LAM: Language Articulated Object Modelers

LAM tackles text-driven articulated 3D object generation by representing both geometry and motion as executable code. Instead of relying on an existing mesh, image, video, or prebuilt 3D asset, the system decomposes a text prompt into an object-part hierarchy and then builds an articulation-ready model procedurally.

The method coordinates specialized LLM/VLM modules: a Link Designer plans the parts and relationships, Geometry and Articulation Coders produce executable structure and joint code, deterministic debuggers catch code issues, and visual checkers iteratively critique rendered geometry and simulated motion. This code-centered representation makes the generated object interpretable and reusable for downstream interactive 3D, robotics, embodied AI, gaming, and VR/AR workflows.

Resources: Project page / Paper / Code / Data

Recommended citation: @inproceedings{gao2026lam, title = {LAM: Language Articulated Object Modelers}, author = {Gao, Yipeng and Ge, Yunhao and Cai, Peilin and Seita, Daniel and Itti, Laurent}, booktitle = {Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)}, year = {2026} }
Download Paper