GIF: Agentic Generation of Interactive and Functional Object Compositions for Robot Learning
Agentic generation of functional object compositions (e.g. parts that must fit): disentangled mesh reconstruction, relative pose refinement under geometric and physical guidance, and a VLM verifier.智能体式生成具备功能关系的物体组合：先生成实例分离的网格，再在几何与物理引导下细化相对位姿，最后由 VLM 验证器挑选最符合描述的候选。
BibTeX
@misc{xu2026gif,
title = {{GIF: Agentic Generation of Interactive and Functional Object Compositions for Robot Learning}},
author = {Long Xu and Zhiqi Zhang and Mi Yan and Shengliang Deng and Chong Xia and Mingyu Dong and Jiayi Chen and Jiangran Lyu and Fei Gao and Zhizheng Zhang and He Wang},
year = {2026},
eprint = {2609.05927},
archivePrefix = {arXiv},
primaryClass = {cs.RO},
url = {https://arxiv.org/abs/2609.05927},
}