Plausible
Preserve task-relevant state over time.
A decision-centered survey of embodied world models
World Models for
Embodied Intelligence
From Plausible to Controllable to Actionable
A world model matters not only because it can imagine a convincing future, but because it preserves the right state, responds correctly to intervention, and improves an embodied agent’s decisions.
Capability ladder
Controllable
Predict how interventions change that state.
Actionable
Turn prediction into measurable downstream gains.
ManipulationNavigationLocomotionAutonomous driving
01 / Motivation
The central question is no longer
“Can the model generate a realistic future?”
but
“What does that prediction enable an agent to do?”
Represent
What state must remain identifiable and measurable?
Intervene
Does changing an action produce the corresponding effect?
Decide
Does prediction improve planning, learning, verification, or recovery?
02 / Capability framework
Three levels.
Three evidence requirements.
The levels describe progressively stronger claims, not mutually exclusive architectures. A method may demonstrate a higher-level use while leaving a lower-level property untested.
01
P
State consistency
Plausible
A plausible model preserves task-relevant temporal, geometric, or physical structure across a rollout—beyond surface-level visual realism.
IdentityGeometryPhysicsMemoryDrift
Evidence: held-out state errors, consistency under occlusion, and performance as a function of horizon.
02
C
Intervention fidelity
Controllable
A controllable model predicts how a commanded intervention changes the future while preserving unrelated aspects of the scene.
Action identityTimingMagnitudeComposition
Evidence: paired or branched interventions from shared initial states, including masked and shuffled-action controls.
03
A
Decision utility
Actionable
An actionable model changes a downstream decision or update and produces a measured gain under matched data, controller, compute, and latency budgets.
PlanningLearningEvaluationRecovery
Evidence: realized task utility against a matched baseline that removes world-model input.
03 / Grounding × improvement
Where prediction
enters the loop.
The matrix connects what grounds a prediction—geometry, physics, or action—to the system component it informs. Select any cell to inspect the interface.
GroundingImprovement loop
01Data
02Reward
03Policy
04Model-self
GGeometry
PPhysics
AAction
04 / Technical landscape
A field in motion.
The survey traces technical progressions within each level, from compact predictive state to intervention-aware modeling and closed-loop use.
Plausible progression
- 01Latent & object-centric state
- 02Metric 3D/4D scene state
- 03Physics-grounded dynamics
- 04Persistent memory
Controllable progression
- 01Action-conditioned prediction
- 02Latent & language actions
- 03Measurable intervention variables
- 04Joint world–action models
Actionable progression
- 01Planning, search & selection
- 02Policy learning from imagination
- 03Learned simulators
- 04Verification & recovery
Across embodiments
The ladder stays fixed.
The decisive variable changes.
Manipulation
Contact, object state, force
Autonomous driving
Multi-agent futures, safety
Navigation
Partial observability, geometry
Locomotion
Fast dynamics, stability
05 / Paper library
Explore the
research landscape.
162 papers
Indexed directly from citations and classifications in the survey body.
Showing 0 of 200 matching papers
All body classificationsNo papers match these filters.
Try a broader keyword or clear one of the classification filters.
06 / Open problems
What still blocks
the next capability.
The most consequential gaps appear at the boundaries between levels—especially when models face longer horizons, new policies, new embodiments, and real-time constraints.
Persistent state & drift
Retain identity, geometry, and unresolved uncertainty across occlusion and long rollouts.
Causal action effects
Separate intervention response from correlations inherited from demonstrations.
Decision utility under budget
Preserve value, feasibility, and safety within real compute and control deadlines.
Joint grounding
Make geometry, physics, and action modules agree when their predictions conflict.
Uncertainty & stable loops
Revalidate confidence after policy updates and prevent shared errors from circulating.
Access & reproducibility
Report action interfaces, budgets, latency, evaluation seeds, and failure cases.
The evaluation shift
The right state.
The right response.
A better decision.
Progress toward actionability should be measured by whether grounded predictions improve closed-loop behavior under explicit data, compute, and latency budgets.
07 / Citation
Build on the framework.
If this survey supports your research, please cite it using the entry below.
@article{yao2026worldmodels,
title = {World Models for Embodied Intelligence:
From Plausible to Controllable to Actionable},
author = {Yao, Nanjie and Wang, Hao and Cheng, Chong and
Chen, Zhikang and Li, Wenzhe and others},
year = {2026}
}