Definition
What is a world model?
A world model is an internal representation that lets an intelligent system model aspects of its environment and predict how that environment changes as a result of possible actions. The idea of connecting learning, an internal model and planning predates the current generation of deep learning: Sutton's 1990 Dyna architecture already planned using a learned model of the world.1
The term regained wide attention within modern deep learning through World Models by David Ha and Jürgen Schmidhuber in 2018. Their system learned a compact spatial and temporal representation of an environment and used that representation to learn behaviour.2 Later approaches have developed this principle further. LeCun, for instance, describes a predictive world model as part of an architecture that lets a system reason and plan,3 while DreamerV3 shows how a learned world model can predict future outcomes of possible actions and train behaviour on that basis.4
A world model for precision manufacturing
Subduxion researches how the same principle can be applied to physical manufacturing processes.
To do that, we translate the general world-model idea into a manufacturing question:
state + possible manufacturing action → expected outcome
A state can include information such as geometry, material, part requirements, available machines and the current production step. A possible action can be an operation on a given machine, with a given setup and process choice.
The model then needs to do more than recognise what a part is: it needs to be able to reason — through process-specific AI models — about what is likely to happen if it is made in a particular way: what geometric state follows, how much time it takes, what quality is achievable, what costs arise, and where uncertainty, risk, or the need for exception handling sits.
None of that requires a visual simulation to count as a world model. The internal representation can be a prediction over structured or learned state rather than an image or animation, and the result can surface as text, a route or process plan, a table, a predicted cycle time, cost or risk figure, a predicted geometry state, or, where that is useful, a visual simulation. The output format is a UI choice, not the definition of the world model itself.
That last part is our own research interpretation of the world-model principle applied to precision manufacturing, developed with human-in-the-loop review built in. It is not an existing, standard definition of a "manufacturing world model".
Why this is more than historical data
Historical production data matters, but does not by itself constitute a world model.
A database can tell you which part was made before, on which machine, and how much time was logged for it. A predictive model needs to go a step further: learning relationships between the part, the chosen actions and the realised outcome, so that it can also reason about new parts and alternative manufacturing routes.
That distinction matters for Subduxion. The goal is not only to retrieve manufacturing history, but to develop models that can predict what different manufacturing decisions are likely to mean in the physical world.
Frequently asked questions
Did Subduxion invent the term world model?
No. World model is an established concept in AI and robotics, used well before Subduxion existed. Subduxion applies the idea to precision manufacturing.
Is a manufacturing world model the same as Blake?
No. Blake is the first application Subduxion is building on this research direction, focused today on estimating and work preparation for milling and turning.
Is the manufacturing world model finished?
No, it is a research and development direction. Current work covers geometry, operation, machine and cycle time for milling and turning; broader process planning and physics are further phases.
Notes & references
- Sutton, R. S. (1990). Integrated Architectures for Learning, Planning, and Reacting Based on Approximating Dynamic Programming. Proceedings of the Seventh International Conference on Machine Learning, pp. 216–224. https://doi.org/10.1016/B978-1-55860-141-3.50030-4 Sutton describes Dyna as an architecture in which learning and planning both use a learned model of the world.
- Ha, D. & Schmidhuber, J. (2018). World Models. arXiv:1803.10122. https://arxiv.org/abs/1803.10122 The authors train a generative neural model to learn a compact spatial and temporal representation of an environment.
- LeCun, Y. (2022). A Path Towards Autonomous Machine Intelligence. Version 0.9.2, OpenReview. https://openreview.net/forum?id=BZ5a1r-kVsf This is explicitly a position paper, not a peer-reviewed research paper; LeCun proposes a configurable predictive world model as part of an architecture for reasoning and planning.
- Hafner, D., Pasukonis, J., Ba, J. & Lillicrap, T. (2025). Mastering diverse control tasks through world models. Nature 640, 647–653. https://doi.org/10.1038/s41586-025-08744-2 DreamerV3 uses a learned world model to predict future outcomes of possible actions.