Imagine trying to describe a complex mechanical engine part using a language that only contains two verbs: "draw" and "push." You could sketch a circle and push it into a cylinder, or sketch a rectangle and push it into a block. But how would you describe a curved exhaust pipe, a rounded safety edge, or a hollowed-out casing?
This is the fundamental bottleneck in modern generative Computer-Aided Design (CAD). While human engineers use a rich, highly specialized vocabulary of geometric operations to build 3D models, early machine learning frameworks have been restricted to a tiny fraction of this language. To bridge the gap between AI-generated shapes and professional engineering, a major research direction focuses on expanding the geometric operation vocabulary of automated CAD systems.
The Sketch-and-Extrude Bottleneck
In professional engineering, 3D models are not stored as uneditable point clouds or meshes. Instead, they are saved as parametric histories—ordered sequences of geometric commands that remain fully editable.
Until recently, automated systems capable of reverse-engineering or generating these sequences were severely limited. As noted by Karl D. D. Willis et al. in Fusion 360 Gallery: A Dataset and Environment for Programmatic CAD Reconstruction and Weijian Ma et al. in Draw Step by Step: Reconstructing CAD Construction Sequences from Point Clouds via Multimodal Diffusion, the vast majority of CAD reconstruction models limit their output to basic "sketch-and-extrude" workflows.
A similar limitation plagues Constructive Solid Geometry (CSG) models, which build shapes by combining basic volumetric primitives via Boolean operations (union, intersection, subtraction). As Kobus Barnard observed in early structural modeling in Learning models of object structure, and as reiterated by the authors of CSG-Stump: A Learning Friendly CSG-Like Representation for Interpretable Shape Parsing, relying on a restricted set of basic primitives like boxes and spheres prevents a model from capturing organic forms or highly detailed mechanical features. The resulting AI-generated models are blocky approximations rather than manufacturing-ready designs.
The Missing Vocabulary of Professional Engineering
To move beyond simple prismatic shapes, modern research focuses on incorporating three major classes of advanced geometric operations into machine learning vocabularies:
1. Edge Treatments (Fillets and Chamfers)
In real-world manufacturing, sharp edges are rare because they cause stress concentrations and pose safety hazards. Engineers use fillets to round off edges and chamfers to bevel them. The authors of CADParser: A Learning Approach of Sequence Modeling for B-Rep CAD and Faez Ahmed et al. in GenCAD: Image-Conditioned Computer-Aided Design Generation with Transformer-Based Contrastive Representation and Diffusion Priors emphasize that omitting these edge treatments prevents neural networks from recovering true design intent, leaving a massive gap between AI outputs and professional engineering standards.
2. Non-Linear Trajectories (Revolves, Sweeps, and Lofts)
Linear extrusion only moves a 2D sketch along a straight line. Real mechanical parts require complex trajectories:
- Revolves: Spinning a 2D profile around an axis to create spherical or cylindrical parts (e.g., wheels, wine glasses).
- Sweeps: Guiding a 2D profile along a curved 3D path (e.g., pipes, springs).
- Lofts: Smoothly transitioning a 3D shape between entirely different cross-sections (e.g., a car fender or an airplane wing).
As argued by Daxuan Ren et al. in ExtrudeNet: Unsupervised Inverse Sketch-and-Extrude for Shape Parsing, restricting models to linear paths fundamentally limits their expressive power. Anis Kacem demonstrates in MiCADangelo: Fine-Grained Reconstruction of Constrained CAD Models from 3D Scans that integrating revolves and sweeps is essential for capturing the design history of sophisticated industrial components.
3. Advanced Surfaces and Shells
Industrial designs frequently employ shelling (hollowing out a solid object to create a uniform wall thickness) and complex mathematical surfaces like Bézier splines or Non-Uniform Rational B-Splines (NURBS). Fenggen Yu et al. point out in D2CSG: Unsupervised Learning of Compact CSG Trees with Dual Complements and Dropouts that without these advanced surface types, generative models cannot accurately represent high-fidelity, organic CAD parts.
Why Expanding the Vocabulary is Hard
If adding these operations is so critical, why were they left out in the first place? The difficulty stems from topological dependency and the resulting combinatorial explosion of the search space.
In a basic sketch-and-extrude sequence, operations are largely independent or stack linearly. However, an operation like a fillet cannot exist in a vacuum; it must reference a specific topological edge created by a previous operation.
[Sketch 2D Profile] ──> [Extrude to 3D Solid] ──> [Identify Edge ID] ──> [Apply Fillet]
│
If Extrude parameters change,
Edge ID may break or disappear!
If a generative model modifies an early extrusion parameter, the downstream edge IDs might change, break, or disappear entirely, causing the geometric engine to crash. Furthermore, as explained by Xianghao Xu et al. in Inferring CAD Modeling Sequences Using Zone Graphs, expanding the command grammar to include revolves, sweeps, and fillets causes the number of valid permutations to explode exponentially. Standard autoregressive token prediction models struggle to maintain geometric validity under such vast, discrete search spaces.
Strategies for Handling Complex Operations
To overcome these constraints, recent research has shifted away from naive sequence generation toward architectures that embed geometric and topological awareness directly into the learning loop.
Space Partitioning and Graph Representations
Instead of guessing commands blindly, some frameworks discretize the infinite search space. Xianghao Xu et al. introduced Zone Graphs to partition 3D space into discrete solid zones. This converts a chaotic sequence-generation problem into a manageable graph-selection problem, allowing algorithms to reliably infer complex operations like revolves and fillets.
In-the-Loop Geometric Engines
To ensure that advanced commands actually produce valid 3D solids, researchers have begun embedding commercial geometric kernels directly into the machine learning training process. For instance, Xiaolong Yin et al. developed RLCAD: Reinforcement Learning Training Gym for Revolution Involved CAD Command Sequence Generation, which integrates the Parasolid geometric engine directly into a reinforcement learning loop. This provides immediate, real-time topological feedback to the agent, ensuring that complex revolution and edge operations resolve into valid boundary representations (B-Reps).
Spatial Reasoning and Multimodal Guidance
Other approaches leverage the massive spatial reasoning capabilities of modern multimodal architectures. Frameworks like CAD-GPT, introduced by Siyu Wang et al. in CAD-GPT: Synthesising CAD Construction Sequence with Spatial Reasoning-Enhanced Multimodal LLMs, and CADCrafter, developed by Cheng Chen et al. in CADCrafter: Generating Computer-Aided Design Models from Unconstrained Images, utilize explicit 3D spatial localization mechanisms and preference optimization. These additions ensure that advanced commands like shells, sweeps, and fillets align perfectly with visual inputs without breaking the underlying parametric structure.
By expanding the vocabulary of generative CAD from basic 2D extrusions to a full suite of professional engineering operations, these frameworks are turning AI from a simple sketching assistant into a competent automated co-designer.
Key Foundational Work
From Kinematics to Shape: An Approach to Innovative Design (1988) — contributed 1 concept elements to this direction.
Symmetry Constraint Inference in Assembly Planning: Automatic Assembly Configuration Specification (1990) — contributed 1 concept elements to this direction.