XDOF: Inside the Robotics Data Startup Building the Infrastructure Behind Physical AI

XDOF has gone from a Berkeley robotics project to a $70 million-funded startup reportedly discussing a $1.2 billion valuation in less than two years. Its real bet is bigger than teleoperation: that the physical-AI boom will require an entirely new industrial supply chain for generating, cleaning, evaluating and learning from robot experience.

Most artificial-intelligence companies building general-purpose robots eventually encounter the same inconvenient problem. A language model can learn from books, websites, code repositories and decades of digitised human communication. A robot that needs to fold a shirt, open a drawer or place an awkward object into a container cannot simply download millions of examples containing every force, movement, camera angle and physical consequence involved in completing the task.

Those examples have to be created.

That shortage has turned robot training data from an academic inconvenience into a commercial opportunity, and few young companies are pursuing it as aggressively as XDOF. Founded in 2024 by UC Berkeley researchers Philipp Wu and Yide “Fred” Shentu alongside Nemo Jin, the San Mateo company is attempting to build the collection systems, software, human operations and evaluation infrastructure required to produce physical-world training data at industrial scale. XDOF describes its ambition more broadly as building infrastructure for “physical intelligence”, supplying robotics companies and AI laboratories with production-scale datasets, robotic systems and the tooling required to train increasingly general machines.

The proposition has already attracted significant financial attention. XDOF emerged publicly on 17 June 2026 alongside a $70 million Series A, with Thrive Capital, Andreessen Horowitz, Spark Capital, Lux Capital and WndrCo among its named backers. On 4 September, TechCrunch reported that the company was already in late-stage discussions for another financing led by 8VC at a valuation of roughly $1.2 billion. The size of that proposed Series B, whether the valuation is pre- or post-money and even whether the transaction will ultimately close remain undisclosed. As of 8 September 2026, Nuvastra has found no public announcement from XDOF or 8VC confirming that it has completed.

That distinction matters. XDOF is not yet a confirmed $1.2 billion company simply because investors are reportedly discussing such a price. What is more interesting is why they appear willing to discuss it at all.

XDOF at a glance

Detail	Current position
Founded	October 2024
Headquarters	San Mateo, California
Founders	Philipp Wu, CEO; Yide “Fred” Shentu, CTO; Nemo Jin, COO
Core business	Robot-training data, teleoperation, collection infrastructure, annotation, evaluation and robotics tooling
Public launch	17 June 2026
Confirmed funding	$70 million Series A
Major named backers	Thrive Capital, Andreessen Horowitz, Spark Capital, Lux Capital and WndrCo
Latest financing	Reported Series B talks led by 8VC at about a $1.2 billion valuation; not publicly confirmed as closed as of 8 September 2026
Reported customers	Around 20, including unnamed frontier AI laboratories
Reported annualised revenue	Approaching $50 million, according to TechCrunch sources
Major open dataset	ABC-130K
ABC-130K scale	134,806 episodes across 195 tasks, totalling 3,553 hours of real bimanual manipulation
Current focus	Infrastructure for training general-purpose and embodied AI systems

The customer and revenue figures require particular qualification. XDOF has not published a customer list or audited revenue statement. The roughly 20 customers figure originated with CEO Philipp Wu in a June TechCrunch interview, while the approximately $50 million annualised revenue figure was subsequently reported by TechCrunch in September citing people familiar with the business. “Annualised revenue” should not automatically be read as contracted recurring revenue; XDOF has not publicly disclosed contract lengths, concentration, gross margins or how much of its income comes from recurring software versus labour- and equipment-intensive data projects.

Where XDOF came from: the GELLO project

While working at UC Berkeley, Wu and Shentu were investigating how robots could learn complex physical skills from human demonstrations. The problem was not merely designing a better learning algorithm. Researchers first needed enough demonstrations on which such algorithms could learn.

That led to GELLO — short for General, Low-Cost and Intuitive Teleoperation — developed by Wu, Shentu, Zhongke Yi, Xingyu Lin and Berkeley professor Pieter Abbeel. The project proposed a relatively simple way for a person to operate a robot arm: build a small physical controller with approximately the same kinematic structure as the robot being controlled. Moving the miniature leader arm produces commands for the full-sized machine, making the interface more physically intuitive than controlling the robot with a conventional joystick, VR controller or 3D mouse.

The original system was deliberately inexpensive. The researchers reported a bill of materials below $300 for each GELLO device and released hardware and software openly. Implementations were demonstrated on Franka, UR5 and xArm robots, including bimanual and contact-heavy tasks such as preparing toast, removing a tea bag, stacking cups and retrieving objects. A 12-person user study reported faster task completion and stronger reliability than the alternative low-cost interfaces tested. The work was published at IROS 2024 after first appearing as a 2023 preprint.

GELLO did not solve general robotics. What it did was make one expensive part of robot learning significantly easier: gathering demonstrations from humans.

That distinction is central to understanding XDOF. The company was not conceived after its founders spotted investor excitement surrounding humanoids. It grew out of a practical research bottleneck they had already encountered themselves.

What does XDOF actually sell?

At its simplest, XDOF is building a data supply chain for robotics.

A robot-learning laboratory may know which model architecture it wants to train yet still need thousands or millions of useful examples of machines interacting with different objects under different conditions. Collecting those examples involves considerably more than recording video. Robots have to be purchased and maintained, sensors calibrated, cameras positioned, teleoperation systems built, operators trained, tasks designed, failures detected, annotations added and the resulting data stored in a form that training systems can consume.

XDOF wants customers to outsource much of that operation.

The company’s public description spans production-scale datasets, robot hardware, collection systems, data infrastructure and policy training rather than simple annotation. Current job advertisements reinforce that positioning: XDOF describes infrastructure including an “exabyte-scale data warehouse”, multi-tenant data systems and large-scale ingestion pipelines, while its data-quality roles cover teleoperation, egocentric capture and multiple sensor modalities.

For readers wanting the wider machine-learning context, Nuvastra’s guide to how machine learning works explains the distinction between training from examples, reinforcement learning and post-deployment inference.

XDOF’s data pyramid

XDOF has described its collection strategy as a pyramid with several forms of data that differ in cost, fidelity and scalability.

At the top sits teleoperation data collected on the actual robot a customer wants to deploy. This offers the closest possible match between training experience and the physical machine that will ultimately perform the work, but it is expensive because every hour requires access to the relevant robot and associated infrastructure.

The next layer is broader teleoperated robot data, including the kind GELLO was designed to generate. Instead of producing demonstrations for one customer's exact machine and application, a collection platform can gather large quantities of generally useful manipulation experience across common robot configurations.

The widest potential layer is egocentric human data: recordings of people performing tasks themselves, captured through cameras or wearable sensors. This is much cheaper to scale because each collector does not require a complete robot, although it introduces an important technical problem known as the embodiment gap. A human hand and a robot gripper have different joints, strength, sensing and movement constraints, meaning human behaviour cannot simply be copied directly into every robot.

That mixture is significant. XDOF is not betting that one kind of dataset will be sufficient. It is trying to build an infrastructure business capable of choosing the right data source for different stages of robot training.

Why robot data is so different from internet data

The comparison with large language models is useful, but only up to a point.

The web provided language-model developers with an extraordinary accidental dataset: text written by billions of humans for reasons that had nothing to do with training AI. Robot interaction data has no comparable reservoir.

Google DeepMind’s Open X-Embodiment project illustrates both the potential and the scarcity. Working with more than 20 research institutions, the project aggregated over one million robot episodes spanning 22 embodiments and hundreds of skills. That was a major research resource, yet even a dataset of that scale remains tiny compared with the information available to language or vision models, and its data was produced by different robots operating under different experimental conditions.

Scale AI, one of XDOF’s most formidable potential competitors, makes the same argument. Its Physical AI business says robot manipulation data cannot simply be scraped from the internet and therefore has to be produced interaction by interaction. Scale now claims its network collects more than 1,000 hours of demonstration data per day and has expanded from autonomous-vehicle annotation into dedicated robotics collection infrastructure.

The bottleneck is consequently more complicated than “not enough video”. Useful robot data may combine camera imagery, joint positions, forces, actions, timestamps, task instructions, sensor readings and information about whether an attempt succeeded. Poor calibration can poison a dataset before model training even starts.

That helps explain why XDOF repeatedly emphasises hardware and operations alongside software. In physical AI, the data pipeline begins in the room where the robot is moving.

ABC-130K: XDOF’s attempt to prove it can operate at scale

XDOF’s most visible public demonstration of this infrastructure is ABC-130K, released alongside researchers from UC Berkeley, MIT, Carnegie Mellon and Amazon FAR as part of the wider ABC behaviour-cloning project.

ABC-130K contains 134,806 episodes covering 195 tasks and 3,553 hours of real-world bimanual manipulation. The broader ABC stack also includes approximately 400 hours of simulated teleoperation and more than 100 hours of real robot evaluations. Tasks range from folding and sorting to tool use, insertion, assembly and fine manipulation. The data and associated tooling have been released openly, with XDOF distributing ABC-130K under Apache 2.0.

The ABC project also publishes robot hardware, training infrastructure, simulation tools, model checkpoints and evaluation procedures. Its researchers report that simulation performance correlated strongly with corresponding real-world performance across the checkpoints tested, potentially allowing researchers to eliminate weaker training configurations before spending time evaluating them on physical robots.

This provides a useful clue about XDOF's intended business. The company does not appear to believe that raw demonstrations are enough. The commercially valuable layer may be everything required to determine which data is useful, how models respond to it and whether the resulting policies actually work on a machine.

More robot data can actually make a model worse

One of the more interesting findings emerging from XDOF-affiliated research cuts directly against the simplistic idea that robot intelligence can be improved merely by collecting ever larger datasets.

A teleoperator folding a shirt may hesitate, adjust a sleeve unnecessarily, re-grip an object or spend several seconds recovering from an awkward move. The final task can still succeed, meaning a crude dataset filter would classify the entire demonstration as good. A behaviour-cloning model, however, may learn the hesitation as readily as the successful fold.

Researchers working with XDOF described exactly this problem in WARP-RM. When increasingly inefficient demonstrations were added to one T-shirt-folding dataset, conventional behaviour cloning deteriorated from successful performance to only two successes in 20 trials. A technique designed to estimate which portions of a demonstration genuinely advanced the task maintained 19 successful trials out of 20 on the intermediate-quality dataset. The authors report throughput improvements of up to 18 times under the experimental setup.

These are research results from the authors involved, not independent replication across the wider robotics industry, so they should not be treated as proof that XDOF possesses a universal data-quality solution. They nonetheless illustrate an economically important point: a robot-data company that can identify which seconds of an expensive demonstration actually improve a model may be more valuable than one that simply produces more hours.

From human demonstrations to robots that practise by themselves

XDOF-affiliated research is also moving beyond data collection towards a more ambitious feedback loop.

The SARM and SARM2 work investigates reward models capable of estimating whether a robot is genuinely making progress during a long task. SARM2 uses common action primitives and a mixture-of-experts architecture to produce progress signals across multiple manipulation tasks. In experiments reported by its authors, the system reduced reward-estimation error substantially relative to tested baselines and supported a reinforcement-learning loop called SPIRAL, in which robots generated new attempts, evaluated those attempts and used the resulting signal to improve their policies.

The reported results included improvement from 58% to 100% success when folding shorts and from 50% to 90% when cleaning a whiteboard. Again, these are results from the research team rather than independent commercial benchmarking, and the system does not amount to a robot that can autonomously teach itself arbitrary physical tasks indefinitely.

But strategically, it shows where XDOF may want to move.

Human teleoperation is expensive. If a company can use human demonstrations to bootstrap a policy, then allow robots to produce and evaluate increasing amounts of their own experience, the economics of data production change dramatically. That places XDOF adjacent to the wider question of whether AI systems can improve through production experience, an area Nuvastra has examined in its coverage of continual-learning startups.

Who founded XDOF?

XDOF was created in October 2024 by Philipp Wu, Yide “Fred” Shentu and Nemo Jin.

Wu serves as chief executive and previously pursued a PhD at UC Berkeley under robotics researcher Pieter Abbeel, with work spanning reinforcement learning, robotics and methods for enabling more general real-world robots. Shentu, now XDOF’s CTO, collaborated with Wu on GELLO and other work connecting language models with robot control. Jin is the company’s chief operating officer.

XDOF says its wider team includes experience from Covariant, Tesla, Google, Uber and Applied Intuition. At its June public launch, Wu told TechCrunch that the business employed roughly 60 people. XDOF's current careers page shows hiring across robotics, perception, whole-body control, infrastructure, mechanical engineering, go-to-market and operations, including roles in San Mateo and Jakarta.

The Jakarta roles are a small but revealing detail. A company planning genuinely global robot-data operations needs not only researchers in California but facilities, managers and collection workforces capable of generating physical experience elsewhere.

How much money has XDOF raised?

The only completed financing that can currently be verified publicly is the $70 million Series A announced on 17 June 2026.

Cooley, the law firm that advised XDOF on the transaction, identifies the financing explicitly as a Series A. XDOF’s website names Thrive Capital, Andreessen Horowitz, Spark Capital, Lux Capital and WndrCo as backers. Wu's own launch announcement additionally thanked Olive Capital, Samsung Next, Scribble Ventures, Abstract and Emerson Collective as investors.

TechCrunch reported on 4 September that XDOF was in late-stage Series B discussions led by 8VC at an approximate $1.2 billion valuation. It reported that XDOF had not initially intended to return to the market so quickly but attracted investor interest as annualised revenue approached $50 million. Neither XDOF nor 8VC commented for that report, and TechCrunch said the terms were not final.

At those headline figures, a $1.2 billion valuation would equal roughly 24 times the reported annualised revenue rate. That is striking for a company founded less than two years ago, but the multiple is considerably less informative than it first appears because public information does not reveal the quality of that revenue. Data contracts can be large without being recurring, and a business employing operators, running warehouses and maintaining robot fleets may have very different margins from conventional software.

The valuation therefore tells us more about investor expectations for the robotics-data category than it does about XDOF's presently disclosed financial fundamentals.

Who are XDOF’s customers?

XDOF says it works with leading robotics laboratories and companies, while Wu told TechCrunch in June that the business had approximately 20 customers, including several frontier AI labs. None of those frontier customers has been publicly identified.

The anonymity is understandable given the competitive sensitivity of frontier robotics programmes, but it leaves investors and outside observers with an evidential gap. Public information does not show how much revenue comes from the largest customer, how many engagements have renewed or whether customers regard XDOF as a strategic long-term platform rather than a temporary route to acquiring datasets quickly.

Those questions become more important if the company is ultimately priced above $1 billion.

XDOF versus Scale AI, Mecka AI and the new physical-data market

XDOF does not have the market to itself.

Scale AI has expanded a decade of data operations into a dedicated Physical AI Data Engine. Its offering spans custom robot collection, annotation, bimanual manipulators, human egocentric collection and production integrations. In March 2026, Scale announced an integration with Universal Robots that allows its software to collect training information directly from production industrial hardware. Its size, customer relationships and existing global data workforce make it the obvious incumbent threat.

Mecka AI is pursuing a closely related infrastructure opportunity but has put greater public emphasis on collecting physical behaviour directly from humans using body sensors, phones and other capture systems. Fortune reported in June that Mecka had raised $60 million across two financings and was projecting a $100 million annual run rate based on signed contracts, although customers and valuation were not disclosed.

micro1, better known for supplying human expertise to AI laboratories, is moving into physical AI from another direction. The company has said it is creating robot pre-training data by having people record ordinary object interactions, giving it another potential route into the same expanding market.

XDOF's present differentiation is therefore unlikely to come from the observation that robots need data; the industry broadly agrees on that. Its stronger claim is full-stack robotics expertise, particularly around teleoperation, data quality, evaluation and policy training.

That is a harder proposition to copy than basic annotation, but it is not impossible to copy.

What could become XDOF’s moat?

The most obvious answer would be proprietary data, but that may not be the company's strongest advantage.

Datasets can lose exclusivity. Customers can collect their own. Hardware changes. Models become more data-efficient. Open research can spread successful collection techniques to competitors.

The more interesting potential moat is operational learning.

Running thousands of physical demonstrations exposes subtle problems that rarely appear in a software-only data pipeline: a joint with slightly different calibration, a camera mounted at the wrong angle, a particular object that consistently causes failures, operators whose technique differs, sensors that gradually drift or task specifications that inadvertently produce useless examples. A company that repeatedly connects those observations back to model performance can accumulate practical knowledge about how to design better data campaigns.

XDOF's research into reward modelling and data quality suggests that it understands this. The goal is not simply to own the video. It is to know why one demonstration teaches a robot more than another.

Open-sourcing ABC-130K can even support that strategy rather than undermine it. Giving researchers access to useful public infrastructure can help establish XDOF-designed systems as common tools, attract researchers and demonstrate technical credibility while the company's highest-value proprietary work remains in customer-specific collection and evaluation.

The risks behind the XDOF story

The most important risk is that XDOF is building infrastructure its biggest customers eventually decide they need to own.

Frontier AI laboratories have enormous capital budgets and powerful incentives to control strategically important training pipelines. If robot data becomes as important to a foundation model as compute, the largest laboratories may bring substantial portions of collection in-house.

A second risk is competition. Scale already operates global data infrastructure and has explicitly targeted physical AI. Mecka is aggressively pursuing human-derived data. Robotics model developers can build internal fleets, while hardware companies can accumulate training experience from deployed machines.

A third is economics. Robot data is physically expensive. Facilities, machines, repairs, calibration and human operators create costs that do not disappear simply because the resulting dataset is digital. XDOF must demonstrate that its software, research and operational efficiency can produce attractive margins rather than leaving it trapped in a high-revenue, high-cost services business.

There is also a technical uncertainty that applies to the entire field: no one yet knows exactly how much real-world data general-purpose robotics ultimately requires, what combination of robot, human and simulated data will dominate, or whether future models will learn physical skills with dramatically greater data efficiency.

The winning data company may therefore need to keep changing what it sells.

Why XDOF matters even if it does not become the “Scale AI of robotics”

Comparing XDOF with Scale AI is tempting because both businesses sit between model builders and the expensive process of creating useful training material. But the analogy can obscure what makes physical AI different.

Language and image data existed before the AI companies arrived. Physical interaction data often has to be manufactured expressly for the model.

That makes the emerging robotics-data industry look less like conventional web annotation and more like a new form of industrial infrastructure. Warehouses filled with robots and operators may become to embodied AI what clusters of GPUs became to generative AI: expensive machinery sitting behind the consumer-facing intelligence.

XDOF is early enough that its ultimate position in that infrastructure stack remains unclear. Its named customers are hidden. Its margins are unknown. Its latest reported valuation has not been confirmed through a closed financing. Several of its most interesting technical results come from research involving its own founders and researchers rather than independent replication.

Yet the company's rapid progress is a meaningful signal.

A startup founded in October 2024 has already raised $70 million, assembled a sizeable robotics team, released one of the largest open bimanual manipulation datasets, attracted roughly 20 customers according to its CEO and, if current reporting is accurate, built a business approaching a $50 million annualised revenue rate.

The important question is no longer whether training data will matter to robotics.

It is who will build the factories that produce it.

XDOF is betting that those factories can become a company.

Frequently asked questions

What is XDOF?

XDOF is a San Mateo robotics infrastructure startup that produces data, teleoperation systems, collection infrastructure, evaluation tooling and related services for companies training physical-AI and general-purpose robot models. It was founded in 2024 and emerged publicly in June 2026.

How do you pronounce XDOF?

XDOF is pronounced roughly “ex-doff”. The name refers to “degrees of freedom”, a robotics term describing the independent ways in which a mechanism can move. CEO Philipp Wu has described the X as representing arbitrary or unlimited degrees of freedom.

Who founded XDOF?

XDOF was founded by Philipp Wu, Yide “Fred” Shentu and Nemo Jin. Wu is CEO, Shentu is CTO and Jin is COO. Wu and Shentu previously conducted robotics research at UC Berkeley and co-authored the GELLO teleoperation work.

How much funding has XDOF raised?

XDOF announced a $70 million Series A on 17 June 2026. Its named backers include Thrive Capital, Andreessen Horowitz, Spark Capital, Lux Capital and WndrCo. A further Series B is reportedly under discussion but had not been publicly confirmed as completed as of 8 September 2026.

Is XDOF worth $1.2 billion?

Not as a confirmed completed valuation at the time of publication. TechCrunch reported that XDOF was in late-stage Series B discussions at approximately $1.2 billion, with 8VC expected to lead. The terms were described as unfinished and neither XDOF nor 8VC confirmed the transaction for that report.

What is ABC-130K?

ABC-130K is an open bimanual robot-manipulation dataset associated with XDOF and researchers from institutions including UC Berkeley, MIT, Carnegie Mellon and Amazon FAR. It contains 134,806 episodes across 195 tasks and approximately 3,553 hours of real-world interaction data.

Does XDOF build robots?

XDOF develops robotic systems and collection hardware as part of its infrastructure, but it is not currently positioning itself primarily as a branded general-purpose humanoid-robot manufacturer. Its business centres on the infrastructure required by companies building and training physical-AI systems.

Who competes with XDOF?

Its closest competitors include Scale AI's Physical AI operation and Mecka AI, while companies such as micro1 are also expanding into physical training data. XDOF additionally competes indirectly with robotics laboratories that build their own internal collection infrastructure.

Why does physical AI need so much data?

Robots have to learn how actions interact with real objects, geometry, friction, force and changing environments. Unlike language and images, much of this data does not already exist online in a form robots can use, so demonstrations frequently have to be generated specifically for training.

Who are XDOF’s customers?

XDOF has not published a customer list. CEO Philipp Wu told TechCrunch in June 2026 that the company had approximately 20 customers, including several frontier AI laboratories. Their identities remain undisclosed.