WorldGrow: Generating Infinite 3D World (github.com)

WorldGrow is a new generative pipeline from Shanghai Jiao Tong University, Huawei, and Huazhong University that synthesizes infinite, explicit 3D worlds by “growing” environments from a single seed block. The method uses a hierarchical, block-wise synthesis strategy with coarse-to-fine refinement: large-scale layout and connectivity are produced first, then local geometry and appearance are refined to yield detailed, textured, walkable meshes. Outputs are explicit reconstructed meshes with textured rendering (not implicit fields), and the authors demonstrate large examples such as a 19×39 indoor environment (~1,800 m²). Generated scenes are physically navigable and intended for navigation, planning, and evaluation tasks. For the AI/ML community this matters because WorldGrow offers a scalable alternative to hand-crafted or limited simulators—an open-ended generator that can produce coherent global structure while preserving fine local detail. That makes it useful for large-scale dataset creation, benchmarking, sim-to-real research, and training RL/robot navigation systems where diverse, walkable 3D worlds are required. The team has released the paper and initialized the repository (active development; interfaces may change) and plans to publish pretrained weights and full training/inference pipelines. The project promises a practical route to procedurally generate explicit 3D environments at scale.