Two weeks ago, OpenAI announced it would relaunch its robotics program that was shuttered in 2021—a pivotal signal that physical AI is entering a new era. The company's return to robotics highlights a critical truth: while large language models have made rapid progress, physical AI faces a different and more complex challenge: data infrastructure.
That gap is creating a new kind of infrastructure business. Unlike LLMs that were trained on a vast sea of publicly available text, robots need data that captures physical interaction, and that kind of data barely exists. YouTube videos and footage captured by gig workers are low-fidelity and hard to reconcile with the physical world.
The absence of quality robotics data has become a bottleneck for frontier AI labs trying to build physical intelligence. As Philipp Wu, co-founder and CEO of XDOF, puts it: "All of the top labs are trying to pursue robotics... you don't want to be in this type of situation where you pursue this technology too late."
The Data Gap Between Digital and Physical AI
The rise of large language models has created an impressive benchmark for digital AI capabilities. Developers now have access to billions of parameters and vast amounts of training data that can be downloaded and processed on standard hardware. Robotics, however, operates in a fundamentally different domain.
Robots require training data that captures the nuances of physical interaction—the weight of objects, the resistance of materials, the unpredictable nature of real-world environments. This data is dirty, unglamorous work that requires hands-on collection rather than web scraping.
XDOF, emerging from stealth today, is betting that the next great bottleneck in AI isn't models or chips, but the data feedback loop needed to teach robots how to interact with the physical world. The startup aims to build the data pipelines, collection tools, and annotation systems that frontier labs and robotics companies can't easily build themselves.
XDOF has raised $70 million from Thrive Capital, Spark Capital, a16z, Lux, and WndrCo to create the data infrastructure for physical AI. The company has about 60 employees and is already working with 20 customers, including several frontier AI labs.
XDOF: Building the Robotics Data Ecosystem
XDOF has raised $70 million from Thrive Capital, Spark Capital, a16z, Lux, and WndrCo to create the data infrastructure for physical AI. The company has about 60 employees and is already working with 20 customers, including several frontier AI labs, though Wu could not name them.
The company's approach involves three tiers of a data pyramid. The most valuable tier is teleoperation data collected on the actual robot being deployed; next comes teleoperated robots gathering more general data, as with their GELLO project; and finally "egocentric" data gathered by humans performing everyday tasks, for which XDOF plans to build its own wearable sensors.
"Your camera choice is going to affect the quality of your data—which is going to affect how your hand-tracking algorithm performs," Wu said. "If you don't design the hardware well from the start, the data you collect might have very specific problems that you didn't anticipate."
From Academic Research to Commercial Infrastructure
The story of XDOF begins in the academic world. Philipp Wu ran into this data problem himself as a PhD student at UC Berkeley, where his focus was on enabling robots to learn skills from large-scale datasets. There was just one problem: they didn't have large-scale data to work with.
"We didn't have large-scale data to work with," he told TechCrunch. "There was this chicken-and-egg problem—we first needed to actually collect data before we could even ask how to train a foundation model for robotics."
Wu and his future XDOF co-founder and CTO, Fred Shentu, worked on a project called GELLO, a low-cost teleoperation system that lets a human operator control a robotic arm to generate training data. "It ended up becoming a very influential paper in robotics, because a lot of people had similar needs and bottlenecks, and many started leveraging this type of device for data collection," Wu said.
Spotting the opportunity, Wu, Shentu, and third co-founder and Chief Operating Officer Nemo Jin launched XDOF in October 2024 to provide a data ecosystem for companies pursuing robotics models. Mindful that data provision alone can be a dead-end business, the company is also focused on data cleaning, tooling, and annotation—creating a self-reinforcing feedback loop for robot trainers.
ABC Dataset: A Milestone for Academic Robotics Research
As a starting point, XDOF is partnering with UC Berkeley's AI Research lab to release what it believes is the largest collection of high-quality robot training data ever assembled, dubbed ABC. It includes 130,000 trajectories of robot manipulation data, 300 hours of simulation, and 100 hours of evaluations.
That kind of scaled-up pre-training data has never been available to academia before. "We've seen in language, image generation, and other fields, that when models and data are released, the community achieves things that you wouldn't necessarily have expected," said David McAllister, a Berkeley PhD student who helped organize the release.
The team has already used the data to train robots on benchmark tasks like folding T-shirts and flattening boxes, or loading AirPods into their cases. These seemingly simple tasks represent significant progress in robotic manipulation—tasks that require understanding of physics, object properties, and precise motor control.
Why Labs Aren't Building This Themselves
The company plans to hire and train armies of teleoperators and egocentric data operators around the world—a labor-intensive model that raises an obvious question: Why aren't the major labs doing this data production work themselves?
The answer lies in specialization. Just as LLM developers rely on specialized tooling and infrastructure providers, robotics companies are finding that data collection is a complex operation requiring expertise in hardware, software, and operations. Building an internal data production capability requires significant investment in both infrastructure and personnel.
"Why aren't the major labs doing this data production work themselves?" Wu acknowledges the question. "The answer is that it's not their core competency. They're experts in model architecture and algorithm development, not in operating large-scale data collection facilities."
The Future of Physical AI Is Data-Driven
OpenAI's return to robotics signals a turning point. The company's focus on physical AI comes after years of dominance in the language model space, and it underscores the belief that multimodal intelligence—combining language understanding with physical interaction—is the next frontier.
For now, the data infrastructure for robotics remains nascent. But companies like XDOF are building the foundational layers that will enable the next generation of physical AI systems. Whether it's household robots, manufacturing automation, or personal assistants that can interact with the physical world, the race is no longer just about algorithms—it's about data.
The next few years will likely see a similar pattern to what we witnessed with LLMs: early dominance by companies that can secure and process the most valuable data, combined with sophisticated model architecture. In robotics, however, that first-mover advantage may depend less on compute power and more on access to high-quality, diverse training data.
As Wu puts it: "Physical AI is the next frontier. The question isn't whether robots will become more capable, but how quickly we can build the data infrastructure needed to get there."
