Back to AI game news
AI game news

WorldStream Builds Persistent 3D Worlds with Macro–Micro AI Planning

By GameSlash

Visit website

WorldStream's October 20 research separates AI planning from scene construction and stores edits beyond loaded geometry, offering design lessons for persistent game worlds.

Researchers from USTC and JD's Joy Future Academy presented WorldStream on October 20, 2026: a 3D-world prototype that expands its terrain as someone explores, accepts natural-language edits and reconstructs previously visited locations with their changes intact. It is research for interactive-world generation, not a released commercial game engine.

Official WorldStream paper Figure 1 showing generated city, rural, forest, desert and coastal regions plus language-grounded interactions.
Figure 1 © Xiaoxiao Ma et al. / WorldStream — paper's overview of five terrain families and runtime actions; author figure, not an independent game test (CC BY 4.0).

Separate AI planning from deterministic scene construction

The system's core decision is architectural. A Macro planner organizes settlements, ecology, roads and water across regions. A Micro planner turns those commitments into tile-level semantic plans. Local code then compiles terrain, buildings, vegetation and entities into 3D geometry. Unlike systems that regenerate a video frame for each action, this design keeps an explicit map of what exists and how it changed.

WorldStream Figure 3 comparing Macro-level regional planning, Micro tile specification and deterministic 3D scene realization.
Figure 3 © Xiaoxiao Ma et al. / WorldStream — authors' three-stage planning-to-geometry diagram; a technical illustration rather than gameplay capture (CC BY 4.0).

What game creators can learn

For AI game builders, WorldStream is a concrete reference for two difficult production problems: coherent infinite maps and persistent player edits. It plans upcoming tiles ahead of motion, compiles only the near-future geometry, and stores tile plans, stable entity identities, random seeds and modifications even after distant meshes are released from memory.

The design also grounds natural-language requests in existing world state. Players can direct nearby actors, alter terrain or describe content for regions that have not yet been generated. This is not evidence that any existing Unreal, Unity or Godot project can import WorldStream as a plug-in; it is a research approach developers can study.

A research reference, not a drop-in game engine

The paper evaluates a Three.js-based WebGL prototype. Its authors note that generation depends on a bounded set of content types and curated geographic rules; long-running worlds still need policies for accumulated records and actors outside the active area. Developers can study the architecture, but the reported results do not establish production readiness for their own game.

Read the WorldStream paper and figures for the method and evaluation limits.

View source