Why Real-World Robot Data Is Suddenly Worth a Billion…

Language models could learn from billions of pages on the internet. Robots have no equivalent archive of hands folding laundry, fitting a cap onto a pen or placing an object without crushing it. That missing physical experience is becoming one of the most valuable bottlenecks in artificial intelligence.

XDOF, a startup founded by researchers with roots at UC Berkeley, is reportedly in late-stage talks for a Series B round at a valuation of about $1.2 billion. The talks come less than three months after the company emerged from stealth with a $70 million Series A. The round is not complete, and neither XDOF nor the reported lead investor 8VC confirmed the terms to TechCrunch. Yet the speed of the discussions is itself a signal: investors are looking beyond humanoid bodies toward the data infrastructure that teaches machines how to move.

What you need to know

XDOF collects real-world robot demonstrations through teleoperation and human-worn sensors.
TechCrunch reports that the company is discussing a Series B at a valuation of roughly $1.2 billion.
The startup reportedly works with about 20 customers and is approaching $50 million in annualized revenue.
Its open ABC-130K dataset contains more than 130,000 episodes across 195 two-arm manipulation tasks.
The financing terms remain unconfirmed and could still change.

Why robots cannot simply learn from the internet

Text and image models benefited from an enormous digital record created by people. A robot needs a different kind of example. It must know not only what a folded shirt looks like, but how forces, joint angles, camera views and timing change throughout the action.

A video can show the result while hiding essential information. The robot may need synchronized images from several cameras, precise motor commands, gripper forces and an indication of whether the attempt succeeded. Collecting that package requires hardware, trained operators, calibration and consistent quality control.

The problem grows when a model is expected to work on different machines. A movement recorded on one pair of arms may not transfer cleanly to another robot with different dimensions, joints or grippers. Data volume therefore matters, but diversity and documentation can matter just as much.

Inside a robot data factory

XDOF grew from GELLO, a comparatively low-cost system that lets a person operate robotic arms remotely. A human performs a task while the machine records the trajectory. Repeating the task under different conditions creates demonstrations from which a model can learn.

The company also plans to use people wearing sensors to capture everyday activity. That could expand collection beyond a fixed robot station, but human motion is not automatically a valid robot command. Software still has to translate hands, arms and body movement into actions a particular machine can execute safely.

ABC-130K illustrates the intended scale. XDOF says the open dataset includes more than 130,000 episodes covering 195 bimanual tasks, including folding, tool use, insertion, assembly and tying. Open publication allows researchers to compare approaches, although a large dataset alone does not guarantee a useful policy.

More demonstrations can sometimes make a robot worse

Robot data contains dead time, hesitation and failed approaches. If a teleoperator pauses before a grasp or makes an unnecessary correction, the model may learn that behavior along with the successful movement. Adding more low-quality examples can amplify the noise instead of improving the skill.

The commercial opportunity is therefore broader than recording. Customers need filtering, annotation, evaluation and repeatable tests showing that a new model version is actually better. A supplier that controls this pipeline could become the robotics equivalent of a data-labeling platform, but with far more hardware and operational complexity.

What a billion-dollar valuation would need to prove

The reported valuation assumes that demand for robot experience will grow rapidly. To justify it, XDOF would need to show that its data improves measurable outcomes: task completion, transfer to new objects, fewer human interventions and safe recovery after mistakes.

There are also governance questions. Customers will want to know who owns demonstrations collected in their facilities, whether videos expose products or employees, and how data from competing companies is separated. A global workforce of teleoperators and sensor-equipped collectors adds questions about training, working conditions and consistency.

The financing report should not be mistaken for a completed deal or proof that general-purpose robots are close. It shows something narrower but important: the market is assigning extraordinary value to the unglamorous work of producing reliable physical experience.

From demonstrations to a usable robot policy

A recorded movement is only the beginning of the learning pipeline. Before a robot can use it, the demonstration has to be synchronized across cameras, joint sensors and grippers, checked for missing frames and connected to a clear description of the task. Engineers then train a policy, test it on situations that were not included in training and inspect the failures. A system that succeeds with one cup in one position may still fail when the lighting changes, the object is rotated or the table is slightly higher.

This is why evaluation can be as valuable as collection. The ABC project does not present data as an isolated archive: XDOF says it also supported real and simulated evaluations at different checkpoints, architectures and hyperparameter settings. That makes progress measurable. For customers, the decisive question is not how many hours were recorded, but whether a new dataset produces a repeatable improvement on tasks that matter.

The transfer problem behind the data boom

Robotics companies would prefer to train one broadly capable model and deploy it across many machines. Reality is less convenient. Camera placement, arm length, joint limits, gripper geometry and control frequency all shape the recorded action. Even the same task can require a different trajectory on another robot. Datasets therefore need enough technical context to show what was measured and on which embodiment.

That creates two possible markets. One is large, relatively standardized data collections that help researchers build general representations. The other is smaller, customer-specific datasets gathered on the exact hardware and workflow used in a factory, laboratory or logistics site. XDOF could participate in both, but the second model demands more field operations and tighter handling of confidential information.

Where the competitive moat could emerge

Robot data is not protected merely because it is expensive to collect. A durable advantage would have to come from a complete operating system for collection: reliable teleoperation hardware, trained operators, automated quality checks, task taxonomies, versioned datasets and evaluation services. The faster this loop identifies weak skills and produces targeted new examples, the more useful it becomes.

Open releases such as ABC-130K do not necessarily weaken that business. They can establish formats, attract researchers and make it easier to compare models. The commercial layer can then focus on scale, customization, privacy and guaranteed quality. The risk is that robot manufacturers build these capabilities internally or converge on synthetic data that reduces the need for expensive physical demonstrations.

Bottom line

The next robotics winner may not manufacture the most human-looking machine. It may build the system that records, cleans and evaluates millions of ordinary actions.

If physical AI is to progress from demonstrations to useful work, robot builders need an experience layer that the open internet cannot provide. XDOF’s rapid rise suggests that this layer is becoming an industry of its own.