What is a world model in robotics? AI that predicts the next scene before the robot moves

A world model in robotics is a learned model that predicts what the robot's camera will see next, given the action the robot is about to take, so the robot can try moves in imagination before it makes them. That definition sits on our [world model technology page](/technologies/world-model); this explainer covers what the models are used for, who is paying for them and what the papers say they can and cannot do.

The money arrived first. AMD agreed on Sep 28 to buy [World Labs for about $8.2bn in stock](/news/amd-to-acquire-world-labs-8-2b-physical-ai-20260928), General Intuition [raised $220m at a $6.2bn valuation](/news/general-intuition-220m-6-2b-gameplay-action-models-20260929) a day later, 2.70 times the valuation it disclosed in June, and Runway, NVIDIA and Odyssey all released world models aimed at robots.

[[table:0]] The open question is whether the technology is ready for the role those prices assume.

A world model does three jobs, and only one of them is the robot's brain

The first job is to act as a simulator. Instead of a physics engine written by hand, the model learns from video how scenes change and lets engineers test policies in it. [Runway's Praxis-1](/news/runway-praxis-1-world-action-model-20260930) reported a 0.95 correlation between how policies scored inside its model and how they did on real robots. The second job is to steer or check a policy: [Dream-to-Remember](/papers/dream-to-remember-continual-steering-of-generalist-robot-policies-with-2610-13816) trained a small steering module inside a world model and lifted a frozen policy to 74.8% from 63.7% across eight tasks learned in sequence, where LoRA fine-tuning collapsed to 11.7%. The third job is to be the policy itself, predicting the next scene and the action together, which is what [UBTech's Thinker-WM](/papers/thinker-wm-a-unified-world-model-with-shared-multimodal-representation-2610-18447) and [Sber's GreenWAM](/papers/greenwam-a-world-foundation-model-for-robotics-2610-18774) claim.

The difference with a [vision-language-action model](/news/what-is-a-vla-model-robot-foundation-models-explained-rt-2-pi0-gr00t-gemini-robotics-2-94bn-raised-2026-20261017) is the prediction step. A VLA maps image and instruction straight to motor commands; a world model adds an explicit guess about the future. This month's [TS-WAM](/papers/ts-wam-temporal-latent-actions-and-world-action-coupling-for-vision-la-2610-21922) fuses the two and beat Physical Intelligence's pi0.5 on real robots, 72.5% against 55.0%.

What the October papers measured

MEKAWIRE has ingested 485 robotics papers since Oct 1, and 54 of them, 11.1%, carry the world-model tag; 44 now have desk briefs.

[[table:1]] The results split cleanly. Where a world model checks or adapts something else, gains are large: [ExecutionBridge](/papers/plans-are-not-executions-adapting-frozen-world-model-planners-to-contr-2610-22041) from Tsinghua and DENSO lifted a driving planner's hard-scene score 44% when the vehicle's controller changed. Where a world model must predict physical outcomes over long horizons, it struggles: [WorldOracle](/papers/worldoracle-benchmarking-action-conditioned-world-models-against-real-2610-18826), a benchmark from AgiBot and NVIDIA researchers, found that all seven models tested drifted on long real-robot runs and often got grasps and spills wrong. Even elegant theory falls short in planning: [LeHami](/papers/lehami-symplectic-latent-dynamics-for-world-models-2610-15267), with Yann LeCun as co-author, cut velocity error 2.8 times yet planned worse than the best baseline, 0.527 against 0.800.