Looking only at descriptions like "building virtual worlds and replicating real scenarios" easily reminds people of the last wave of the "metaverse" technology trend.

Written by: Zen, PANews

On September 8, a U.S. startup named Antioch completed a $32 million Series A financing round, led by the well-known Silicon Valley venture capital firm Greylock. Along with the $8.5 million seed round completed earlier this year, the company, which was founded just over a year ago, has raised a total of $40.5 million.

Antioch's main approach is to bring robots, sensors, software systems, and operational environments from the real world into computers, where machines can undergo large-scale training, testing, and validation in a digital world that closely resembles reality. Failures and extreme scenarios that are costly, hard to reproduce, or even dangerous in reality can be repeated here.

Antioch is not an isolated case; as Physical AI becomes a new investment direction in the AI industry, a number of companies focused on 3D spaces, digital twins, physical simulations, and synthetic data are regaining capital attention. What they attempt to solve is essentially how to create a world in a computer that is close enough to reality.

Looking only at descriptions like "building virtual worlds and replicating real scenarios" easily brings back memories of the last wave of the "metaverse" technology trend. However, the difference is that Antioch is not enabling users to swim through a virtual world wearing VR headsets; this time, the ones that truly need to "live" long-term in the virtual world are the robots.

The virtual world has not disappeared; it is just that this time, the "residents" have changed.

Humans Did Not Move into the Metaverse

In August 2021, at the most important computer graphics conference in the world, SIGGRAPH, NVIDIA announced the expansion of the Omniverse platform. At that time, NVIDIA's positioning was very clear: Omniverse is a platform for simulation and 3D collaboration and is "providing a foundation for the metaverse."

Its core idea is to allow 3D content originally scattered across different software to enter the same virtual world.

One very important underlying technology is called USD, which stands for Universal Scene Description. This is a set of 3D scene description standards originally developed by Pixar for complex animated films and later open-sourced. Simply put, it is like a "universal language" for the 3D world: the objects in a virtual scene, where they are located, what materials they use, how they move, and what the relationships between them are can all be organized through USD.

Therefore, at that time, designers using 3D modeling software like Blender, artists creating textures and materials with Adobe Substance 3D, and engineers using various engineering software could connect their respective creations to the Omniverse via USD and collaborate within the same 3D world.

NVIDIA even described this future as a kind of "3D Internet": Today, the internet connects web pages, and in the future, the internet may connect countless 3D spaces that can be entered and interacted with in real time.

In October of the same year, Facebook changed its company name to Meta, marking the entry of the metaverse into its most fervent phase.

Compared to NVIDIA, Zuckerberg's vision of the future was even grander. In his concept, people would no longer use the internet merely through two-dimensional screens but enter a digital world with a sense of space and "presence," where they could meet, work, entertain, shop, and socialize. Meta predicted at the time that in about a decade, the Metaverse could cover a billion people and form a huge digital economy.

The metaverse is a bucket where anything can be thrown in. In the following two years, VR, AR, digital humans, virtual offices, 3D games, digital twins, and NFTs as digital assets were all placed into a unified imagination about a "large-scale digital world."

However, the consumer-grade metaverse industry has yet to materialize, and a large number of initial participants have quietly exited. Meta has not given up on VR, AR, and smart glasses, but the Reality Labs that support these businesses remains an extremely expensive long-term investment. In 2025, Reality Labs reported an annual revenue of approximately $2.207 billion, while its operating loss reached $19.193 billion. Meta also anticipates that the operating loss for this department will roughly maintain the same level as in 2025 in 2026.

In contrast, the changes in Omniverse are more intriguing. Five years ago, NVIDIA referred to Omniverse as "the foundation of the metaverse." Now, the definition provided on NVIDIA's official website has changed to "a set of libraries and microservices for developing Physical AI applications."

The upper narrative and product focus of Omniverse have shifted toward industrial digital twins, robot simulation, and autonomous driving development. It is linked to a series of tools for robotics and Physical AI, such as Isaac Sim, Isaac Lab, Cosmos, PhysX, and Warp. The overarching narrative it carries has changed significantly.

In fact, the history of robot simulation, industrial digital twins, and physics engines predates the "metaverse's inaugural year." The previous wave of the metaverse merely put together relatively scattered technologies such as 3D modeling, real-time rendering, digital twins, virtual collaboration, open 3D standards, and NFTs into a unified vision of a "large-scale digital world."

Thus, after the tide of the metaverse receded, these infrastructures did not vanish. People have gradually realized that robots may need a virtual world even more than humans do.

Why Robots Need a Virtual World

This year, at the GTC conference, NVIDIA once again demonstrated a very intuitive "robot data pyramid" when explaining the human-robot training system.

The bottom layer consists of the largest quantities of internet and human data, such as web pages, images, videos, and records of human behavior. These data are nearly inexhaustible and can inform robots about what the world is like and what people are doing, but they usually lack the joint angles, forces, tactile sensations, and control signals that robots actually need when performing actions.

The top of the pyramid consists of real robot data. Allowing robots to actually grasp, move, and walk through remote operation or autonomous running can yield the data closest to the final task, but it is also the most expensive: a robot can only truly experience a maximum of 24 hours in one day and involves machines, venues, operators, and equipment maintenance.

Between these two levels is simulation and synthetic data. It is not as accurate as real robot data but can be replicated and generated at high speed on GPUs; it does not simply "see what humans are doing" like internet videos but can directly produce the actions, states, and sensor data required for robots to perform tasks.

NVIDIA thus proposed the future goal of continually expanding the synthetic data layer in the pyramid, aiming to make it an important data source for robot training.

The pros and cons of the three types of data are quite simple: internet and human data are cheap, large in scale, but far from true robotic execution; real robot data is closest to actual deployment and has the highest physical authenticity, but it is expensive and slow; simulation data occupies the middle ground—some authenticity is sacrificed, but its scale, cost, and controllability greatly exceed that of the real machine.

For this reason, simulation is becoming one of the most favored directions for infrastructure investments in Physical AI in the United States.

Moreover, it addresses not just the issue of data quantity. The biggest characteristic of the real world is that it cannot be arbitrarily replicated. A company with 100 robots can only operate 100 robots simultaneously; a warehouse cannot possibly rearrange its shelves, lights, floors, and goods every ten minutes just for training robots. However, in a virtual environment, the same robot can be copied into thousands or even tens of thousands of digital copies.

A robotic arm can face 1,000 different table layouts simultaneously; robots can continuously change the position, weight, and material of objects; lighting, camera positions, ground friction, and even sensor errors can be intentionally randomized. This training method is known in the robotics field as Domain Randomization; it does not require robots to memorize a perfect virtual environment but intentionally alters the environment continuously, allowing the model to learn how to complete the same task under various changes.

Additionally, apart from needing "normal operation" data, developers must also pay attention to data for extreme situations. A sudden camera failure, a slippery surface under the robot's feet, or the robot falling are all scenarios to be considered. The real world cannot create accidents every day just to collect data, but the virtual world can.

This is also why some believe that robot development cannot rely solely on real test fields. After all, real testing requires equipment, venues, and engineering time, and extreme failures are hard to replicate. But in simulation, a single system update can immediately face thousands of conditions for parallel testing.

The biggest question the metaverse once faced was, why would people need to enter a virtual world? Today, robots have given a completely different answer. For humans, the virtual world is merely an alternative to real life. But for robots, the virtual world can become a training ground before entering reality.

Rebecoming a Business of "Virtual World"

A new "virtual world infrastructure" industry is emerging around Physical AI, and the division of labor among these companies is becoming increasingly refined.

In August 2026, NavVis, based in Munich, Germany, completed a $85 million financing round.

NavVis is a spatial data company. Its core business is not training robots but rapidly scanning real factories, buildings, and infrastructure into high-precision 3D data using mobile scanning devices. The company disclosed that in 2025 alone, the reality spaces processed and distributed by its system exceeded one billion square meters.

Now, NavVis has directly positioned this business as the "data foundation" of Physical AI—if robots are to work in real factories in the future, computers must first accurately know what the factory looks like.

The South Korean startup NdotLight addresses another layer of issues. NdotLight is a Physical AI data company specializing in creating robot simulation 3D assets. In August of this year, it secured new financing of 15 billion Korean Won, approximately $10.6 million, led by the Korea Development Bank (KDB).

Ordinary 3D models usually just need to be realistic enough to be used in games or animations. However, this is far from sufficient for robot training, as the models also need to specify how heavy a chair is, where collisions will occur, the surface friction, and which parts can move.

NdotLight has developed TRINIX, which converts ordinary 3D content into so-called SimReady assets that can directly enter a robot simulator. Besides appearance, it also adds mass, friction, joint structure, and collision data, integrating with NVIDIA Omniverse and Isaac Sim. It resolves how to make virtual objects not only replicate reality but also be usable by robots like real objects.

On September 17, another company from Iceland, Treble, secured $18 million in financing. Treble is an acoustic simulation company. After establishing virtual spaces, it can simulate how sound propagates in rooms, buildings, and other environments and generate synthetic acoustic data.

As robots, smart glasses, and other Physical AI devices increasingly rely on microphones to understand the real environment, companies are beginning to apply their acoustic digital twin technology to Physical AI training, primarily addressing how robots can "hear" sounds that are close to reality in the virtual world.

These types of companies combined are beginning to piece together a digital world that is very close to what people described as a "virtual world" five years ago. The difference is that while the metaverse pursued immersion, Physical AI pursues authenticity.

This leads to one of the most important concepts in the field of robot simulation, "Sim-to-Real," which explores whether what robots learn in the virtual world can still be applied when they return to the real world.

This is also why NVIDIA's Isaac Sim emphasizes not only visual rendering but also rigid body dynamics, joints, collisions, and various virtual sensors to minimize this gap; Isaac Lab further allows developers to run large numbers of robotic environments simultaneously, training robotic policies through GPU parallelization.

Therefore, the today's robotic virtual world fundamentally differs from the core of many consumer-grade metaverses in that the former's value must ultimately return to the real world for verification.

Humans ultimately did not move into the metaverse as originally envisioned. However, the paths paved for the virtual world have not been completely abandoned, as robots are walking back to reality along them.