← Zinan Yang · lab

Two ways a machine can know the world

Reflex&Imagination

A vision-language-action model and a world model both sit inside modern robots — but they learn opposite things. One learns a mapping. The other learns a simulator.

VLA · vision-language-action
“What should I do, right now?”
A policy. Pixels and words in, motor commands out. No prediction of the future — it acts, then looks again.
World model
“What will happen next?”
A simulator. Given a state and an action, it predicts the next state — so futures can be rehearsed before anything moves.

The fundamental difference

Where the loop closes

Every intelligent controller runs a loop: sense, decide, act, sense again. The deepest difference between these two families is where that loop closes.

A VLA closes its loop through reality. It maps what it sees and what it is told directly to an action, executes it on the real world, and only then observes the consequence. All of its knowledge is compressed into that single mapping — a trained reflex.

A world model closes its loop inside itself. It learns the dynamics of the environment — next state = f(state, action) — so it can feed its own predictions back in as input and roll the world forward in its head, many steps, many alternatives, without touching anything real.

VLA — a trained reflexone forward pass per action
pixels camera frame words “pick the red block” policy π action = π(pixels, words) VLM backbone + action head action Δ pose, gripper REALITY the actual world the loop closes through reality — act first, see what happened after
World model — a learned simulatormany imagined steps per real action
pixels observation encoder compress z latent state dynamics z′ = f(z, a) imagination — feed predictions back in, no reality required decoder imagined frames reward how good is z′? planner / policy pick the best imagined future

Notice what the world model diagram does not output directly: an action. On its own, a world model doesn’t act at all — it only predicts. It needs a planner or a policy bolted on to choose between the futures it imagines. And notice what the VLA never produces: a prediction. It cannot tell you what the world will look like after its action — it simply doesn’t model that.

Simulation · run it yourself

The blocked corridor

Both agents were “trained” on the same maze: demonstrations that go straight down the middle corridor. Run them, then block the corridor and run again — this is where the two philosophies split.

VLA · reflexacts from the learned mapping
idle — press run
World model · imaginationrehearses futures, then acts
idle — press run

With the corridor open, the reflex is superb — fewer computations, straight to the goal. Block it, and the VLA keeps emitting the action its training data prescribed for “corridor-like” observations: it walks into the new wall, because nothing in its mapping represents consequences. The world model never memorised a route at all — it rehearses candidate futures against its internal simulator (the faint violet trails), scores them, and simply imagines its way around the change. (Toy rendering, of course: real VLAs generalise better than a lookup table, and real world models imagine in latent space, not on a grid — but the asymmetry is exactly this.)

The price of imagination

Dreams drift

If world models can rehearse the future, why not imagine hundreds of steps ahead and plan everything? Because a learned simulator is never exact. Each imagined step adds a small error, and errors compound — the dream slowly parts company with reality. Drag the horizon and watch twelve imagined futures of the same bouncing ball fan away from the true trajectory (white).

imagination horizon 60 steps · spread

This is why practical systems imagine in short bursts: plan a handful of steps, execute the first one, observe reality, and re-plan from the fresh observation. Reality keeps correcting the dream before it drifts too far.

If you come from control engineering

This whole page may feel familiar. A VLA is a learned direct controller — a reflex law mapping measurement to actuation, like a tuned PID or a cam table: fast, fixed, no plant knowledge. A world model is the plant model inside an MPC: the imagination horizon is the prediction horizon, and “imagine a few steps, execute one, re-observe, re-plan” is precisely receding-horizon control.

And training a policy inside a world model (Dreamer-style) is commissioning your controller against the digital twin before you ever download to the real machine.

Side by side

The anatomy of the difference

VLA
World model
Core question
What should I do?
What will happen?
What is learnt
A policy — a direct mapping from observation + instruction to action
The dynamics of the environment — a predictive simulator of state transitions
Mapping
a = π(o, language)
s′ = f(s, a)
Loop closes
Through reality — act, then observe
Inside the model — imagine, then act
At inference
One forward pass per action — fast, reactive, System 1
Many imagined rollouts per decision — deliberate, System 2
Fed by
Demonstrations (teleoperation, imitation) + internet-scale vision-language pre-training
Any experience at all — video, interaction logs — no labels or demonstrations needed
Great at
Following open-ended language, transferring semantic knowledge (“the red block”), low-latency control
Counterfactuals, planning routes never demonstrated, safe rehearsal, learning without a real robot
Fails by
Confident nonsense off-distribution — repeats the reflex when the world changed
Compounding prediction error — the dream drifts; hallucinated physics rewards impossible plans
Flag-bearers
RT-2OpenVLAπ0HelixGemini Robotics
Dreamer V3MuZeroGenie 3V-JEPA 2Cosmos

Where the field is heading

Rivals on paper, partners in practice

The honest answer to “which is right?” is that they solve different halves of the same problem — and the strongest recent systems are hybrids. Three couplings keep appearing:

Train the reflex in the dream

Dream training

Dreamer-style: learn a world model from experience, then train a fast policy entirely inside its imagination — millions of rehearsals, zero wear on real hardware. The deployed artefact behaves like a reflex; it was raised in a simulator.

Manufacture experience

Synthetic data

Video world models (Genie, Cosmos) act as data factories: generate plausible, controllable worlds and roll robots through them, then distil the result into a VLA. The scarcest resource in robotics — real interaction data — gets synthesised.

Reflex proposes, model vets

Imagination as a safety check

A VLA proposes an action instantly; a world model rolls it forward a few steps and vetoes futures that end badly. System 1 supplies speed, System 2 supplies foresight — much like how you catch a falling cup by reflex but plan a house move in your head.