The Robotics Playbook Is Starting To Look Familiar, And That’s Not Necessarily Good News

Alex de Vigan, CEO & Founder of Physicl, building world-ready data infrastructure powering robotics, world models, and Physical AI systems.

getty

​In my last article, I argued that physical AI is approaching a data bottleneck. The industry is moving quickly toward robots, world models and embodied systems adoption, but the infrastructure required to train those systems is still a fraction of what it needs to be.

Since then, something interesting has happened: Some of those at the forefront of building the future of robotics have started describing the roadmap in almost identical terms.

Recently, Jim Fan, who leads Nvidia’s embodied autonomous research group, described what he called “the great parallel”: the idea that robotics is beginning to follow the same developmental arc as large language models (LLMs).

The only difference here is that the raw material is no longer language. It is the physical world itself.

The Internet Was The Dataset For LLMs; Reality Is The Dataset For Robotics

One reason the LLM explosion happened so quickly is that the internet already existed as a massive training sandbox. Humanity had unknowingly spent decades generating structured text, images and videos at enormous scale.

Physical AI does not have that advantage, and robots cannot train on language alone. They must learn not just what the world looks like, but how it behaves when interacted with.

This is where the comparison with language models becomes useful. The AI industry now increasingly understands that robotics may require its own equivalent of pretraining at internet scale, except the “tokens” are actions, environments and physical interactions. The problem is that reality is much harder to digitize than language.

A warehouse, for example, is a constantly shifting environment filled with edge cases and unpredictable interactions. A robot operating in that environment cannot rely on approximation alone. It requires representations of the world that are physically coherent enough to support reliable decision-making. That’s the main reason why simulation and synthetic data are becoming so strategically important across the industry.

Compute Is No Longer Just Compute

One of the most important ideas emerging from the robotics community is that compute is the key to everything. Historically, robotics faced a scaling problem: collecting real-world training data was slow, expensive and difficult to scale. Simulation changed that equation.

As NVIDIA's Jim Fan recently put it, "compute now equals environment equals data." The idea here is simple: More compute allows developers to generate and simulate more environments, and those environments become the training data that teaches machines how the world works.

The bottleneck shifts from collecting physical experience manually to generating physically consistent environments at scale.

This is the main reason why NVIDIA has invested so heavily into world simulation infrastructure—and another proof that the industry recognizes that physical AI will be dependent on scalable world generation.

But there is an important caveat here. Synthetic data is only valuable if the simulation itself reflects reality closely enough. Otherwise, models risk learning behaviors that collapse the moment they encounter the unpredictability of the real world.

There is also another challenge emerging beneath the simulation conversation. Most discussions assume that enough 3D data already exists to train these systems. In reality, much of the world's physical data was never created, structured or licensed for AI training in the first place.

Language models benefited from vast, internet-scale datasets such as Common Crawl, which gave AI developers access to billions of web pages for training. Physical AI has no equivalent. There is no comprehensive, standardized dataset of the real world that robots and world models can learn from at scale.

And this is where much of the industry conversation still feels underdeveloped.

World Models Need A World To Learn From

The current generation of robotics demos is genuinely impressive. Systems that once seemed years away are now navigating warehouses, manipulating objects and performing complex tasks in real-world environments. But there is a danger in confusing impressive demonstrations with scalable deployment.

A robot completing a task in one environment is a technical achievement. A robot performing that same task reliably across thousands of unpredictable environments is an entirely different challenge.

That is why the companies shaping the next phase of physical AI may not be the ones building the most impressive robots, or the largest models. I believe the future of physical AI relies on those building the underlying invisible data infrastructure: the digital twins, simulation-ready 3D assets and world-generation systems that allow machines to learn from reality before they are asked to operate within it.

The Part Of The Stack Nobody Is Talking About

One of the reasons this moment feels important is that robotics is increasingly starting to resemble mainstream AI development rather than traditional hardware engineering.

Today’s conversation is shifting from hardware and mechanics toward training pipelines, world models and scaling laws.

That does not mean hardware stops mattering. It means the center of gravity is shifting. The new limiting factor becomes the quality of the environments systems are trained on, the diversity of edge cases they encounter and the realism of the simulations they learn from.

This also explains why the pace of progress suddenly feels faster. Once robotics began benefiting from the same compounding effects that accelerated language AI (scale, synthetic data, pretraining and simulation) development cycles compressed dramatically. That doesn’t guarantee success, but it does suggest that physical AI is entering a very different phase than the one we knew even a few years ago.

For me, the most important takeaway here is not that robots are coming. It is that the industry is finally starting to understand what they will require to work reliably in the real world.

The industry now understands the value of compute, and is beginning to understand the value of simulation. What many teams still underestimate is the cost of building the data layer beneath both.

The main question for AI developers is no longer, “Can we build smarter models?” but rather, “Can we give those models enough of the world to learn from?”

How the industry answers that question could define the next decade of physical AI.​

Forbes Technology Council is an invitation-only community for world-class CIOs, CTOs and technology executives. Do I qualify?