Physical AI
XDOF Raises $70M to Build the Data Layer Robots Still Lack
XDOF emerged from stealth with $70 million in funding and ABC-130K, a large manipulation dataset built with UC Berkeley AI Research lab. The launch points to a less glamorous but central robotics bottleneck: collecting, cleaning, and evaluating the physical interaction data needed to train useful robot models.
XDOF emerged from stealth on June 17 with $70 million in funding and a blunt thesis for physical AI: the next constraint is not a flashier humanoid body, it is the shortage of clean robot interaction data.
The startup is building data pipelines, collection hardware, teleoperation operations, and annotation systems for AI labs and robotics companies. Its first public release, ABC-130K, includes more than 130,000 manipulation trajectories, 300 hours of simulation, and 100 hours of evaluations.
That makes XDOF less cinematic than a walking robot demo, but probably more important. If humanoids are going to move from stage events to reliable factory work, the industry needs a data supply chain that can survive contact with real objects, real failure modes, and real labor costs.
Key Stats
$70M
Funding Raised
130K+
Robot Trajectories
20
Reported Customers
60
Employees
The News
XDOF, pronounced "ecks-doff," launched publicly with backing from Thrive Capital, Spark Capital, Andreessen Horowitz, Lux Capital, and WndrCo. The company was founded in October 2024 by Philipp Wu, Fred Shentu, and Nemo Jin, after Wu and Shentu worked on GELLO, a low-cost teleoperation system that lets a human operator drive a robotic arm to produce manipulation data.
The core pitch is simple. Language models scaled because the internet had massive text corpora. Image models scaled because the web had massive image corpora. Robots do not get the same free dataset. They need trajectories, force-sensitive manipulation records, timing, camera views, failure labels, recoveries, and action histories tied to the physical world.
XDOF says it is already working with 20 customers, including several frontier AI labs that it has not named. That detail matters. If accurate, the business is not just selling a dataset to robotics researchers. It is selling the operational layer that AI labs need if they want robot models but do not want to run warehouses full of calibrated hardware and trained teleoperators.
The company is also releasing ABC-130K with UC Berkeley's AI Research lab and collaborators. The dataset contains more than 130,000 trajectories across complex manipulation tasks, along with simulation and evaluation material. For an industry full of demos that are hard to compare, a larger shared training and evaluation corpus is useful because it gives researchers something concrete to test against.
Why this matters
Robot foundation models will not be trained by scraping the open web alone. They need expensive physical traces of bodies interacting with objects. XDOF is trying to turn that scarce, messy input into a repeatable infrastructure business.
Why Robot Data Is Different
The phrase robot training data sounds tidy until the work begins. A model cannot learn how to zip a bag, insert a cable, flatten a box, or align a part from a detached image alone. It needs the action sequence that produced the result, the sensory stream that guided the action, and the recovery path when the first attempt failed.
That is why YouTube-scale video is a weak substitute. It shows what humans or machines did, but often lacks exact actuator commands, calibrated depth, joint states, gripper force, synchronized camera geometry, and clean failure labels. A robot model trained only on passive video can learn visual priors, but it still has to translate observation into control.
Physical AI also has a data distribution problem. A humanoid in an auto plant, a bimanual tabletop rig in a lab, a warehouse arm, and a mobile manipulator all produce different data. Even when the task looks similar to a person, the embodiment changes the policy. Grip geometry, reach envelope, compliance, wrist torque, sensor placement, and latency all affect whether a learned action transfers.
XDOF is positioning itself at that uncomfortable intersection. It wants to collect data across robots, clean and annotate it, and build tools that make the output useful to model teams. That is harder than producing a slick benchmark video, but it is closer to the plumbing physical AI needs.
AI-generated image
Robot data operations depend on calibration, repeatability, and synchronized sensing. Source: Biped.News AI illustration.
The Three-Layer Data Stack
XDOF describes a data pyramid with three tiers. At the top is teleoperation data collected on the actual robot that will be deployed. This is the most valuable because it matches the target body, sensors, workcell, and task. It is also the most expensive, since each customer robot has to be maintained, calibrated, and operated.
The middle tier uses lower-cost teleoperation systems such as GELLO-like rigs to collect more general manipulation data. This can produce scale without tying every demonstration to a production robot. It also lets model builders test ideas before spending time on a more expensive embodiment.
The bottom tier is egocentric human data, gathered from people performing tasks while wearing sensors. That approach can scale faster, but it introduces another translation problem. Human hand motion, wrist compliance, tactile feedback, and object understanding are not the same as a robot's control space. The best use case may be task structure and perception, not direct imitation.
| Data Tier | What It Captures | Strength | Main Risk |
|---|---|---|---|
| Target-robot teleoperation | Actions on the deployed robot body | Best embodiment match | High cost and slow scaling |
| General teleoperation rigs | Manipulation demonstrations on lower-cost systems | Better volume and repeatability | Transfer gaps to production robots |
| Egocentric human sensing | First-person task demonstrations | Large task diversity | Hard action-space translation |
ABC-130K Is a Signal, Not the Whole Business
ABC-130K gives the market a public artifact to inspect. The reported scale, more than 130,000 trajectories and hundreds of hours of related simulation and evaluation material, is large by open robotics dataset standards. It also arrives at the right moment, as labs try to adapt foundation model habits to physical tasks.
Still, open data is not automatically a moat. The real commercial question is whether XDOF can produce new customer-specific data faster, cleaner, and cheaper than in-house robotics teams. For large AI labs, the calculation may be similar to data labeling in the early computer vision era. Owning the model is strategic. Owning every data collection operation may be a distraction unless the collection itself becomes a source of product advantage.
XDOF seems aware of that trap. Its public framing is not just "we sell data." It is building collection tools, annotation systems, feedback loops, and operations. That matters because raw robot demonstrations can be noisy. A training run needs metadata about the scene, object identity, operator behavior, success criteria, failure modes, recovery attempts, and sensor quality.
The company also has to solve workforce design. Its CEO has described the need for large facilities, many robots, maintenance teams, calibration, and trained operators. That sounds more like an industrial services company than a pure software startup. It may be exactly the kind of unglamorous business that makes glamorous humanoid demos possible.
Data
Trajectories, sensor streams, task labels, timing, and failures captured from physical interaction.
Operations
Robot maintenance, calibration, teleoperator training, workcell design, and quality control.
Models
Robot policies and foundation models that use the data to improve manipulation and recovery.
What This Means for Humanoid Robots
Humanoid companies often talk about fleet learning. The promise is that every robot improves the model for every other robot. In practice, fleet learning only works if the fleet produces data that can be trusted, searched, labeled, and converted into better behavior. That is a systems problem, not just a model-size problem.
For bipedal robots, the data challenge is even sharper. Locomotion and manipulation interact. A robot reaching into a bin, opening a door, or loading a cart is not just moving an arm. It is managing balance, foot placement, torso posture, collision risk, and recovery. A small force error at the wrist can become a whole-body correction.
That does not mean every useful training dataset has to come from a full humanoid. Many manipulation skills can start on cheaper rigs, then transfer through simulation, retargeting, or post-training. But the final mile will still need embodiment-specific data. The question is how much expensive target-robot data is needed after broad pre-training.
If XDOF can reduce that amount, it becomes strategically useful to humanoid builders. If it cannot, the company may still serve labs that need general manipulation data, but the direct path to production humanoid autonomy will stay expensive.
AI-generated image
Physical AI will require data operations as much as model research. Source: Biped.News AI illustration.
The Business Test
XDOF's funding round is a sign that investors see robot data as an infrastructure category. The tougher test is whether customers keep buying after the first dataset or pilot. Robotics data work is labor-heavy, hardware-heavy, and sensitive to customer secrecy. Frontier labs may want help, but they may also guard their robot programs closely.
There are three business signals to watch over the next year. First, whether XDOF names major customers or publishes repeat contracts. Second, whether ABC-130K produces measurable model improvements in outside labs. Third, whether the company can show that its data collection loop shortens deployment timelines for real robots.
The competitive pressure will be intense. Large robotics companies can build internal data operations. AI labs can acquire teams. Hardware startups can bundle data services around their own robots. Universities can publish more open datasets. The category is attractive precisely because everyone has the same problem.
XDOF's advantage, if it has one, is focus. Most model labs do not want their best researchers spending days debugging camera sync, swapping grippers, training teleoperators, or cleaning failed demonstration logs. If XDOF can own that machinery, it can become the quiet supplier behind a large part of the physical AI race.
FAQ
What did XDOF announce?
XDOF emerged from stealth with $70 million in funding and announced ABC-130K, a large open robotics manipulation dataset built with UC Berkeley AI Research lab and collaborators.
Why is robot training data hard to collect?
Robots need synchronized physical interaction records, including camera views, action histories, timing, task outcomes, failures, and embodiment-specific sensor data. Passive internet video usually lacks those signals.
Does this replace humanoid robot testing?
No. Broad manipulation data can help pre-train models, but production humanoids still need data from their own bodies, sensors, tasks, and work environments.
Is this a deployment story?
Not in the robot fleet sense. XDOF is selling data infrastructure and says it has customers, but the announcement is about funding, dataset release, and infrastructure, not robots deployed into production work.
The Bottom Line
XDOF is a reminder that physical AI may be bottlenecked by logistics before intelligence. The robotics market has no shortage of models, bodies, investor decks, or lab demos. What it lacks is high-quality, high-volume physical interaction data that can be used again and again.
That makes the company worth watching even if it never builds a humanoid. A robot that works in the real world is the visible product. The data operation behind it may be the more durable asset.