semantic operation histories, program reconstructions, and assembly annotations, enabling a wide range of discriminative and generative learning tasks.

1. Dataset Structure and Components

The Fusion 360 Gallery Dataset comprises three major, task-specific components, each rigorously defined and released in conjunction with key research publications:

Segmentation Dataset: 35,858 B-rep CAD bodies (∼390,000 faces), each annotated at the face-level by the CAD modeling operation most responsible for its creation. Supported operations: Extrude (side/end), Cut (side/end), Fillet, Chamfer, Revolve (side/end) (Lambourne et al., 2021).
Programmatic Design Sequences: 8,625 full human-authored CAD programs expressed in a domain-specific sketch + extrude language. Each includes modeling sequence (operation-by-operation), geometric/constraint parameters, and final geometry as B-rep, mesh, and STEP (Willis et al., 2020).
Assembly and Joint Annotations: 8,251 multi-part assemblies, 154,468 bodies, and 32,148 parametric joint annotations (including joint-class, axis, pose, and hole/contact metadata). Accompanied by graph-based representations for learning over part connectivity (Willis et al., 2021).

Each dataset includes standard train/test/validation splits and is serialized in both native CAD (Fusion 360 .f3d/.smt), neutral geometry (.step, .obj), and richly indexed JSON metadata for direct integration into ML pipelines.

2. Segmentation: Structure, Annotation Schema, and Statistics

The segmentation component focuses on operation-aware labeling of B-rep bodies. Each face in a B-rep is assigned one of eight mutually exclusive labels, corresponding to operations in the native Fusion 360 workflow:

Label	Description	Proportion (%) of All Faces
ExtrudeSide	Lateral faces of additive extrusion	50.6
ExtrudeEnd	End faces of additive extrusion	15.7
CutSide	Lateral faces of subtractive extrusion	17.5
CutEnd	End faces of subtractive extrusion	2.35
Fillet	Created by a fillet operation	10.22
Chamfer	Created by a chamfer operation	2.94
RevolveSide	Lateral faces of a revolve	0.62
RevolveEnd	End faces of a partial revolve	0.08

Per-body modeling complexity is skewed toward simplicity: 49% of CAD bodies are created by a single operation, 31% have two or more, and only 1% involve more than ten operations. The dataset captures a wide tail in geometric complexity (max 421 faces per model, median 9, mean ≈10.9). File formats ensure direct mapping from native B-rep topologies (explicit face/edge/loop/coedge/vertex lists) to network input, mesh (OBJ, ~3,000 triangles/body), and point cloud (PLY, 2,048 uniformly sampled points) representations (Lambourne et al., 2021).

3. Programmatic CAD Sequences and the Fusion 360 Gym

The design sequence dataset supplies 8,625 parametric CAD programs in a purposely minimal DSL, limited to sketch and extrude operations for tractability of program induction tasks. Each program is a stepwise sequence with:

Sketch primitives: Lines, arcs, and circles dominate (∼95%), with sequences averaging ~9.7 curves and ∼6.3 geometric constraints per design.
Extrusions: Signed extrusion depths (mostly [1 mm, 20 mm]), with each operation specified by sketch source, profile, distance, and semantic Boolean (new body, join, cut, intersect).

Programs are stored as JSON, with corresponding ground-truth geometries in native (.smt), neutral (.step), and mesh (.obj) formats. The Fusion 360 Gym formulates CAD construction as a Markov decision process (MDP), representing state as the current/target geometry pair encoded as a B-rep face adjacency graph. The Gym supports both sketch extrusion and face extrusion action spaces and exposes deterministic transitions via the Fusion 360 API. Evaluation tasks include program reconstruction—recovering a minimal-length program matching a given target geometry with exact volumetric agreement (IoU, conciseness metrics) (Willis et al., 2020).

4. Assemblies, Joints, and Graph-Based Representations

The assembly dataset provides detailed annotations for mechanical assemblies, focused on the bottom-up prediction of parametric jointed assemblies. Key statistics:

8,251 assemblies, mean ≈18.7 bodies per assembly (distribution: 1–5 bodies ≈30%, 6–20 ≈45%, >20 ≈25%)
19,156 unique part-pair joint-sets, 32,148 total joints (joint taxonomy: rigid, revolute, slider, cylindrical, pin-slot, planar, ball joints—reflecting Fusion 360 API classes)
>80% of annotated joints involve circular/cylindrical geometry (holes/shafts), and 47.5% of joints directly mate holes to features on the other part
82% of part-pairs contain at least one hole

Contact surfaces and holes are algorithmically annotated (coincident B-rep faces within 0.1 mm, full metric diameter/depth/type for holes). Data is structured in JSON and as node-link graphs (NetworkX) capturing both intra-part (face/edge adjacency) and inter-part (joint axis, labels) connectivity, enabling graph neural network approaches to joint prediction, pose estimation, and contact reasoning (Willis et al., 2021).

5. Methods, Benchmarks, and Results

Segmentation, reconstruction, and assembly prediction have all been explicitly benchmarked:

BRepNet operates directly on B-rep coedge topology, incorporating permutation/incidence matrices to perform topological message passing. Face, edge, and coedge features encode surface curvature type, area, convexity, and direction flags. The architecture attains 92.5% accuracy and 77.1% mean IoU on operation segmentation, exceeding mesh- and point-cloud-based baselines, demonstrating the necessity of direct B-rep input for high-fidelity reasoning (Lambourne et al., 2021).
CAD program synthesis utilizes neural policies (Message-Passing Networks with graph, GCN/GAT/GIN variants) trained by imitation learning, optimized for program induction via search policies (random rollout, beam, best-first). Top approaches achieve ∼67% exact reconstruction (IoU=1) with concise programs on the holdout test set in face extrusion tasks (Willis et al., 2020).
Assembly/joint prediction leverages graph-based encodings with MLP + GATv2 message passing, cross-entropy and symmetric losses over dense joint connectivity bipartite graphs. Joint localization is refined by geometric axis and pose search minimizing overlap and surface contact cost functions (Willis et al., 2021).

A comparison of segmentation approach performance is summarized as follows:

Method	Accuracy (%)	Mean IoU (%)
BRepNet (winged edge)	92.52 ±0.15	77.10 ±0.54
ECC graph network	90.36 ±0.23	72.08 ±0.50
Rule-based CAD kernel	64.57	49.53
PointNet++	74.00 ±0.84	34.78 ±2.23
MeshCNN	62.99 ±0.37	20.59 ±0.41

6. Limitations, Biases, and Future Extensions

While the Fusion 360 Gallery datasets offer unparalleled scale, richness, and accessibility for CAD research, several limitations are acknowledged:

Extreme label imbalance (rare classes such as RevolveEnd at 0.08%) and dependence on high-quality watertight B-reps; non-manifold or noisy topologies degrade model performance (Lambourne et al., 2021).
The operation set for segmentation and program induction is currently restricted (primarily sketch, extrude, cut, revolve); planned extensions include loft, sweep, draft, and shell operations alongside operation order/group metadata, enabling enhanced parameter-history recovery and reverse engineering (Lambourne et al., 2021).
Joint data is limited to rigid, two-part parametric joints, with weak/partial supervision (positive-unlabeled challenge). Approximately 20% of public Gallery assemblies include explicit joints, and only rigid mechanical coupling is modeled (Willis et al., 2021).
Dataset content is skewed toward hobbyist and small-scale mechanical part designs.

Future work includes grafting the B-rep/topological architecture onto other cell complexes (polygonal meshes, superpixels, polyhedral graphs) and extending learning tasks to parameterization, generative synthesis, and robotics integration.

7. Accessibility, Usage, and Research Applications

The datasets are available via GitHub (see project pages in referenced papers) for non-commercial research use. Each component supports a variety of academic use cases:

Geometric deep learning on CAD B-reps, operation-aware segmentation, and face labeling
Programmatic reconstruction, inverse sketch/extrude program synthesis, and parameter recovery
Assembly reasoning, joint pose estimation, contact search, and assembly/disassembly planning
Cross-domain transfer between B-rep and mesh/point cloud representations

The Fusion 360 Gallery Dataset has established itself as the de facto benchmark for research on structured CAD model understanding and program induction, serving as the gold standard for reproducibility and rigorous benchmarking in this domain (Lambourne et al., 2021, Willis et al., 2020, Willis et al., 2021).