The "Scale AI for Robotics" Is Nearing a $1.2B Valuation Just 3 Months After Stealth

XDOF in Late-Stage Talks for Series B Led by 8VC at $1.2 Billion Valuation.

XDOF, Just Three Months Out of Stealth, Is in Talks for a Series B at a $1.2 Billion Valuation:

XDOF is betting that the next major bottleneck in artificial intelligence won't be language data—it will be real-world data for robots.

The race to build truly capable AI-powered robots is entering a new phase, and one relatively young startup is suddenly attracting serious attention from investors. XDOF, a robotics data company that collects real-world teleoperation data for training general-purpose robots, is reportedly in late-stage discussions to raise a Series B at a valuation of around $1.2 billion, with 8VC expected to lead the round, according to people familiar with the deal.

What makes the story particularly remarkable is the timing.

XDOF emerged from stealth less than three months ago.

The company had already raised a $70 million Series A in June, with backing from major venture capital firms including Thrive Capital, Andreessen Horowitz, Lux, and Spark Capital.

At the time, another funding round was reportedly not part of the immediate plan. But things appear to have moved much faster than expected.

According to people familiar with the company, XDOF's rapid growth and annualized revenue approaching $50 million have attracted investors who are now pushing the startup toward another financing round.

The exact size of the Series B has not been disclosed, and it remains unclear whether the reported $1.2 billion valuation includes the new investment. The deal is still being negotiated, meaning the terms could change.

Why XDOF Is Suddenly Worth So Much:

At first glance, collecting data for robots might not sound like the most glamorous business in artificial intelligence.

But that perception is changing quickly.

Large AI models such as ChatGPT benefited enormously from the enormous amount of text, images, code, and other information already available online. Robots don't have that luxury.

A robot needs to understand the physical world.

It needs to know how to pick up a cup without dropping it, fold a shirt, open a drawer, move objects around a kitchen, manipulate tools, navigate unfamiliar environments, and perform thousands of other physical tasks.

There isn't a giant internet archive containing all of that information in a format robots can simply learn from.

That creates one of the biggest challenges in robotics AI:

Where does the training data come from?

That's the problem XDOF is trying to solve.

XDOF Wants to Become the Data Infrastructure for Robotics:

XDOF's business is built around creating the infrastructure needed to collect, organize, and annotate real-world robotics data.

The startup provides data pipelines, robotics data collection tools, teleoperation systems, and annotation infrastructure that AI laboratories and robotics companies would otherwise have to develop themselves.

In simple terms, XDOF wants to become the data-supply chain behind the physical AI industry.

The opportunity is enormous.

As companies move from AI that works primarily with text and screens toward embodied AI, physical AI, and general-purpose robots, demand for high-quality real-world training datasets could increase dramatically.

Investors have therefore begun comparing XDOF with companies such as Scale AI and Mercor—businesses that became important infrastructure providers during the rise of generative AI.

The difference is that XDOF is applying the same basic idea to the physical world.

From a UC Berkeley Research Project to a Billion-Dollar Startup:

XDOF was founded in 2024 by Philipp Wu, who serves as CEO, and Fred Shentu, the company's CTO. Both are researchers associated with UC Berkeley.

The company's origins can be traced back to Wu's doctoral research.

While studying how robots could learn from large datasets, Wu encountered a fundamental problem: there simply wasn't enough large-scale, high-quality robotics data available. The challenge led Wu and Shentu to develop GELLO, a relatively low-cost teleoperation system that allows a human operator to control a robotic arm remotely.

The concept is straightforward but powerful.

Instead of waiting for robots to independently learn how to perform complex tasks, humans can demonstrate those tasks while the system records the movements and interactions.

Those demonstrations can then become valuable robot training data.

The research eventually resulted in an influential robotics paper and helped establish the technological foundation for XDOF.

Human Operators Are at the Center of the System:

One of the most interesting aspects of XDOF's approach is that humans remain heavily involved in collecting the data.

The company combines robot teleoperation with human data collectors who use sensors to record everyday physical activities.

Imagine someone folding clothes, flattening cardboard boxes, manipulating household objects, or performing another ordinary task.

Sensors can capture information about how the person's body and hands move while the activity is being performed.

At the same time, teleoperators can remotely control robots and demonstrate how machines should perform similar tasks.

The result is a richer dataset designed to help AI systems understand the relationship between human movement, physical objects, and robotic actions.

This could become increasingly important as researchers attempt to build robots capable of operating in unpredictable real-world environments.

XDOF and UC Berkeley Are Building a Massive Robotics Dataset:

XDOF is also partnering with UC Berkeley's AI Research Lab on a project called ABC, which the company believes could become the largest collection of high-quality robot training data ever assembled.

The goal is ambitious: create a large-scale dataset that researchers and robotics companies can use to improve the intelligence and capabilities of physical machines.

The importance of projects like ABC goes beyond simply collecting more data.

For robotics companies, the quality, diversity, and variety of training data can directly influence what a robot is capable of learning.

A dataset containing thousands of examples of one task may not be enough.

General-purpose robots need exposure to a huge range of environments, objects, movements, people, and situations.

That means the robotics industry needs something approaching an internet-scale data collection operation for the physical world.

XDOF Plans to Build a Global Human Data-Collection Network:

To reach that scale, XDOF plans to recruit and train data collectors around the world. The company expects to work with different categories of human operators,

including:

• Teleoperators who remotely control robots.

• Egocentric operators who wear body sensors.

• Human data collectors recording everyday physical tasks.

• Workers who help generate and annotate robotics datasets.

This approach effectively creates a human-powered data network for robotics AI training. It also highlights something that is sometimes overlooked in discussions about artificial intelligence.

Even as AI becomes increasingly automated, humans are still essential for creating the datasets that allow these systems to learn.

XDOF Already Has 20 Customers:

The company's growth is particularly notable because it appears to be happening before the broader robotics market has fully matured.

That customer base provides an important clue about where the demand is coming from. Leading AI companies increasingly see robotics as one of the next major frontiers for artificial intelligence.

The technology needed to build general-purpose robots is improving rapidly, but the availability of high-quality physical-world training data remains a major limitation.

If XDOF can solve that problem at scale, its position in the robotics ecosystem could become extremely valuable.

The Robotics Data Race Is Just Getting Started:

XDOF isn't the only company chasing this opportunity.

Other startups, including Mecka AI, are working on collecting real-world data for robot training.

Meanwhile, established human-data companies such as Scale AI and Micro1 are also expanding beyond traditional large language model data into physical-world and robotics applications.

That competition could become intense.

The companies that control the infrastructure for collecting and labeling high-quality robotics data could become critical suppliers to the next generation of AI companies. It's a business model that looks surprisingly similar to the data-labeling industry that emerged during the generative AI boom.

But this time, the data isn't primarily text.

It's movement, manipulation, perception, and interaction with the physical world. Why Robotics Training Data Could Become the Next AI Gold Rush

The biggest opportunity for XDOF may be that robotics data is still in its early stages. Generative AI companies were able to build powerful language models partly because enormous quantities of digital information already existed.

Robotics has no equivalent.

There is no single "internet of physical actions" containing billions of examples of people opening doors, preparing food, moving furniture, folding laundry, using tools, or interacting with unfamiliar objects.

• Someone has to collect that information.

• Someone has to structure it.

• Someone has to label it.

And someone has to build the infrastructure that allows AI researchers to turn it into useful training datasets.

That is precisely the market XDOF is targeting.

What the $1.2 Billion Valuation Says About Physical AI:

If the reported Series B valuation of approximately $1.2 billion becomes official, it would represent a remarkable jump for a company founded only in 2024 and emerging from stealth just months ago.

But the valuation also tells us something larger about where venture capital is moving. Investors are increasingly looking beyond traditional generative AI applications and toward the infrastructure required for physical AI, humanoid robots, autonomous machines, and general-purpose robotics.

The AI industry spent the past several years building models that could understand language, images, and code.

The next phase may be about teaching machines how to understand and interact with the real world.

And that requires an entirely new type of data.

The Bigger Picture:

XDOF's story is ultimately not just about another AI startup raising money.

• It reflects a much larger shift happening across artificial intelligence.

• The industry is moving from digital intelligence to embodied intelligence.

• Large language models learned from the digital world.

• Robots have to learn from the physical one.

That difference creates an enormous data problem—and potentially an enormous business opportunity.

If XDOF succeeds in building a global infrastructure for collecting high-quality robotics training data, the company could become an important piece of the emerging physical AI ecosystem.

The reported $1.2 billion valuation may therefore be less about what XDOF has accomplished so far and more about what investors believe the robotics data market could become.

The next AI gold rush might not be about generating more text.

It might be about teaching machines how to touch, move, manipulate, navigate, and operate in the real world.

And XDOF wants to supply the data that makes that possible.

Key Takeaways:

• XDOF: Robotics data infrastructure startup founded by UC Berkeley researchers Philipp Wu and Fred Shentu.

• Reported Series B valuation: Approximately $1.2 billion.

• Potential Series B lead: 8VC.

• Previous funding: $70 million Series A.

• Reported annualized revenue: Approaching $50 million.

• Customers: Approximately 20, including frontier AI laboratories.

• Core technology: Robot teleoperation, human data collection, sensors, data pipelines, and robotics data annotation.

• Major project: ABC, a collaboration with UC Berkeley's AI Research Lab.

• Market opportunity: Robotics training data, embodied AI, physical AI, humanoid robots, and general-purpose robotics.