BrickWright: Agentic Creation of Physically Feasible Brick Builds
Abstract Overview
BrickWright is a framework for generating physically feasible LEGO brick builds from text or image prompts by guiding pre-trained coding agents to write BrickLayer code in a plan-execute-validate loop. BrickLayer provides part discovery over the LDraw catalog, connector-aware assembly functions, and validation for collisions, floating parts, and disconnected components. The paper defines physical feasibility primarily as using real catalog parts, avoiding collisions and floaters, and forming as few connected components as possible. Experiments on 512 prompts compare BrickWright against vanilla coding agents and specialized brick-generation models using both feasibility and prompt-adherence metrics.
Novelty
The main novelty is pairing general-purpose pre-trained coding agents with a brick-specific toolkit that combines an ergonomic assembly API and an executable physical-feasibility validator, instead of training a task-specific generator. The framework also extends this agentic workflow to articulated, image-conditioned, and constrained brick generation.
Results
On the BrickNet-512 benchmark, BrickWright-guided agents achieved 100% validity for all tested agent backbones, compared with baseline validity rates ranging from 9% for GPT-6 Luna to 69% for GPT-6 Astra. The full system improved prompt adherence for lightweight agents and maintained it for stronger agents, while outperforming specialized text-to-brick models on overall feasibility. Ablations showed that the validator is critical for high validity, and the assembly API halves inference cost relative to working directly in raw LDraw.
Key Points
- BrickLayer introduces searchable part discovery, connector-typed assembly operations, and validation checks for collisions, floating parts, and connectivity within a unified Python representation.
- BrickWright uses a lightweight plan-execute-validate loop to let pre-trained agents iteratively write, test, render, and revise brick-build code until feasibility checks pass.
- Empirically, the framework reached 100% valid builds on 512 prompts across agent sizes, and the ablations indicate that both the assembly API and validator materially contribute to performance.
References
- arXiv: https://arxiv.org/abs/2610.31651v1
- Fugu-MT: https://fugumt.com/fugumt/paper_check/2610.31651v1
- Hugging Face Papers: https://huggingface.co/papers/2610.31651
- Project: https://avalovelace1.github.io/brickwright/