A note before you start
This post is more technical than the other pieces in our physical AI series. It is written for readers who want the details beneath our position. If you prefer the plain-language route, start with Two ways to predict the world and The real bottleneck in physical AI, then return here.
Hybridisation is real, but task- and access-dependent
Both world models (WMs) and physics engines (PEs) predict the future state of the agent’s body and immediate environment (the “state of the world”) as a result of agent actions. Strictly speaking, WMs and PEs are distinct: a WM predicts by using patterns learned from examples, whereas a PE predicts by numerically solving the physics equations. Learned WMs and explicit physics simulators are increasingly being combined, but they are not collapsing into a single technology. The best description of this trend is hybridisation and cross-fertilisation, rather than a complete merger of two cleanly separated fields.
Based on our internal experimentation and our reading and evaluation of the published literature, our view is that hybrid systems will become more common, while access remains uneven. Some high-performing systems depend on large, closed models and proprietary data, while projects such as OpenVLA and openpi have reduced barriers to fine-tuning and deployment, at least for VLAs (see below). Increasingly, strong pre-trained models are released openly, even as the very frontier remains closed. More importantly, however, the openness of a model does not imply ease of use. The more accurate models tend to be very large, and require extensive effort and compute to utilise properly.
On the simulation side, differentiable physics provides gradients through a simulator, enabling gradient-based system identification, design, planning, and control. This can improve data or sample efficiency in some tasks, but it’s not a universal guarantee: collisions and friction can produce discontinuous or inaccurate gradients. Learned simulators commonly use graph neural networks, neural operators, or related architectures to approximate state evolution.
On the generative side, some recent systems incorporate explicit 3D supervision, physical constraints, or numerical simulation. These additions can improve cross-view consistency, controllability, and physical plausibility. They do not by themselves guarantee physical correctness: current learned visual-prediction systems continue to struggle with latent material properties, contact-rich interactions, novel or unfamiliar (“out-of-distribution”) conditions, and adherence to physical rules. Geometric consistency, visual realism, and dynamical accuracy should therefore be assessed separately.
The practical boundary is consequently one of comparative advantage. Explicit and hybrid simulators are especially useful when forces, constraints, contact behaviour, parameter control, repeatability, and quantitative validation are central. Generative WMs are especially promising when photorealism, semantic diversity, and multiple plausible futures matter. Neither advantage is exclusive: learned models can predict contact dynamics, while procedural simulation and domain randomisation can generate extensive visual diversity.
Embodied AI and autonomous driving are prominent areas of intersection, but they are not the only ones. In gaming, learned interactive engines have appeared so far as standalone experimental titles and research prototypes, while game developers are beginning to adopt world models alongside traditional physics-based engines in their tool-chains, and digital-twin research increasingly combines physics-based models with learned surrogates, corrections, and data-driven decision systems. The relevant leadership question is therefore not whether to choose “physics” or “AI” in the abstract. It is the combination that delivers the required accuracy, coverage, latency, controllability, validation, and accessibility for the task at hand.
A note on a loaded term
The term “world model” is loaded, and precision matters. Several products marketed as WMs are, on close inspection, something else. Often, they are a vision-language-action (VLA) model that maps a language instruction and a video input to an action or a series of actions. The distinction is simple: a true WM predicts the next state of the world as a result of an action, whereas a VLA merely predicts the next action or actions. Some recent WM architectures, however, jointly learn future-observation prediction and action generation. Another group of not-really-world-models are image and video predictors or generators: they predict or generate the next state of the visible world, or the next frames of a video, but they are not action-conditioned. Aicadium applies a strict definition: a true WM is action-conditioned and predicts the next state of an environment you can interact with. Holding to that definition helps leaders separate genuine capability from confident labelling. We set out the three paradigms of world models in our plain-language explainer.
The word “world” (as in “WM” and “state of the world”) refers to the state of the agent’s immediate environment and body. For example, it may comprise the positions and velocities of the robot joints and the surrounding objects, predicted either explicitly or implicitly as the WM’s internal state (a “latent representation”). One should further distinguish between the “state” (e.g., the actual positions of objects and joints) and an “observation” of the state (what a camera sees or a sensor reads). WMs may be trained to predict one or the other, or both. PEs compute the future state (with some accuracy), and a separate piece of software then renders this state into an observation: for example, Autodesk Maya or NVIDIA Isaac Sim can render PhysX predictions as a camera would see them.
The access asymmetry few are discussing
Many of the most capable models are now openly released, yet they remain practically out of reach for smaller teams. In our experiments, the largest and most realistic models performed clearly better at following instructions and producing physically plausible results. They were also the hardest to use. Running a model to make a single prediction is far less demanding than training or fine-tuning it. The best models require server-class computing to fine-tune.
This is the access asymmetry. Open weights do not equal open access when the compute bill is the real barrier. The encouraging side is that large labs now release strong pre-trained models. A smaller team often needs only to adapt one rather than train from the beginning. That lowers the barrier, though it does not remove it. Hardware keeps improving, too, which widens what a small team can attempt each year.
Where the field goes next
A clear architectural trend is emerging from the convergence: hierarchy. Fast control and slow reasoning pull in opposite directions. Manipulation needs very low latency, while planning needs a larger, slower model that understands cause and consequence over time. The promising design places a fast action model (usually a VLA) underneath a slower reasoning model. The central challenge is how the two communicate, especially when something goes wrong or something unexpected happens.
It is worth keeping perspective. Today’s physical AI resembles early large language models: impressive, but prone to errors and inconsistency. Much of the progress since then came from the engineering built around those models, not from the models alone. A similar trajectory in physical AI could make robots viable in many domains that are out of reach today.
Are WMs and PEs merging into one technology?
They are converging but not merging into a single technology. Physics engines are adopting learned and differentiable components, and world models are incorporating physics, yet each still leads in different tasks. Contact-accuracy work favours physics engines, while visual-diversity work favours world models.
Why does the access asymmetry matter for strategy?
The access asymmetry matters because capability and accessibility are diverging. The strongest models are open in principle but expensive to adapt in practice. A realistic strategy plans for adapting pre-trained models on accessible hardware rather than training from scratch.
Aicadium’s view is that the winners in physical AI will not be those with the largest models. They will be those who match the right method to the task and plan honestly around the cost of using it. Convergence rewards judgment, and judgment is something an organisation can build.
Read the rest of the series
This piece is Aicadium’s position on where the field is heading. If you have not yet read the companions, they are shorter and less technical. Start with Two ways to predict the world and The real bottleneck in physical AI. Then finish with our ten-minute guide for chief executives.
