World Labs
What changed
Atlas broadens spatial output
World Labs reports one model spanning text, image, video, point clouds, and Gaussian splats.
WORLD MODELS / 2026
From language to physical intelligence
A living research atlas tracing how world models connect perception, prediction, planning, and action through primary sources.
15Tracked models
TECHNOLOGY FIELD
SelectedLayers of prediction, planning, and action
This sequence is a reading path, not a mandatory architecture; systems may skip layers or create feedback loops.
- 01
Models statistical structure in language and produces or interprets text.
- 02
Connects visual observations with language representations and instructions.
- 03
Represents how an environment may change, including the consequences of actions.
Inspect evidence for this layer - 04
Searches possible actions or trajectories against a goal and predicted outcomes.
- 05
Observes, chooses, acts, and updates its behavior across a feedback loop.
- 06
Applies learned perception, prediction, and control in embodied systems.
- 07
Maintains an operational digital representation linked to a physical system.
- LLM informs VLM.
- VLM informs the world model.
- The world model predicts for the planner.
- The planner plans with the agent.
- The agent acts through physical AI.
- The digital twin feeds back to the world model.
WML · World Model Laboratory
Explore the connection between perception, prediction, planning, and action studied in WFM through experiments in WML. Compare predictions with simulated outcomes in interactive physics scenes, then explore a different action from the same decision point.
Educational simulation: WML uses simplified predictors; it does not validate the performance of learned foundation world models or transfer to the real world.
Open the experiment in WMLWFM within aserdargun.com
WFM is a research atlas in the physical AI layer of the AI Learning System. Read the sources here, explore prediction and action in WML, then follow asset records, evidence, and human oversight in ITL.
These links describe a learning path. The applications run independently; they do not imply shared telemetry or automatic control. ITL examples are synthetic teaching records.
What changed
All signalsNVIDIA
What changed
Cosmos 3 opens its development stack
The technical report links multimodal world generation with released development artifacts.
Waymo
What changed
Waymo adds a driving world model
The announced simulator targets controllable and counterfactual autonomous-driving scenes.
Runway
What changed
Runway frames a general world model
GWM-1 extends generative video research toward controllable environments and robotics.
NVIDIA Research
What changed
Cosmos 1.0 launches a world foundation model platform
NVIDIA presents the first Cosmos release as a development platform for physical AI.
Decart & Etched
What changed
Oasis opens real-time interactive worlds
Decart and Etched publish a transformer-based world that responds to user input in real time.
Meta AI
What changed
V-JEPA introduces latent video pretraining
Meta's first V-JEPA work establishes self-supervised joint-embedding prediction for video as a world-model pretraining objective.
arXiv
What changed
GAIA-1 frames driving as a generative world
Wayve's GAIA-1 paper defines video-based generative world modeling for autonomous-driving research.
Decart
What changed
Oasis 3 targets physical-AI training
Decart reports real-time, multi-view simulation environments controlled by physical signals and available through an API.