World Models in Production: Simulation, Evaluation, Limits

Q: What is a world model, and how is it different from a video generator or a physics simulator?

A world model is a generative model whose output is an environment: given a prompt, an image, a video or a sequence of actions, it predicts what the world looks like and does next, and in the interactive variants such as Genie 3 and Cosmos 3 it does so fast enough for an agent to act inside it. The distinguishing property is action conditioning plus learned rather than authored dynamics. A video generator predicts the next frame from past frames and a prompt but does not respond to an agent's actions. A physics simulator such as Isaac Sim, CARLA or MuJoCo computes the next state from equations and asset definitions you wrote, which makes it reproducible with a seed and inspectable down to each contact force. A digital twin is a simulator bound to a specific real asset by live telemetry. Because a world model's dynamics are learned, it can produce scenes nobody built, which is its value, and nobody can audit how it produced them, which is its limit. That inspectability gap is why physics simulators and twins remain where reproducible evidence for incident investigation and safety cases must live.

Q: What are the three production jobs for a world model, and which should a team start with?

Data generation, closed-loop evaluation and in-the-loop planning, in rising order of risk. Start with data generation: the world model expands coverage of the long tail, rain on a route recorded in sun, a pedestrian where none was, a pallet orientation the warehouse never produced, and a bad synthetic frame is caught by the same plausibility filters and lineage tracking that catch any bad training sample. Move to closed-loop evaluation second and only alongside a physics simulator and a real-data baseline, because the score a policy earns inside a generated world is a function of the policy and the world model together, and a generated world that flatters a policy is indistinguishable from a good policy until the field test. Treat planning in the loop, where an agent imagines futures inside the model and picks an action, as narrow in 2026: offline, short-horizon, non-safety-critical decisions verified by a physical sensor before commitment, because the vendors' own limitation statements on action range, multi-agent modelling and consistency horizon mean an action chosen on an imagined future is an action chosen on fiction.

Q: What is the two-simulator architecture and why does it need a real-data arbiter?

The policy runs in two environments and an arbiter decides which to believe per scenario class. The physics simulator owns scenarios you can specify exactly and must reproduce: regression suites, safety cases, contact physics, anything a certifier will ask to see again. The world model owns scenarios you cannot specify and need many of: weather, actors, locations, the long tail. The arbiter is a growing set of real recorded episodes in which both environments are asked to reproduce the recorded outcome from the recorded starting state; whichever diverges more from reality loses trust weight in that scenario class, and promotion of a policy requires the weighted scores to agree within a tolerance that has an owner. Without the arbiter, two scores disagree and the louder engineer wins. With it, disagreement is a measured quantity, trust is earned where it has been checked against reality and withheld where it has not, and a vendor's model update, which changes the physics under every scenario at once, triggers recalibration instead of silently invalidating last quarter's weights. Divergence must be measured on outcomes the policy cares about, such as time to collision or grasp success, not on pixel similarity.

Q: Which world model should a robotics or autonomy team choose in 2026?

It depends on the job and on whether you need weights. NVIDIA's Cosmos 3, launched on 31 May 2026, is an open omni-model that reasons, generates worlds and predicts actions in one mixture-of-transformers architecture, with Super and Nano variants on Hugging Face and build.nvidia.com and an Edge variant announced; it is the choice for teams that need to post-train on their own embodiment and run on their own infrastructure, with the caveat that its benchmark leadership claims are NVIDIA's own and its optimised tooling assumes NVIDIA hardware. Google DeepMind's Genie 3, available through Project Genie to Google AI Ultra subscribers since 29 January 2026, generates interactive worlds at 720p and 20 to 24 frames per second and suits agent exploration and curriculum diversity, but it is hosted only and your environments live inside Google's product. Waymo's World Model, described on 6 February 2026, is Genie 3 post-trained to emit camera and lidar matching Waymo's sensors; it is not a product but it is the template for what sensor-accurate closed-loop evaluation requires, which is post-training on your own fleet data. World Labs' Marble, generally available since November 2025, generates 3D scenes exported as Gaussian splats and meshes and fits teams whose deliverable is an environment to run in their own engine.

Q: What limits do the vendors themselves state for world models?

Google DeepMind's Genie 3 page lists them most explicitly: a limited range of actions agents can carry out, difficulty accurately modelling interactions between multiple independent agents, inability to simulate real-world locations with perfect accuracy, clear text often generated only when it is in the input description, and support for a few minutes of continuous interaction rather than extended hours, with changes recalled for roughly a minute. NVIDIA describes Cosmos 3's physics accuracy as leading among open models, which is a benchmark claim rather than a guarantee of exactness, and lists its Edge variant as coming soon. Waymo's post describes capabilities and does not publish validation figures against real outcomes. World Labs notes that single-image generation in Marble must invent all the detail not present in the input. Read as architecture, these statements mean episodes must fit inside the consistency horizon, multi-agent scenarios cannot be trusted without checking how other actors react to the policy under test, generated environments cannot stand in for a specific real site, and every vendor release changes the physics under every scenario and should trigger recalibration.

Q: Can a world model replace our physics simulator?

No, and the attempt is the most common way a world-model programme goes wrong. A physics simulator is reproducible bit for bit with a seed and inspectable down to each contact, which is what regression suites, incident investigation and safety-case evidence require; a world model is neither, and its own vendors document a consistency horizon of minutes. Retiring the physics stack because the world model looks better means regression tests stop being reproducible, the safety case loses its evidence, and nobody can explain a collision. The right relationship is complementary: the physics simulator owns the measured and repeatable, the world model owns the diverse and unmeasured, and a real-data arbiter decides how much to trust each in each scenario class. Marble's export of generated scenes as meshes with colliders points at a useful hybrid, where generated geometry runs inside a real engine with authored physics, keeping inspectability while gaining diversity.

Q: How should generated training data from a world model be governed?

The same way any upstream model artefact is governed, plus lineage per sample. Every generated sample should carry the world model's version, the prompt or conditioning that produced it and the seed where the model exposes one, so that a regression in a trained policy can be traced to a generation batch and that batch removed from the training set. A physics plausibility filter should reject dynamics that violate known limits, because a world model will generate a vehicle decelerating faster than tyres allow and a policy trained on it learns that it can. Licensing and provenance matter because a policy trained on generated data inherits whatever the world model learned from its training video, and a product shipping into a regulated domain will be asked where its training environments came from; open weights such as Cosmos 3 can be post-trained and audited as far as weights allow, hosted models such as Genie 3 cannot. And the real-data collection programme for the scenario classes you ship must continue, because generated long-tail data is so cheap that cancelling real collection is the natural economic move, and it is the move that stops the arbiter from ever catching the world model's errors.

Q: What does a mature world-model programme look like?

Five stages. Stage 0 is the demo: someone generated a warehouse and walked around in it, useful for a pitch and nothing else. Stage 1 is generation with lineage: the world model feeds the synthetic data pipeline, every sample carries model version, conditioning and seed, a plausibility filter rejects impossible dynamics, and regressions trace to batches. Stage 2 is two simulators and one arbiter: the physics simulator and the world model both run the evaluation suite, scenario classes are defined in advance, real episodes calibrate per-class trust weights on outcome metrics, and promotion requires weighted agreement within an owned tolerance. Stage 3 is domain post-training and governance: an open model post-trained on the product's own sensor data so outputs match the hardware, following the pattern Waymo described; vendor releases treated as a change to every scenario that triggers recalibration; licensing and provenance retained for every generated dataset. Stage 4 is bounded planning: world-model rollouts enter the action loop only for short-horizon, non-safety-critical decisions verified by a physical sensor before commitment, the safety function stays on authored, certified logic, and the arbiter's disagreement history is a first-class metric on the autonomy dashboard.