Before a humanoid robot can open a door without breaking the key in the lock, lift a glass without shattering it, or cross a street without startling a driver, it needs to train extensively. Similarly, before a factory robot learns to react to a bolt falling from a conveyor or another robot suddenly slowing in the work path, it must experience these scenarios repeatedly – thousands of times in situations no one would want to test around humans.
The robot accomplishes all this in one place: the simulator. Nvidia's simulation world, Omniverse (the company's virtual environment platform), serves as the environment where robots are "born." It functions as a cognitive kindergarten where humanoid robots learn to walk, operate, understand, react, fall, and rise. Just as an infant develops cumulative motor and cognitive abilities, the robot learns within an artificial world governed by real-world physical laws.
The simulator generates thousands of situational variations: a glass falling at a different angle, a slightly higher step, weak lighting, a person crossing too quickly in the movement path – to teach the robot to react to as many scenarios as possible.
"If we want to build intelligence that understands the physical world and operates within it, we need to teach it in a world similar enough to reality so it can function within it safely, efficiently, and controllably," Rev Lebaredian, Nvidia's vice president of simulation technologies and Omniverse, said in an exclusive conversation with Israel Hayom.
A defining moment in the journey
Lebaredian joined Nvidia in 2002, after working in the film industry. Early in his career, he worked at production houses like Disney and Warner Bros., and later founded a startup developing advanced rendering technologies. In cinema, the rendering process transforms raw graphics into realistic images that appear as if filmed by a camera – a process that was particularly slow and demanding in the early 2000s, sometimes requiring hours of computation for each frame.
As part of his work, he contributed to creating effects in films like "Armageddon," "X-Men," "The Sum of All Fears," and Disney's "Mighty Joe Young," a film nominated for an Oscar for effects thanks to the digital gorilla character at the story's center.
In the early 2000s, Nvidia was primarily a gaming chip manufacturer, far from the AI giant it is today, valued at approximately $5 trillion. Lebaredian joined exactly when Nvidia's flagship product, the graphics processing unit (GPU), began transforming, and he accompanied the company from the crude computer games era of the early 2000s to today's AI revolution, changing the world at rapid speed.
"I joined Nvidia at a defining moment in its journey, precisely when we launched the ability to program shaders (programmable graphics functions) directly on the GPU. This significantly accelerated rendering capabilities, but more importantly, this was the moment the GPU opened for the first time to free programming. I worked then on the first programming language for graphics processors, CG, which became the first brick on the path to CUDA (Nvidia's parallel computing platform), the language dominating parallel computing today," he recounted.
Today, as head of the company's simulation division – Omniverse – Lebaredian is among the handful of senior executives leading simulation and physical intelligence at the company. Nvidia believes this field will drive the next major technological revolution, bringing artificial intelligence into the physical space of daily life. In this revolution, the division Lebaredian heads will have one of the most significant roles.
"Nvidia CEO and founder Jensen Huang said years ago that the most important algorithms will be those understanding the physical world and capable of influencing it," Lebaredian stated.
From language understanding to world understanding
Those algorithms Huang discussed years ago are materializing today in a new field of artificial intelligence: the world model. Just as a language model learns from billions of sentences to predict which word will come next with the highest probability, and thus essentially understand language, meaning, and context – a world model learns to predict what will happen next in the physical world. Namely, how an object will move, how force will affect, what will happen if a door opens too quickly, or where an object placed at this or that angle will roll.
"A world model is the central foundation of the next revolution: physical intelligence, meaning AI that understands not just words, but the universe," Lebaredian explained. According to him, this is a statistical model developing a probabilistic understanding of dynamic reality, not of text. This model will essentially be the robot's "brain," decoding the environment's visual information and knowing how to operate, where to turn to avoid an obstacle, and what force to apply to crack an egg while making an omelet, for example.
But to do this, it needs data of a type that doesn't exist on the internet. Not words, but material, movement, acceleration, friction, light, temperature, interactions, human environments, and physical infrastructures. The training process is fundamentally similar to that of language models – learning from countless examples and situations – except that here the examples must come from the physical world itself.
"The major problem with physical intelligence," Lebaredian explained, "is that we don't have a digital archive of physics. We need to capture it from reality – and that's expensive, dangerous, and limited. The solution is to recreate reality in simulation, and then produce synthetic data from it."
According to Lebaredian, Nvidia's simulation world is not merely a three-dimensional model. It is an engine of natural laws. A city where every lamppost, sidewalk, car, and tree branch is coded to behave as in reality. In this environment, a robot can walk thousands of simulated years in a short time, accumulating experience impossible in the real world.
Releasing the "genie" from the GPU
To understand Nvidia's role in the AI revolution and the magnitude of the mission the company placed on Lebaredian's shoulders, one must return to the story's beginning – and trace the development of one of recent decades' most influential components: the graphics processing unit.
This development did not amount to gradual increases in performance. This is deep evolution, where each new GPU generation changed the computer's very nature. To such an extent that some believe that without Nvidia, not only would a large language model not function at the required speed, but we might not have imagined the very possibility.
Language models, world models, and advanced robotics all feed on enormous parallel computing power, the kind that needed to be born before theoretical thinking about them became possible. Twenty years ago, the GPU was a dedicated graphics unit designed to accelerate computer games. It was designed as a "drawing machine," receiving a series of fixed commands defining how a three-dimensional object should appear on screen. All stages were rigid: how light falls, how reflection forms, whether the material is shiny or matte. The processor could execute these tasks quickly, but nothing existed beyond this.
"In the early 2000s, everything was very simple and limited," Lebaredian recalled. "You couldn't write your own code. Performance was high, but flexibility didn't exist." According to him, the field's first significant revolution occurred when Nvidia opened the shading stage to programming. Instead of built-in models, developers could write their own functions, recreate light and material laws, and build graphic worlds as they imagined them. The change then appeared as a breakthrough for the gaming world alone, but in practice, it freed the GPU from its initial engineering constraints.
The drawing machine became a machine that understood somewhat more about how the world behaves. The hardware ceased being a black box and became an open platform. This was the moment the seed was planted that later became a computing superplatform.
"I've been at Nvidia for 23 years," Lebaredian said, "and almost throughout this entire period, the company has dealt with the question of what else the GPU can be beyond what it was designed for."
"Far beyond what we imagined"
Lebaredian recounted that as shader programs became more flexible, more and more developers identified potential within the GPU far exceeding graphics. Thus, for example, academic researchers began using the graphics processor for physics calculations – they took the same shading function that calculates light and adapted it to compute airflow, water movement, or particle dynamics. The graphics processor's essence as a computer with powerful parallel computing capabilities gradually became clear.
"We saw researchers using it for things completely unrelated to graphics – physical simulations, fluid dynamics, molecules. This was the moment we understood our processors could serve far beyond what we imagined," he stated.
At this stage, Nvidia understood it must change direction and give this computing body a new form. In 2006, CUDA (Nvidia's parallel computing platform) launched, a software environment allowing regular code to run on the GPU. No more disguising scientific problems as graphics, no more manipulating textures or pixels – but a complete computer capable of processing large arrays, running loops, and executing complex algorithms quickly. Historically, this was the turning point at which the GPU ceased to be a graphics accelerator and became a general-purpose computing engine.
The network that learned to "see"
Here arrived another defining moment in the development of artificial intelligence, made possible by Nvidia's programming language. AlexNet – that groundbreaking 2012 neural network learning to identify objects in images with high accuracy like cats, dogs, cars – ran on CUDA. AlexNet marked the beginning of the past decade's computer vision era, with countless applications from smart security cameras to facial recognition systems in smartphones. That same processor, previously drawing shadows, became a machine learning model to identify complex patterns – learning to "see."
Here, it became clear how critical this link was. Those telling AI's history usually emphasize algorithmics but almost always ignore the fact that behind all this stood infrastructure that realized the vision: parallel computation of enormous data quantities at speeds and prices that enabled the very idea of large models.
In a sense, had the GPU not first freed itself from its graphic constraints, we might not have been able to think about a language model as a feasible project. In retrospect, the GPU appears to have undergone the most dramatic transformation chain in computing history: from drawing machine to scientific computer, from graphics accelerator to global AI engine, and from imaging system to virtual reality source, raising the next generation's robots.
Nvidia did not merely improve the GPU. It reinvented it repeatedly until it became the foundation supporting today's entire artificial intelligence revolution – and likely will be tomorrow's as well. "We are only at the beginning of the process of creating foundational world models. No one will 'own' them or be their exclusive owner – this is a project all humanity will need to contribute to," Lebaredian concluded.
