How we build it, in the open.

For the past few years, robotics has been organized around a single bet: that scaling Vision-Language-Action models would deliver general-purpose physical intelligence the way scaling LLMs delivered general-purpose language. Larger models, more demonstrations, more diverse tasks — the recipe seemed familiar.

As these systems enter warehouses and factories, a subtler bottleneck is emerging: not the quality of action generation, but the absence of explicit reasoning about consequences. Generating an action is not the same as deciding on one. The field's response has been a steady architectural shift — from reactive action generation toward predictive decision making — with the world model at its center: first attached to policies as an external simulator, now increasingly fused with action generation into world action models.

Reactive VLA Systems and Their Limits

Most of today's VLA systems — OpenVLA, SmolVLA, the π0 family — follow this recipe, and it works remarkably well. But their reasoning about the future is implicit at best, embedded in the training distribution rather than expressed as an explicit step.

Consider grasping a deformable polybag. Several grip forces look reasonable from the current observation alone — but will the bag slip, deform unevenly, stay stable through the lift? A VLA produces an action without separately evaluating what it implies a half-second later. For repetitive, well-bounded tasks that is often enough. For contact-rich, long-horizon manipulation, it leaves a class of failures that pure scaling has struggled to close.

What Is a World Model — and Why Now

A world model is a learned model of dynamics: given the current state of the world and a candidate action, it predicts the future state that would result. A VLA answers "what action fits this observation?"; a world model answers "what happens if I take this action?" — the question that separates a reflex from a plan.

The idea is old — model-based RL and MPC have built control on dynamics models, hand-derived or learned, for decades. What changed is scale and medium: video models trained on internet-scale footage turn out to encode a great deal of physical interaction and scene evolution — imperfectly, but at a scale no robot fleet can match. If next-token prediction was the first pre-training paradigm, next-physical-state prediction is shaping up to be the second.

This framing also explains the VLA ceiling. VLA backbones come from image–text pretraining — strong on what things are, weak on how things move. Video pretraining supplies exactly that missing signal, and unlocks a data source robots can't generate themselves: internet-scale and egocentric human footage, orders of magnitude cheaper than robot demonstrations.

First Role: The World Model as Simulator

In its first wave of adoption, the world model sat beside the policy as an external simulator: generate several candidate actions, roll each forward through learned dynamics, inspect the predicted futures — classical MPC in structure, with dynamics learned from data. The same model doubles as a synthetic data engine and a policy evaluation environment; NVIDIA's Cosmos platform and Wayve's GAIA-1 embody this role, and Qwen-RobotWorld states the job description outright — synthetic data generation, policy evaluation, and planning signals for downstream control.

Its cost shows at deployment: rendering photorealistic video for every candidate is slow and compute-hungry, and video fidelity is not decision quality. The robot doesn't need to render the future — it needs to know whether the grasp will hold.

From Simulators to World Action Models

The answer is to fold prediction and action into one network. A world action model (WAM) jointly learns to predict future states and the actions that bring them about, starting from a backbone pretrained to predict how scenes evolve.

NVIDIA is converging its Physical AI stack on this architecture from both ends. Its world foundation models are growing into policies — Cosmos Policy post-trains a video world model into one, and Cosmos 3 extends an omnimodal world model directly into a world-action model — while its flagship robot policy line is being rebuilt on world modeling: DreamZero, which reports over 2× better generalization to new tasks and environments than state-of-the-art VLAs, and GR00T N2, its productization previewed at GTC 2026 for release by year end. The two directions meet in the middle.

τ0-WM and Being-H0.7, from AgiBot and BeingBeyond, bracket the key deployment question this leaves open: what should prediction cost at runtime? τ0-WM spends test-time compute on it deliberately, rolling candidate actions forward and scoring them before anything touches the world. Being-H0.7 spends none: it compresses future-aware reasoning into a latent state compact enough to live inside the control loop, with no rollout at all. Both land on the same lesson — prediction that improves control, not video fidelity.

Fusion isn't the only endpoint, either. π0.7 keeps its VLA intact and places a world model above it, generating subgoal images that steer the policy asynchronously from outside the control loop. The field hasn't converged on where the world model should live: inside the policy, beside it, or above it. What has converged, at least among the systems here, is that each carries an explicit component whose job is to predict the future.

From Prediction to Decision

Prediction alone doesn't decide. A world model can tell you the polybag will slip under one grasp and hold under another; it cannot tell you which outcome is acceptable. That takes a value signal that scores futures against the task, and a procedure that selects among candidates accordingly.

The loop is familiar from game-playing AI — something proposes, something predicts, something judges, and a selection rule binds them. AlphaGo ran it on a perfect simulator of Go's rules; MuZero replaced that simulator with a learned model. Robotics is now assembling the same loop around a learned model of the physical world. Cosmos Policy generates candidate actions, their predicted futures, and their values from a single network, then executes the highest-value one — best-of-N, in the paper's own terms; given its own rollout data, it can refine those predictions from experience. τ0-WM splits the roles across one shared backbone: its video-action model proposes, and its simulator's task-progress scores evaluate, driving a propose–evaluate–revise cycle before execution. In both, the world model is no longer just a predictor; it is an action-consequence evaluator. That is the difference between generating actions and deciding on them.

Why This Matters for Industrial Robotics

The consequences are clearest in industrial settings, where the marginal failure is rarely a missing capability — it's an unreliable one. Modern VLAs can generate plausible grasping motions for polybags and flexible packaging; the hard question is whether a grasp survives contact, lift, and transport. That depends on contact geometry under load, grip force against material stiffness, slip during acceleration — properties of futures the robot has not yet caused.

This is where the latent-space direction pays off. For many industrial tasks, the future can be collapsed to a single decision-relevant property — will this grasp hold — and a compact model that predicts just that, fast enough to run inside the control loop on edge hardware, is worth more than a photorealistic video model. The goal isn't to render the future. It's to choose well within it.

One Flywheel Becomes Three

The shift also reshapes data strategy. In the pure-VLA paradigm, operational data trains one thing: the policy. In the predictive paradigm, the same data feeds three learners. Successful picks refine the policy. Every transition — success or failure alike — is free supervision for the world model: a recorded mapping from state and action to what happened next, no labeling required. And the outcome labels, with failures and near-misses carrying the most information, train the value function that separates grasps that hold from grasps that slip.

The fleet becomes a substrate for compound learning, where each robot's experience improves several layers of intelligence at once. The companies that build that compounding structure may hold a more durable advantage than those scaling a single policy.

From Acting to Deciding

Robotics taught its models to act. It is now teaching them to imagine — and increasingly, the models that imagine best are the same ones that act. What remains is teaching them to decide.

The defining question of Physical AI is shifting from "what action should the robot take?" to "which future should the robot choose?" — from reactive action generation toward predictive decision making. That is not a single breakthrough waiting to happen but a restructuring of how robotic systems are built, and the systems that internalize it earliest will be the ones that scale.