Sep 13, 2026 · 24 min
World models turn 3D space into AI infrastructure
World Models, Robotics, and the Future of 3D AI
The conversation frames spatially grounded world models as a bridge between generative media, robotics, games, and physical intelligence.
- 1Atlas combines environment generation, image-based reconstruction, and simulation in one foundational world-model approach.
- 2Persistent 3D context and camera control could make generated environments more navigable, consistent, and useful than ordinary video outputs.
- 3Real-to-sim-to-real workflows could let robots train and evaluate against simulations built from footage of their actual operating environments.
Don't miss
Johnson describes how real video could be converted into environment-specific simulations for evaluating and training robots.
The brief
Justin Johnson of World Labs presents Atlas as foundational infrastructure: a model that can generate environments, reconstruct real spaces from images, and simulate how objects and robots behave.
The central argument is that world models should serve many fields rather than one product category, spanning creative tools, games, VR, robotics, and physical intelligence.
Atlas grounds reference images in 3D space, giving users persistent context and camera control that can prevent generated environments from drifting like ordinary video outputs.
The robotics use case is especially concrete: footage from a real environment could become a simulation for testing robots, enabling a reel-to-sim-to-reel training loop.
The episode’s larger tension is control versus spectacle: world models become useful when humans and agents can direct them, not when they behave like opaque generative slot machines.
Featuring
Justin Johnson
World Labs
Sora
Flight Simulator
Grand Theft AutoListen to the full episode and explore every guest, topic, and moment on PodLume.