StableText2Lego
收藏资源简介:
StableText2Lego 数据集是一个包含超过 47,000 个不同的 LEGO 结构的大型数据集,这些结构覆盖了 ShapeNetCore 数据集中的 28,000 个独特的 3D 对象,共分为 21 个常见的对象类别。每个结构都配有一组文本描述和一个稳定性分数,用于指示其物理稳定性和可构建性。该数据集的构建过程涉及将 ShapeNetCore 中的 3D 形状转换为 LEGO 结构,然后进行结构稳定性和文本描述的生成。该数据集可用于训练生成物理稳定和可构建的 LEGO 设计的模型,并可用于解决实际物体设计中存在的构建难度和结构稳定性问题。
The StableText2Lego dataset is a large-scale collection comprising over 47,000 distinct LEGO constructions, which cover 28,000 unique 3D objects from the ShapeNetCore dataset and are categorized into 21 common object categories. Each construction is paired with a textual description and a stability score that indicates its physical stability and buildability. The construction of this dataset involves converting 3D shapes from ShapeNetCore into LEGO structures, followed by the generation of structural stability metrics and textual descriptions. This dataset can be used to train models that generate physically stable and buildable LEGO designs, as well as to address challenges related to construction difficulty and structural stability in real-world object design.
数据集概述:Generating Physically Stable and Buildable LEGO® Designs from Text
基本信息
- 标题: Generating Physically Stable and Buildable LEGO® Designs from Text
- 作者: Ava Pun*, Kangle Deng*, Ruixuan Liu*, Deva Ramanan, Changliu Liu, Jun-Yan Zhu (* denotes equal contribution)
- 机构: Carnegie Mellon University
- 论文链接: arXiv
- 代码: 提供
- Demo: 提供
数据集描述
- 数据集名称: StableText2Lego
- 规模: 超过47,000个LEGO结构,涵盖28,000个独特的3D对象
- 数据生成流程:
- 从ShapeNetCore网格生成LEGO设计,通过体素化到20×20×20网格并应用legolization确定砖块布局。
- 通过随机化砖块布局增强每个形状的多个结构变体,同时保留整体形状。
- 对每个变体进行稳定性分析,过滤掉物理不稳定的设计。
- 通过24个不同视角渲染LEGO设计,使用GPT-4o生成详细的几何描述。
- 特点: 每个形状附带详细描述,确保多样性和稳定性。
方法概述
- 模型名称: LegoGPT
- 方法:
- 将LEGO设计标记化为文本标记序列,按从下到上的光栅扫描顺序排列。
- 创建指令数据集,将砖块序列与描述配对,微调LLaMA-3.2-Instruct-1B。
- 在推理时,LegoGPT根据文本提示逐步生成LEGO设计,每次预测一个砖块。
- 对每个生成的砖块进行有效性检查,确保格式正确、存在于砖块库中且不与现有砖块碰撞。
- 完成设计后验证其物理稳定性,若不稳定则回滚到稳定状态并继续生成。
应用示例
- 文本到LEGO结构生成:
- "A streamlined vessel with a long, narrow hull"
- "A classical guitar"
- "A basic sofa"
- "A bookshelf with horizontal tiers"
- "A high-backed chair"
- "A backless bench with armrest"
- 机器人自动组装: 展示生成结构的自动化组装过程(8倍速)。
- 纹理和彩色LEGO模型生成:
- "Rustic stone bench with moss growth [...]"
- "Hot rod with flame paintwork [...]"
- "Electric guitar in metallic purple [...]"
- "Sunburst Les Paul with amber finish [...]"
引用
bibtex @article{pun2025legogpt, title = {Generating Physically Stable and Buildable LEGO Designs from Text}, author = {Pun, Ava and Deng, Kangle and Liu, Ruixuan and Ramanan, Deva and Liu, Changliu and Zhu, Jun-Yan}, journal = {arXiv preprint arXiv:2505.05469}, year = {2025} }
致谢
- 感谢Minchen Li, Ken Goldberg, Nupur Kumari, Ruihan Gao, Yihao Shi的讨论和帮助。
- 感谢Jiaoyang Li, Philip Huang, Shobhit Aggarwal开发的双手机器人系统。
- 支持机构: Packard Foundation, Cisco Research Grant, Amazon Faculty Award, Manufacturing Futures Institute, Carnegie Mellon University, Richard King Mellon Foundation, Microsoft Research PhD Fellowship。




