World Models, Metaverse, Digital Twins, and Physical AI: Are They the Same?

Over the past few years, concepts such as the metaverse, Web3.0, simulation data platforms, digital twins, and physical AI have emerged one after another, making it easy for the average person to become confused.

What is their relationship to world models?

The answer is: They are not exactly the same thing, but both point to the broader trend of the blurring boundary between the digital and physical worlds.

The world model acts more like the "cognitive layer" or "underlying operating system" of these concepts, responsible for enabling AI to understand and simulate the world.

Over the past few years, tech industry buzzwords can generally be divided into three categories.

The first category is "spatial experience," exemplified by the metaverse. It aims to enable humans to socialize, work, consume, and live in virtual spaces.

The second category is "relations of production," exemplified by Web3.0, which aims to reconstruct data ownership, identity, and incentive mechanisms using blockchain technology.

The third category is "technical capabilities," including simulation data platforms, digital twins, physical AI, and world models. All of them aim to understand, simulate, predict, or generate the physical world through digital means.

World models belong to the third category, but they are more fundamental.

It is not a specific application, but rather a capability that enables AI to build a logically inferable world in its mind. The metaverse may rely on it, simulation data platforms are its predecessors, digital twins are its close relatives, physical AI is its host, and Web3.0 essentially operates on a different technological layer altogether.

Let me break it down one by one.

At the height of the metaverse craze, people envisioned an immersive virtual society featuring avatars, virtual real estate, digital assets, online concerts, and remote work. Its core is a spatial experience: people can enter, socialize, consume, and create.

However, the biggest bottleneck for the metaverse at the time was content creation. Building a virtual city required massive amounts of art and engineering resources, at extremely high costs, yet the experience remained rudimentary. Many projects ultimately became empty showrooms or speculative land trades, leaving users unsure of what to do after a brief walkthrough.

If mature, world models could directly generate interactive 3D worlds from text, effectively acting as an "auto-generator" for the metaverse. Google Genie 3 has already demonstrated a prototype: input a single sentence, and it generates a world you can explore in real time. In the future, you might simply say, “I want to take a walk along the Bund in 1920s Shanghai,” and the world model will generate a street, a set of NPCs, and a storyline for you.

So they are not the same thing. The metaverse is the “destination,” while world models are the “tools for building roads and cities.” World models don’t have to be turned into a metaverse, but for a metaverse to achieve low-cost, large-scale, and interactive experiences, it likely cannot do without world models. The aspects of the metaverse that remain unrealized may be fulfilled by world models.

The core of Web3.0 is blockchain, decentralization, token economics, and user ownership of data. It aims to address issues of ownership and incentives on the internet, not “how the world is understood and simulated by machines.”

For example, world models explore how AI mentally simulates the world, while Web3.0 investigates who owns digital assets in this world and how they are traded. These two can be combined—for instance, trading land via NFTs within a virtual world generated by a world model, or using DAOs to govern rules for virtual cities—but their core technologies are fundamentally different.

So Web3.0 and world models are not fundamentally the same thing. Their relationship is more like this: Web3.0 could be the "economic rules" of future virtual worlds, while world models are the "physical rules." One is a social science issue, the other an engineering problem.

This is the closest. Over the past few years, autonomous driving companies have invested heavily in simulation platforms such as CARLA, 51World, Unity Autonomous Driving Simulation, and NVIDIA DRIVE Sim. Their core value lies in generating extreme scenarios in a virtual environment to enable cost-effective training of autonomous driving algorithms.

The problem with these platforms is that most scenarios require manual setup or rule-based generation. Corner cases such as heavy rain, snowstorms, irregular obstacles, and pedestrians suddenly crossing the road must be modeled one by one by designers, resulting in low efficiency. Additionally, rule-generated scenarios often lack naturalness, causing algorithms to overfit to artificial patterns after extensive training.

What the world model does is use AI to automatically generate these scenarios. Instead of relying on designers to manually place obstacles, it learns physical laws from real-world data and generates variants that are nearly indistinguishable from reality. XPeng claims that its world model enables simulation tests equivalent to 30 million kilometers driven per day, while Horizon can generate a controllable driving video in just 30 seconds.

Therefore, the simulation data platform and the world model can be viewed as version 1.0 and 2.0 of the same concept. The former relies on human effort and rules, while the latter leverages AI generation. The world model does not negate the value of the simulation data platform; rather, it intelligently automates and scales it.

Digital twins have gained significant popularity in recent years across industries such as manufacturing, urban planning, and energy. At their core, they create high-precision, 1:1 digital replicas of the physical world. For example, building a digital twin of a factory allows real-time synchronization of equipment status for monitoring, maintenance, and optimization. Similarly, creating a digital twin of a city enables simulation of traffic flow, pipeline pressure, and disaster response.

Digital twin is a mirror of the present. It answers the question: How is the real world right now?

The world model is a “simulation sandbox for the future.” It must not only understand the current state of the factory but also predict: if this production line accelerates, will the equipment overheat? If the robot moves this way, will it collide with the shelves? If a typhoon arrives tomorrow, how will the power grid load change? It answers the questions: What will the real world become, and how should I act?

Therefore, a world model incorporates some capabilities of a digital twin but goes one step further: from “replicating reality” to “simulating the future.” You can think of a digital twin as a component or prerequisite of a world model, but the ambitions of a world model are greater.

Jensen Huang and NVIDIA have recently been promoting "Physical AI"—AI that can act in the physical world—such as autonomous vehicles, humanoid robots, industrial robotic arms, and drones.

For a physical AI to act, it needs three things: - Perception: seeing the world; - Understanding: knowing the laws of the world; - Decision-making: choosing actions.

The world model handles the middle layer—understanding the laws of the world and predicting the future. It enables AI not just to detect obstacles ahead, but to anticipate how those obstacles will move next and what outcomes different actions will produce.

So you could say that the world model is a core component of physical AI, but not the entirety of it. Physical AI also includes sensors, actuators, control algorithms, safety systems, and more. The world model serves as the “cerebral cortex” of physical AI, responsible for simulating and reasoning ahead before taking action.

If arranged in a hierarchy, it would look something like this:

Underlying infrastructure: Computing power, GPUs, cloud, sensors, data collection

Cognitive layer: World model—understanding and simulating the laws of the physical world

Application Tool Layer: Simulation Data Platform, Digital Twin—Transforming Cognitive Capabilities into Training or Monitoring Tools

Action Layer: Physical AI—robots, autonomous vehicles, and other entities that act in the real world

Experience Layer: The Metaverse — A Virtual Space Where Humans Are Immersed

Rule Layer: Web3.0 — Ownership, Identity, and Economic Incentive Rules

The world model resides at the "cognitive layer," supporting application tools, action systems, and virtual experiences from above, while depending on computing power and data from below. It is not any single concept itself, but may serve as a common foundation for many concepts.

These concepts are easy to confuse because they all point to the same overarching trend: the blurring of boundaries between the digital and physical worlds.

The metaverse aims to have humans spend more of their lives in the digital world;

Web3.0 aims to make digital assets owned by individuals;

The simulation data platform aims to train AI for the physical world using the digital world;

Digital twin aims to synchronize two worlds in real time;

Physical AI aims to enable AI to act in the physical world;

A world model enables AI to possess an internal, simulatable representation of the world, serving as a cognitive layer that connects the digital and physical realms.

The world model doesn't necessarily replace these concepts, but it could become the underlying infrastructure for many of them. Just as an operating system doesn't replace apps, but all apps run on top of it, metaverses, simulation platforms, digital twins, and physical AI—these apps—will ultimately need the world model as their operating system to coordinate their understanding of the world.

So, are the concepts that were previously hyped the same as world models?

Strictly speaking, no.

But many of the concepts once hyped may ultimately require world models to be realized.

—END—

This article is from the WeChat public account "IT Juzi" (ID: itjuzi521), authored by Judy.