SHORT ANSWER
Policy asks “what should I do?” World model asks “what happens if this occurs?”
A VLA policy conditions on observations, language and often robot state to generate actions. A world model represents or predicts transitions: how the environment and robot state may evolve over time, potentially under candidate actions.
FUNCTIONAL DIFFERENCE
The two roles optimize different outputs.
Primary output
Robot action, action token, action vector or action chunk.
Primary question
Given this observation, instruction and state, what action should the policy produce?
Typical training signal
Robot demonstrations and action-conditioned trajectory data.
Primary output
Predicted next state, future latent representation, video, dynamics or transition estimate.
Primary question
Given the current state and possibly an action, how is the world likely to evolve?
Typical training signal
Temporal transitions, future observations, dynamics or predictive consistency.
WHY COMBINE THEM
A policy can benefit from a model of consequences.
Long-horizon robot behavior benefits from predicting whether an action will make progress, cause a collision, lose contact or move the scene into a harder state. A world model can provide that predictive structure while a policy remains responsible for selecting or generating the action.
CURRENT CONVERGENCE
The boundary is already becoming architectural rather than model-level.
NVIDIA GR00T N1.5 is a concrete example: it keeps flow-matching policy learning while adding Future Latent Representation Alignment, a predictive objective that aligns current representations with future target embeddings. The result is still a robot policy, but it learns an additional world-model-like representation of future state.
WORLD ACTION MODELS
Some research increasingly treats prediction and action as one coupled problem.
The field is moving toward architectures where learned dynamics, future prediction and policy generation share representations. That does not erase the conceptual distinction: prediction tells a system what may happen, while an action policy determines what the robot will attempt.
OPENRGD VIEW
OpenRGD keeps the semantic roles separable even when one neural model implements both.
OpenRGD research treats world-model state, physics context, cognition proposals and execution as distinct semantic responsibilities. A future implementation could use one foundation model for several of those jobs, but the graph can still record which evidence is predictive, which output is an intended action and which component owns physical execution.