
The virtual worlds left behind by the metaverse welcome robot "natives".
By Zen, PANews
On September 8, an American startup named Antioch completed a $32 million Series A funding round led by well-known Silicon Valley venture capital firm Greylock. Adding an $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 core approach involves digitizing real-world robots, sensors, software systems, and operational environments, enabling large-scale training, testing, and validation of machines within a digital world that closely mirrors reality. Here, failure modes and extreme scenarios that are prohibitively expensive, difficult to replicate, or even dangerous in the physical world can be simulated repeatedly.
Antioch is hardly alone. As Physical AI emerges as a new investment direction in the AI industry, a wave of companies centered around 3D spaces, digital twins, physics simulation, and synthetic data are regaining capital attention. At their core, they all attempt to solve the same problem: how to create a world close enough to reality inside a computer.
Looking only at descriptions like "building virtual worlds and replicating real-world scenes," it's easy to recall the previous tech wave of the "Metaverse." But the difference lies in this: Antioch isn't creating a space for users to roam freely wearing VR headsets. This time, the entities that truly need to "live" in a virtual world long-term are robots.
Virtual worlds haven't disappeared; it's just that the "residents" living in them have changed this time.
Humans Never Moved into the Metaverse
In August 2021, at SIGGRAPH, one of the world's most important conferences on computer graphics, NVIDIA announced an expansion of its Omniverse platform. At the time, NVIDIA's positioning for it was very clear: Omniverse is a platform for simulation and 3D collaboration, and it is "laying the foundation for the metaverse."
Its core concept is to bring 3D content originally scattered across different software applications into a single virtual world.
One of its crucial underlying technologies is called USD, or Universal Scene Description. Originally developed by Pixar for producing complex animated films before being open-sourced, it serves as a 3D scene description standard. Simply put, it acts like a "universal language" for the 3D world: what objects exist in a virtual scene, where they are located, what materials they use, how they move, and how they relate to each other can all be organized through USD.
Consequently, designers using 3D modeling software like Blender, artists crafting textures and materials with Adobe Substance 3D, and engineers working with various engineering tools could connect their respective outputs to Omniverse via USD, collaborating within the same 3D environment.
NVIDIA even described this future as a "3D Internet": while today's internet connects web pages, the internet of the future might connect countless 3D spaces that users can enter and interact with in real time.
Later that October, Facebook rebranded itself as Meta, pushing the metaverse into its most feverish phase.
Compared to NVIDIA, Zuckerberg envisioned a far grander future. In his conception, people would no longer just use the internet through two-dimensional screens but would step into a digital world with spatial depth and a sense of "presence," holding meetings, working, playing, shopping, and socializing within it. At the time, Meta projected that the Metaverse could reach one billion people within a decade, spawning a massive digital economy.
The metaverse became an umbrella term capable of containing everything. Over the following two years, narratives around VR, AR, digital humans, virtual offices, 3D gaming, digital twins, and digital assets like NFTs were all folded into a unified vision of a "massive digital world."
However, the consumer-grade metaverse industry has yet to materialize, with a significant number of early participants quietly fading out. While Meta hasn't abandoned VR, AR, and smart glasses, Reality Labs—the division backing these initiatives—remains an extremely costly long-term investment. In 2025, Reality Labs generated approximately $2.207 billion in annual revenue but recorded an operating loss of $19.193 billion. Meta simultaneously forecasted that the segment's operating losses would remain largely at the 2025 level in 2026.
By contrast, the evolution of Omniverse is more intriguing. Five years ago, NVIDIA dubbed Omniverse the "foundation for the metaverse." Today, NVIDIA's official website defines it as "a set of libraries and microservices for developing Physical AI applications."
Omniverse's overarching narrative and product focus have shifted toward industrial digital twins, robotics simulation, and autonomous driving development. Beneath it lie a suite of robotics and Physical AI tools including Isaac Sim, Isaac Lab, Cosmos, PhysX, and Warp. The dominant narrative carried by the same platform has fundamentally changed.
In reality, the history of robotics simulation, industrial digital twins, and physics engines predates the "metaverse origin year" by far. The previous metaverse wave merely bundled relatively disparate technologies—such as 3D modeling, real-time rendering, digital twins, virtual collaboration, open 3D standards, and NFTs—into a unified imagination of a "massive digital world."
Thus, after the metaverse bubble deflated, these infrastructures did not vanish. People simply gradually realized that robots might need a virtual world far more than humans do.
Why Robots Need a Virtual World
This year at the GTC conference, NVIDIA once again showcased a highly intuitive "robotics data pyramid" while explaining its humanoid robot training framework.
The bottom layer consists of the largest volume of internet and human data, such as web pages, images, videos, and behavioral records. Virtually inexhaustible, they inform robots about what the world looks like and what humans do, but typically lack the joint angles, forces, tactile feedback, and control signals required when a robot actually performs movements.
Topping the pyramid are real-world robot data. Having robots genuinely grasp, transport, or walk via teleoperation or autonomous operation yields data closest to the final task, but it is also the most expensive: a single robot can physically experience at most 24 hours per day, and it entails costs related to hardware, facilities, operators, and equipment maintenance.
Bridging these extremes are simulation and synthetic data. Less precise than real-world robot data, it can be rapidly replicated and generated in parallel within GPUs. Unlike internet videos that merely let us "see what humans are doing," it can directly generate the actions, states, and sensor data necessary for robots to execute tasks.
Hence, NVIDIA proposes continuously expanding the synthetic data layer in the middle of the pyramid, aiming to make it a critical data source for robot training.
The pros and cons of these three data types are straightforward: internet and human data are cheap and vast in scale but far removed from actual robot execution; real-world robot data is closest to physical deployment and offers the highest physical fidelity, yet it is expensive and slow; simulation data sits in the middle—sacrificing some authenticity, but vastly outperforming real hardware in scale, cost, and controllability.
Precisely because of this, simulation is emerging as one of the most heavily backed directions in U.S. Physical AI infrastructure by both capital and major tech firms.
What it solves is not just a question of data quantity. The defining feature of the real world is its lack of infinite reproducibility. If a company owns 100 robots, it can only run 100 robots simultaneously; a warehouse cannot rearrange shelves, lighting, flooring, and merchandise every ten minutes purely for robot training. In a virtual environment, however, a single robot can be duplicated into thousands or even tens of thousands of digital avatars.
A robotic arm can simultaneously confront 1,000 different desktop layouts; robots can endlessly alter object positions, masses, and materials; lighting, camera angles, ground friction, and even sensor errors can be manually randomized. This training methodology in robotics is known as Domain Randomization. Instead of forcing robots to memorize a single perfect virtual environment, it deliberately varies conditions continuously, enabling models to learn how to accomplish the same task across diverse variations.
Additionally, beyond data captured during "normal operations," developers must also prioritize edge-case scenarios. Sudden camera failures, floors unexpectedly becoming slippery, or robots falling are all situations that need to be anticipated and overcome. It's impossible to engineer daily accidents in the physical world just to collect data, but a virtual world can accommodate that effortlessly.
This is also why some argue that robotics development cannot rely indefinitely on physical test ranges. Real-world testing demands equipment, facilities, and engineering hours, and extreme failures are hard to replicate. In simulation, however, a single system update can instantly trigger parallel testing across thousands of conditions.
The biggest hurdle the metaverse once faced was asking why humans should necessarily enter a virtual world. Robots now provide a completely different answer. For people, a virtual world is merely an optional alternative to real life. For robots, however, a virtual world can serve as a proving ground before entering the physical realm.
The "Virtual World" Reemerging as a Business
Centered around Physical AI, a new "virtual world infrastructure" industry is emerging, with these companies' divisions of labor growing increasingly specialized.
In August 2026, Munich-based NavVis completed a $85 million funding round.
NavVis is a spatial data company. Its core business isn't training robots, but rather using mobile scanning devices to rapidly capture real-world factories, buildings, and infrastructure into high-precision 3D data. The company disclosed that in 2025 alone, its systems processed and distributed physical spaces exceeding one billion square meters.
Today, NavVis explicitly positions this business as the "data foundation" for Physical AI—if robots are to work in real factories in the future, computers must first accurately know what those factories look like.
South Korean startup NdotLight tackles another layer of the puzzle. Specializing in generating 3D assets for robotics simulation, NdotLight is a Physical AI data company. This August, it secured 15 billion KRW (approximately $10.6 million) in new funding, led by Korea Development Bank (KDB).
Standard 3D models usually only need to look convincing enough to be used in games or animations. But that falls drastically short for robot training, which also requires knowing how heavy a chair is, where collisions will occur, the surface friction coefficient, and which components are movable.
NdotLight's TRINIX transforms standard 3D content into what are known as SimReady assets—3D models capable of being fed directly into robot simulators. Beyond visual fidelity, it augments objects with mass, friction, joint structures, and collision data, integrating seamlessly with NVIDIA Omniverse and Isaac Sim. It solves how to ensure virtual objects not only mirror reality but can also be manipulated by robots just like physical items.
On September 17, another Icelandic startup, Treble, secured $18 million in funding. Treble specializes in acoustic simulation. Once it constructs a virtual space, it can model sound propagation across rooms, buildings, and other environments, generating synthetic acoustic data.
As robots, smart glasses, and other Physical AI devices grow increasingly reliant on microphones to comprehend their surroundings, the company is deploying its acoustic digital twin technology for Physical AI training. Its primary focus is answering how robots can "hear" acoustically realistic sounds within virtual environments.
Collectively, these companies are effectively reconstructing a digital world, bringing it remarkably close to what people called a "virtual world" five years ago. The distinction, however, is that while the metaverse chased immersion, Physical AI pursues fidelity.
This ultimately circles back to the most critical concept in robotics simulation: "Sim-to-Real." It explores a fundamental question: Can knowledge acquired by robots in virtual environments actually hold up once applied in the real world?
This explains why NVIDIA's Isaac Sim emphasizes more than just visual rendering, incorporating rigid body dynamics, joints, collisions, and various virtual sensors to bridge this gap as much as possible. Meanwhile, Isaac Lab further empowers developers to run numerous robot environments simultaneously, leveraging GPUs to train robot strategies in parallel.
Therefore, the core distinction between today's robotics virtual worlds and many consumer-focused metaverses lies in the fact that the former's ultimate value must be validated upon return to reality.
Humans never moved into the metaverse as originally envisioned. Yet the roads paved for virtual worlds haven't been entirely abandoned; instead, robots are walking along them, stepping back into reality.















