GAMESLASHAI 3D models look fine until pieces collide: BrickWright checks each connection
By GameSlash
BrickWright gives AI coding agents rules to catch floating and overlapping 3D pieces. A research method game makers can learn from for modular construction.
AI can generate a good-looking 3D build that breaks the moment you inspect its parts: pieces overlap, float or never connect. The same kind of mistake can frustrate creators making modular game worlds, where every wall, platform or tile needs to fit.
In a study posted October 22, 2026, BrickWright gives AI coding agents a parts catalogue and automatic assembly checks instead of trusting the final render alone. Its official project site includes real interactive 3D examples and source code.
What does BrickWright actually create?
The system takes a text or image description and writes Python code to assemble actual LEGO-type brick parts. The paper shows a lighthouse attacked by a sea serpent, vehicles and articulated objects, saved in LDraw format. It is designed for brick assemblies, not for exporting ready-made levels to Unity, Unreal or Godot.
The AI can search a catalogue, snap compatible connectors together, and inspect the result from different angles. The important advance is the checking process it uses while building—not the claim that it can draw a model from words.
It checks the mistakes an attractive 3D image can hide
With BrickLayer, an agent checks whether parts intersect, float without support or split into disconnected groups. BrickWright then loops through writing code, validating the object, rendering previews and fixing failed placements until the required checks pass.
These checks are geometric. The paper explicitly does not claim to test all real-world stability or whether a human could assemble each model in a practical order.
512 prompts: what improved?
Across 512 benchmark requests, the researchers reported 100% physical feasibility by their defined checks for BrickWright-guided agents. Agent baselines without BrickWright ranged from about 9% to 69%, depending on the model. This is a paper benchmark rather than a GameSlash test, and passing the rules does not imply beautiful or commercially ready art.
What can someone making a game with AI learn?
For voxel games, modular building systems or AI-assisted map tools, the transferable idea is to let AI suggest placements and have game-side code reject collisions, floating platforms and invalid connections before the scene is accepted.
This is an engineering lesson rather than a plug-in you can install into a game. Using the same pattern requires your own game-part library, snapping rules and export pipeline. The current BrickWright system specifically targets LEGO brick construction.
Sources: research paper · interactive demos and code · original figures by Pun et al. under CC BY 4.0. The LEGO Group is not affiliated with the project.