Artificial intelligence (AI) can now write software, analyse contracts and describe complex scenes with remarkable accuracy. Ask it how to pick up a red cup from a kitchen counter and place it on a shelf, and it will produce a perfectly sensible sequence of steps.
But there is a catch, and it is the defining challenge of the next era of AI. A model that can describe those steps cannot yet produce the motor commands to carry them out. Understanding the physical world is not the same as acting in it.
That is why world models are becoming a central part of the AI conversation. They are the systems designed to help machines build an internal representation of how physical reality works – how objects move, how space changes and how actions lead to what is likely to happen next. Deloitte’s 2026 enterprise AI research suggests that an operational shift is already underway, with world models emerging as a practical layer for simulation, physical AI and decision-making. In the race towards unlocking general intelligence, world models could be the missing piece to the puzzle.
Why language is not enough
For the past decade, AI has been trained mainly on text and images. That has produced astonishing results, but language is still a poor substitute for lived experience. Words can describe gravity, friction and balance, but they do not provide direct exposure to weight, force or timing.
Physical intelligence demands something different. A robot navigating a kitchen, a vehicle interpreting traffic or an arm assembling a component must constantly update its understanding of the world in response to motion, contact and uncertainty. The requirement to make this happen successfully, doesn’t just rely on lightning-fast perception, but also accurate predictions.
World models are the answer to that prediction requirement. They enable an AI system to infer what comes next before acting, rather than learning only after the fact. That makes them a crucial bridge between perception, planning and control, and it is why many researchers now treat them as foundational to physical AI and a plausible route toward more general intelligence.
Why robots are harder than chatbots
Part of the difficulty is data. Language models were trained on the internet, which offered an almost limitless supply of text, all of it in more or less the same format. The data needed to teach machines physical skills is nothing like that. It is expensive to collect, narrow in scope and wildly inconsistent: a navigation route, a tele-operated grasp and a dashcam clip all record the world in incompatible formats, captured by machines with entirely different bodies.
Finding ways to align these incompatible sources into something machines can actually learn from has become one of the most significant problems in AI research, and recent open-source releases show what the solutions are starting to look like.
What embodied AI is becoming
Alibaba’s recent research provides a useful example. In February, DAMO Academy open-sourced RynnBrain, an embodied foundation model built on Qwen3-VL that improves environmental understanding, spatial reasoning and task planning. By releasing it on Hugging Face, GitHub and ModelScope, Alibaba also opened it to researchers and developers around the world.
June brought a more direct attempt at the seeing-versus-acting gap: the Qwen-Robot Suite, three foundation models that each translate language into a different domain of physical action.
The architecture is surprisingly simple. Qwen-RobotNav handles mobility, from spoken navigation instructions to autonomous driving. Qwen-RobotManip tackles physical interaction through a shared representation of states and actions that allows data from very different robots to be trained together. That unlocked more than 38,100 hours of previously fragmented open-source training data.
The third model is the most intriguing. Qwen-RobotWorld is a world model, a system that predicts what the physical environment will look like a moment from now. In practice, it gives a robot something like foresight: the ability to rehearse the consequences of an action before committing to it, rather than finding out the hard way.
This is important because it allows experience gained in one domain to strengthen another. Manipulation improves a model’s understanding of contact physics, driving contributes to large-scale geometry, while navigation develops richer spatial reasoning. Rather than training separate systems for every task, embodied AI starts to build a shared understanding of how the physical world works.
From instruction to action
The larger significance lies in what happens when these pieces are put to work together. Because each model accepts plain-language instructions, a general-purpose AI can orchestrate them much as it would any other tool. One model plans, another navigates, another manipulates, allowing higher-level reasoning to translate into physical action.
In one demonstration, a high-level model takes the instruction “clean the table”, breaks it into simple steps, hands each one to the manipulation model, watches progress and replans when something stalls.
In another, an agent sent to find an open restroom in an unfamiliar building spots a “cleaning in progress” sign on the first one it reaches and reroutes to search the other side of the building. These are small moments, but telling ones: general intelligence is beginning to translate directly into physical action.
The bigger prize
In many ways, this is what the next phase of AI is really about. Not just systems that can explain the world, but systems that can model it, anticipate it and operate within it.
If language models taught AI to talk, world models may teach it to understand reality. And if that understanding can be transferred across robots, environments and tasks, then the industry may be moving toward something even larger: general intelligence that is not confined to text, but grounded in the physical world.
