01Known worlds
Games and metaverse taught the interface.
People know avatars, spaces, inventories, maps, shared places.
World model concept map
World model concept map: AI video to spatial worlds.
Use the world model concept map to connect AI video, spatial computing, digital twins, physical AI, and generated worlds.
Virtual worldsAI videoSpatial computingPhysical AI
Concept mapCore sentence
Virtual worlds were human-built. Now AI learns to generate, control, simulate them.
Minecraft and Roblox explain the mental model, the metaverse persistence and social space, Vision Pro spatial computing. EMO, Veo, Wan, Kling, and Ray explain controllable video; Cosmos and digital twins simulation.
Scene explainer
Three steps.
01
02Generated media
AI video made the world visible.
Synthetic scenes became easy to watch and share, but behaved like clips.
03World models
The next layer: enterable and controllable.
Scenes remember space, respond to action, and support agents or simulation.
Concept flow
How the concept map connects familiar ideas.
01
Past interface
Human-built virtual worlds
02
Current surface
AI-generated video and humans
03
Spatial interface
Spatial computing and immersive access
04
Industrial layer
Simulation, digital twins, and physical AI
05
Core capability
World models
Past interface
Human-built virtual worlds
People already understood avatars, sandbox worlds, social rooms, user-built spaces.
Minecraft-style identity
Blocky avatars
Simple characters make presence easy: a person enters a world.
MinecraftRoblox avatarVoxel worlds
Who inhabits the generated world, and can identity persist?Buildable spaces
Sandbox worlds
Minecraft and Roblox trained users to expect modifiable, shareable worlds.
Minecraft blocksRoblox experiencesUGC worlds
Generated spaces become more valuable when editable, not disposable.Persistent social space
Metaverse
Framed virtual worlds as social, persistent, identity-driven, despite manual tooling.
Meta Horizon WorldsVR roomsSocial worlds
Worlds generated on demand, not only built by hand.Current surface
AI-generated video and humans
The visible surface; deeper issues are control, consistency, and memory.
Audio-driven identity
Expressive humans
The same identity must move, emote, sing, and stay coherent.
EMODigital humansTalking avatars
If generated people cannot persist, generated worlds feel unstable.Prompt-to-motion
Video models
Turn text, images, audio, and references into moving scenes.
Veo 3.1Wan2.7-VideoKlingRaySora
Can those scenes be controlled, extended, and interacted with?From avatar to actor
Digital characters
Generated characters will need continuity across avatars and portraits.
MetaHumanRoblox avatarEMO portraitRunway Characters
Characters are the social layer of generated worlds.Spatial interface
Spatial computing and immersive access
They are how generated worlds may be seen and operated.
Computer as environment
Spatial computing
Reframes computing as placed into space, not a flat screen.
Apple Vision ProSpatial video3D interfaces
Need interfaces where generated space can be inspected, edited, inhabited.Scene as data
3D reconstruction
Make real or imagined spaces computable.
NeRF3D Gaussian SplattingLingBot-MapHY-World 2.0
Generated worlds need spatial structure, not only pixels.From screen to place
Immersive worlds
Users feel located inside a generated or captured environment.
Meta QuestVision ProImmersive video
More legible when users can enter and manipulate the output.Industrial layer
Simulation, digital twins, and physical AI
Not entertainment; simulation for robots, vehicles, factories, and cities.
Real world mirror
Digital twins
Model real systems so teams test changes before touching reality.
NVIDIA OmniverseFactory twinsCity simulation
Simulations become cheaper to create and easier to vary.AI for embodied systems
Physical AI
Models of how environments respond to motion, contact, and decisions.
CosmosHY-Embodied-0.5LingBot-VALingBot-VLA
World models become training infrastructure, not just media.World-scale spatial memory
Geospatial models
Connect AI to real-world places, maps, and location-aware behavior.
Niantic spatial AIMapsAR location layers
They turn the real world into a modelable environment.Core capability
World models
Modeling how worlds change under time, viewpoint, and action.
World responds to action
Interactive generation
Preserve coherent state when the user moves, edits, or acts.
Genie 3MarbleHappyOysterHY-World 2.0
The difference between watching a clip and entering a system.Base models for simulation
World foundation models
Reusable infrastructure for generating, predicting, and testing world states.
CosmosGWM-1World APIHY-World 2.0LingBot-VALingBot-VLA
Where creative, spatial, and physical generation share a vocabulary.World as training ground
Agent environments
Agents need environments to observe, act, fail, and learn.
Game worldsRobot simulatorsInteractive scenes
A substrate for training and evaluating future AI agents.Bridge table
What each concept contributes.
| Entry concept | Known for | Connects to | Meaning in world models |
|---|---|---|---|
| Blocky avatars / Minecraft | Simple identity inside a buildable world | Avatars, sandbox worlds, UGC | Generated worlds need persistent users, objects, and editable structure. |
| Metaverse | Persistent social virtual spaces | VR, Horizon Worlds, social identity | Automate world creation instead of relying on manual building. |
| Vision Pro | Spatial computing and immersive interface | AR, spatial video, 3D interaction | Generated worlds need spatial interfaces for viewing, editing, operation. |
| AI video | Generated motion, characters, and scenes | EMO, Veo 3.1, Wan2.7-Video, Kling, Ray | The video layer must become controllable, continuous, and stateful. |
| Digital twins | Simulation of real systems | Omniverse, robotics, LingBot-VA, LingBot-VLA, city and factory models | Useful when they predict and test real-world behavior. |
| World model | Predicting and generating world state | Genie 3, Marble, Cosmos, GWM-1 | Not a place or device; the model making worlds behave. |