[论文] BrickWright: Agentic Creation of Physically Feasible Brick Builds

论文概要

研究领域: CV
作者: Ava Pun, Kevin Dai, Sally Chen, Maxwell Jones, Ruixuan Liu, Kangle Deng, Changliu Liu, Maneesh Agrawala, Deva Ramanan, Jun-Yan Zhu
发布时间: 2026-10-22
arXiv: 2610.31651

中文摘要

前沿编码智能体在 3D 建模方面越来越熟练，但在构建需要在物理世界中组装的物体时仍面临挑战，经常创造出存在悬浮、碰撞或断开部件的无效组装。我们认为，为智能体提供组装构建的工具和物理可行性验证能力，就足以让它们生成既物理有效又符合输入提示的组装体。聚焦于乐高积木领域，我们提出 BrickWright——一个使用预训练智能体从文本或图像提示生成物理可行积木构建的框架。BrickWright 通过 BrickLayer 引导智能体编写代码来创建构建，BrickLayer 是一个 Python 工具包，包含两个主要组件：一个符合人体工程学的积木组装 API（相比原始 LDraw 将推理成本降低一半），以及一个提供物理可行性反馈的验证器。基线智能体的物理有效率仅从 9%（GPT-6 Luna）到 69%（GPT-6 Astra）不等，而 BrickWright 引导的智能体在 512 个提示样本上达到了 100% 有效率，同时在轻量级智能体上改善了提示遵循度，在更强智能体上保持了提示遵循度。

原文摘要

Frontier coding agents are becoming increasingly adept at 3D modeling. Yet, they still face challenges when building objects meant to be assembled in the physical world, often creating invalid assemblies with floating, colliding, or disconnected parts. We present BrickWright, a framework for generating physically feasible LEGO brick builds from text or image prompts using pre-trained agents. BrickWright guides agents to create builds by writing code with BrickLayer, a Python toolkit with an ergonomic brick assembly API and a validator that provides feedback on physical feasibility. Whereas baseline agents achieve physical validity rates ranging from only 9% to 69%, BrickWright-guided agents achieve 100% validity on 512 prompts.

自动采集于 2026-10-24

#论文 #arXiv #CV #小凯