What Is Generalist AI and Why It Matters

Generalist AI is an American startup with bases in the San Francisco Bay Area and Boston. Its ambition is to create what it calls "physical AGI" — a universal artificial intelligence capable of understanding the physical world and controlling various types of robots within it. And these are no newcomers. The founding team has behind it the development of models such as PaLM-E (the first large multimodal model for robotics at Google), RT-2 (a model translating images and language into robotic actions), and the scaling of ChatGPT and GPT-4 to hundreds of millions of users. Engineers from Boston Dynamics brought experience from building the Atlas, Spot, and Stretch robots.

In other words: the group that stood at the birth of the most important robotics and language models of recent years has come together under one roof — and investors have just given them another $400 million.

Who Invested in Generalist AI

The funding round was led by Radical Ventures, a fund specializing in AI startups. It was joined by 8VC, Union Square Ventures, Hanabi Capital, Norwest, and all major existing investors including Nvidia NVentures, Boldstart Ventures, Spark Capital, Bezos Expeditions (Jeff Bezos's investment entity), and NFDG. Among the new angel investors are names like Fei-Fei Li (an icon of computer vision from Stanford), Bin Lin (co-founder of Xiaomi), and entrepreneur Naval Ravikant. In total, the startup has already raised over $500 million.

Nvidia's presence among the investors makes sense — Generalist AI needs massive computing power to train models on petabytes of physical interaction data. And Nvidia, in return, sees in robotic foundation models yet another giant market for its GPUs.

From GEN-0 to GEN-1: How Robots Had Their "ChatGPT Moment"

The development story of Generalist AI is surprisingly analogous to what we know from large language models.

In November 2025, the team released the GEN-0 model, which was the first to demonstrate the existence of scaling laws in robotics — a mathematically grounded principle that a larger model trained on more real-world data produces a more capable robot. This is the same principle behind the explosive growth in capabilities of models like GPT-4 or Claude. GEN-0's average success rate on tested tasks was 64% — not enough for production deployment, but it showed the path forward.

In April 2026, the GEN-1 model arrived, pushing the boundaries significantly. Across six different manual tasks — from assembling car parts to folding T-shirts to packaging phones — it achieved an average 99% success rate. And at a speed roughly 3× higher than the best previous systems (including the competing π₀ model from Physical Intelligence). For example, the robot folds a box in 12 seconds — while GEN-0 or π₀ needed 34 seconds for the same box.

The Data Doesn't Come From Robots, but From Humans

One of Generalist AI's key innovations lies in the way training data is collected. While the traditional approach requires expensive teleoperation devices (a human remotely controls a robot and the system learns from it), Generalist AI trains its foundation model on data from cheap wearable devices on humans performing millions of everyday activities. The dataset now contains over half a million hours of high-quality physiological interaction recordings — and contains no robotic data whatsoever.

When the model is then adapted to a specific robot and a specific task, it needs only about a single hour of robotic data. This is a radical decrease compared to earlier methods, which required hundreds to thousands of hours of teleoperation. And it is precisely this data efficiency that makes the entire approach economically viable.

Improvisation: When the Robot Starts "Thinking"

Perhaps the most interesting capability of GEN-1 is what researchers call "improvisational intelligence". During car parts assembly, for example, a washer might slip out of the gripper. Instead of stopping and waiting for human intervention, GEN-1 can spontaneously choose one of several strategies: place the washer back and grasp it again, use the edge of the slot to recapture it, or even engage the second hand for bimanual manipulation.

This behavior was never trained — it emerged as an emergent property of the model, similar to how language models began to "understand" context even though no one explicitly explained it to them. It is precisely this ability to react to unexpected situations that separates a lab demo from a system that can actually work alongside humans.

Competition Isn't Waiting: Physical AI as the Next Battleground

Generalist AI is not alone in the field of robotic foundation models. Physical Intelligence (π), a startup led by Karol Hausman, has raised over $400 million and is developing its own π₀ and π*₀.₆ models. Skild AI raised $300 million at a $1.5 billion valuation. And let's not forget Figure AI (humanoid robots, $2.6 billion valuation), Tesla Optimus, or Chinese companies like AgiBot and Unitree, whose humanoids are already heading to Europe.

Interestingly, Generalist AI differentiates itself philosophically: it is not tied to a specific type of robot. Their models are meant to work across humanoids, industrial arms, warehouse robots, and even autonomous systems in space. As they say: "The future of robotics is bigger than any single robot."

What This Means for the Czech Republic and Europe

The Czech Republic has one of the highest densities of industrial robots in the world — according to the IFR (International Federation of Robotics), we have long ranked among the top twenty in manufacturing robotization. At the same time, we face the same problem as the rest of Europe: a labor shortage in logistics and manufacturing.

Robotic foundation models like GEN-1 could solve this problem. If a robot can handle tasks like parts assembly, goods packaging, or machine operation with 99% reliability and without costly reprogramming, it opens the door to automation even in smaller Czech companies, for which traditional industrial robots are too expensive and inflexible.

From a regulatory standpoint, these systems fall under the EU AI Act, which classifies robots in industrial settings as high-risk applications. Generalist AI does not yet operate in Europe — but if it wanted to, it would have to go through certification and meet strict safety and transparency requirements. This could also play into the hands of European alternatives that emerge in the regulated market.

The Flywheel Effect: Why Another Hundreds of Millions Are Heading Here

Generalist AI describes its strategy as a "flywheel": scaling robotic learning creates better models, better models can do more useful physical work, and data from real-world deployment powers the next generation of even more capable models. It's the same loop that drives the exponential growth in capabilities of large language models.

The four hundred million dollars is meant to go toward developing the next generation of models, expanding infrastructure for physical data collection, increasing computing capacity, and first commercial deployments with industrial partners. If the loop starts spinning the way the startup expects, we may witness a moment when a robotic foundation model crosses the utility threshold in the same way GPT-3 did for text.