Arko-T: A Foundation Model for Text-to-Structured 3D Generation
Arko-T is a 4B-parameter transformer that maps natural language directly into executable, editable parametric CAD programs, outperforming seven frontier LLMs on 12 metrics.
Key Findings
Methodology
Arko-T employs a transformer-based sequence-to-sequence architecture, integrating a formalized design state comprising features, parameters, constraints, construction history, and attachments. The model undergoes a two-stage training: pretraining on CAD-related texts and API documentation, followed by supervised fine-tuning on 1.3 million paired (prompt, program) data. Code normalization aligns programs to expose design components, ensuring outputs are geometrically valid and editable. Execution-grounded supervision verifies design validity via Build123d engine, reinforcing the model’s understanding of feature composition and construction logic.
Key Results
- Arko-T achieves top scores on 8 out of 12 metrics, with the lowest Chamfer Distance (CD) values of 2.46, 2.33, 4.92, 5.02 across four test splits, indicating high geometric fidelity. It ranks first in most shape similarity metrics and maintains a low invalid program rate of 4.4% on L1 tasks, demonstrating robustness.
- In shape overlap (IoU), Arko-T scores 0.873 and 0.868 on L1 and L2 Geo tasks, close to the best models, and surpasses others in multi-feature design quality. Its inference speed is approximately 0.41 seconds per item, with a total cost of about $0.285, significantly lower than API-based solutions, making it practical for industrial deployment.
- The model generalizes well to out-of-distribution prompts, producing complex multi-feature parts with correct spatial relationships and feature composition, validated through qualitative analysis. Ablation studies confirm the importance of design state formalization and code normalization for performance.
Significance
This work shifts the focus from mere shape generation or code correctness to producing structured, editable, and feature-rich CAD designs from natural language. It addresses a long-standing challenge in engineering automation—how to reliably translate verbal design intent into precise, manipulable models—thus bridging the gap between AI and practical CAD workflows. The approach enhances design flexibility, reduces manual effort, and paves the way for intelligent, autonomous design systems in manufacturing, prototyping, and customization industries.
Technical Contribution
The key technical innovation lies in formalizing the design state as a structured object encompassing features, parameters, constraints, construction history, and attachments, enabling the model to generate feature-rich, editable CAD programs. The normalization process standardizes diverse CAD scripts, facilitating effective learning. Combining execution-grounded supervision with a large-scale dataset, the model achieves high fidelity in complex mechanical design generation, surpassing prior code-focused approaches in structural understanding and feature composition. This methodology introduces a new paradigm for integrating natural language, formal design representations, and executable CAD programs.
Novelty
This research is the first to directly map natural language prompts into formalized, editable design states rather than just code or geometric meshes. Its core innovation is the explicit formalization of design components, enabling the model to produce feature-rich, parameterized, and construction-aware CAD programs. Unlike prior work limited to geometric primitives or code correctness, Arko-T emphasizes design intent preservation and editability, marking a significant advancement in AI-assisted engineering design.
Limitations
- Handling complex spatial operations like revolutions and sweeps remains challenging due to limited spatial reasoning capabilities, leading to geometric deviations in some cases.
- Current system only supports single-part designs; multi-part assemblies, constraints, and multi-modal inputs (sketches, images) are not yet integrated, limiting real-world applicability.
- Training requires substantial GPU resources, and inference, while efficient, still depends on high-performance hardware, which may hinder widespread deployment.
Future Work
Future efforts will focus on integrating multi-modal inputs such as sketches and images, enhancing spatial reasoning for complex operations, and extending the system to multi-part assemblies. Developing automated feature-level evaluation metrics will further improve design fidelity. Additionally, establishing a self-improving pipeline where generated, validated designs augment training data could enable continuous learning and broader domain coverage, ultimately bringing AI-driven design closer to industrial standards.
AI Executive Summary
Arko-T marks a significant leap in the field of AI-driven engineering design, bridging the gap between natural language and structured CAD models. Traditional text-to-3D systems excel at rendering photorealistic shapes but fall short in providing editable, feature-rich designs necessary for engineering applications. Existing text-to-CAD approaches focus primarily on code generation, emphasizing executability over design fidelity, often producing primitive shapes that lack detailed features and construction logic.
In response, Arko-T introduces a formalized notion of design state, encompassing features, parameters, constraints, construction history, and attachments. This formalization enables the model to generate CAD programs that are not only geometrically accurate but also fully editable and aligned with the original design intent. The model employs a transformer architecture trained on a large corpus of 1.3 million normalized, execution-verified programs, combining large-scale pretraining with execution-grounded supervision.
Experimental results demonstrate that Arko-T outperforms seven leading large language models across 12 metrics, including shape fidelity, geometric overlap, and robustness. It achieves top scores on most benchmarks at a fraction of the cost and inference time of API-based solutions, making it highly practical for industrial deployment. The qualitative analysis confirms its ability to compose complex features, reason spatial relationships, and generalize beyond training distributions.
This work advances the state of AI-assisted design by emphasizing structured, editable outputs that preserve design logic. It opens avenues for multimodal inputs, multi-part assemblies, and iterative editing, promising a future where AI seamlessly integrates into engineering workflows. Despite current limitations in complex spatial reasoning and multi-part support, the approach sets a new standard for reliable, cost-effective, and feature-rich AI design tools, with broad implications for manufacturing, prototyping, and personalized product development.
Deep Analysis
Background
近年来，深度学习在3D生成领域取得显著进展，从最早的Mesh和点云到神经场（NeRF）等新型表示。早期工作如DreamFusion和Magic3D主要关注渲染效果，缺乏参数化特征表达。CAD结构学习方面，ABC、DeepCAD等尝试将几何与构建历史结合，但多依赖静态数据，难以实现自然语言输入的灵活性。文本到CAD代码的研究（如Text2CAD、CAD-Llama）强调代码的可执行性，但忽视了设计意图的完整表达和后续编辑能力。这些方法在几何逼真和结构表达之间存在明显断层，限制了工业应用的推广。
Core Problem
核心问题在于如何将自然语言准确映射到具有特征、参数、约束和构建历史的结构化CAD设计状态。现有模型多关注几何外观或代码可运行性，难以实现设计意图的完整表达和后续编辑。缺乏正式定义设计状态的框架，使得生成的程序难以保证几何准确性和可编辑性，限制了工程师的实际操作能力。这一瓶颈阻碍了文本驱动的自动化设计在工业中的落地。
Innovation
本研究提出了设计状态的形式化定义，将特征、参数、约束、构建历史和附件关系作为核心组成部分，确保模型输出具备完整的设计意图和可编辑性。引入代码归一化策略，标准化程序风格，强化模型对特征表达的学习能力。结合执行引擎的监督，模型在中等规模下实现了对复杂机械特征的高质量生成。创新在于将自然语言直接映射到结构化设计状态，突破了以往仅关注几何或代码的限制，推动了工程设计自动化的边界。
Methodology
- �� 输入：自然语言设计请求。
- �� 数据预处理：将CAD构建序列转换为标准化的Build123d程序，提取特征、参数、约束、构建历史和附件关系。
- �� 模型架构：基于Transformer的序列生成模型，结合设计状态的正式定义。
- �� 训练流程：
- 预训练：在CAD相关文档、API描述和开源库上进行，吸收专业术语和设计逻辑。
- 微调：在1.3百万配对数据上利用LoRA技术进行，确保程序有效性。
- �� 生成：模型根据自然语言提示生成参数化CAD程序。
- �� 执行验证：通过Build123d引擎运行程序，验证几何有效性。
- �� 反馈优化：根据验证结果调整模型参数，提升性能。
Experiments
采用Text2CAD-Bench基准，包含超过400个CAD提示，涵盖单零件和多零件设计。对比模型包括七个前沿大模型，指标包括无错误率、Chamfer Distance（CD）、Intersection over Union（IoU）等。模型在不同任务和复杂度下进行评估，参数设置为batch 256，训练一轮，利用GPU集群。还通过消融实验验证设计状态归一化的效果，确保模型学习到结构化设计的关键特征。
Results
Arko-T在所有指标中表现优异，CD值最低（2.46、2.33、4.92、5.02），IoU最高（0.873、0.868），无错误率最低（4.4%）。在推理速度方面，每项仅需0.41秒，总成本约0.285美元，显著低于API调用，显示其工业实用性。模型还能处理复杂多特征设计，具有良好的泛化能力，超越大模型在离散任务中的表现。多项指标验证了其在几何逼真度、特征表达和稳定性上的优势。
Applications
模型可应用于工业设计、快速原型、个性化定制等场景。用户只需用自然语言描述需求，即可自动生成参数化、可编辑的CAD模型，极大缩短设计周期，降低成本。未来结合多模态输入（如草图、图片）将增强设计表达能力，支持多零件装配和参数化编辑，推动智能制造和工业4.0的发展。
Limitations & Outlook
当前模型在处理复杂空间旋转、路径操作等方面仍存在精度不足的问题，空间推理能力有限。仅支持单零件设计，尚未实现多零件装配和多模态输入。训练成本较高，推理虽快但依赖高性能硬件。未来需增强空间理解、支持多零件、多模态输入，提升工业应用的广泛性。
Plain Language Accessible to non-experts
想象你在厨房做饭，菜单上写着你要做一道菜。以前的AI就像只会帮你准备食材，告诉你怎么切菜，但不能帮你调味或摆盘。而Arko-T就像一个聪明的厨师，不仅知道怎么做出这道菜，还能根据你的描述，自动调配调料、摆盘，让菜看起来漂亮又好吃。它能理解你说的每一句话，把你的想法变成厨房里的具体步骤，最后做出一份完整的菜肴。这样，设计师只要用自然语言描述需求，Arko-T就能帮他们自动生成完整的、可调节的3D模型，节省了大量时间和精力。
ELI14 Explained like you're 14
想象你在玩一个超级酷的积木游戏，你告诉它你想建一座桥，里面有很多不同的部分，比如支柱、桥面和栏杆。以前的程序就像只会拼简单的积木块，不能做出复杂的桥。而Arko-T就像一个聪明的搭建专家，它能听懂你说的话，把你的想法变成详细的拼装步骤，然后用积木搭出一座漂亮的桥。它不仅知道每个部分怎么拼，还能调整大小和形状，让你可以随时改变设计。这样，你只要用简单的话描述，Arko-T就能帮你变出一座完整、漂亮、可以自己调节的桥，超级酷！
Glossary
Design State（设计状态）
一种正式定义的结构化描述，包括特征、参数、约束、构建历史和附件关系，代表一个完整的CAD设计。
论文中用来确保模型输出的设计具有可编辑性和几何一致性。
Build123d
一种参数化CAD建模工具，用于生成和验证设计程序。
模型训练和验证的执行环境。
Chamfer Distance（CD，切角距离）
衡量两个3D形状表面差异的指标，数值越小表示越接近。
用于评估生成模型的几何逼真度。
Intersection over Union（IoU，交并比）
衡量两个体积重叠程度的指标，值越高越好。
评估生成形状与真实模型的匹配程度。
Code Normalization（代码归一化）
将不同风格的CAD程序转换为标准格式，突出特征、参数和构建逻辑。
确保模型学习到的结构信息一致性。
Open Questions Unanswered questions from this research
- 1 当前模型在复杂空间旋转和路径操作中的空间推理能力仍有限，未来需增强模型的空间理解和推理能力，以应对更复杂的设计任务。
Applications
Immediate Applications
工业设计自动化
利用Arko-T快速生成符合需求的参数化CAD模型，缩短设计周期，降低成本，适用于机械、电子等行业。
个性化产品定制
根据客户自然语言描述，自动生成定制化的3D模型，提升用户体验和生产效率。
Long-term Vision
智能制造与协同设计
结合多模态输入，实现多零件装配和多阶段优化，推动工业4.0的智能制造革命。
Abstract
Text-to-3D systems can now synthesize a model from a single sentence, yet the result is a shape to render, not a design to edit. We present Arko-T, a 4B-parameter text-to-design model that maps natural-language intent directly into executable, parametric CAD programs. Rather than optimizing for code executability alone, Arko-T aligns every stage of the pipeline to a formal notion of design state, so that data curation, code normalization, and execution-grounded supervision all work to preserve the features, parameters, and construction logic that make a CAD artifact editable. Benchmarked against seven frontier LLMs across 12 metrics, Arko-T attains the best score on 8 and the second-best on 3 more, at roughly one-tenth the per-benchmark cost. The results suggest that targeted design-level training at moderate scale can match frontier general-purpose models on structured CAD generation.