XDOF supplies robotics developers with recorded demonstrations, data collection equipment, and evaluation services for training AI to control physical machines. Its customers include robotics companies and frontier AI laboratories; in June 2026, the company reported working with 20 customers, without naming them.

A developer training a robot to manipulate objects needs more than video of a person doing the task. It needs observations paired with the robot's actions, so a model can learn which movements produce the desired result. XDOF's teleoperation systems let a human guide robot arms while the equipment records demonstrations, and its data tools organize and annotate those recordings for training.

For illustration, a hypothetical customer teaching a robot to pack objects could commission demonstrations of picking up different items and placing them in a box. Human operators would perform the task through the robot, generating examples that the customer could use to train its control model. Changing the objects, camera views, or robot hardware changes what must be collected; a generic video library is not automatically a substitute for demonstrations on the relevant machine.

XDOF's public data portal separates browsable samples from complete, versioned datasets provisioned to an organization. Its ABC-130K release, developed with academic collaborators, contains more than 130,000 episodes across 195 two-arm manipulation tasks, including folding, tool use, and assembly, on accessible YAM robot arms. The dataset is available under Apache 2.0, making this release an open resource rather than an exclusive commercial catalog.

XDOF also runs evaluations to show whether a trained model can perform a chosen task. For the ABC research project, it worked with researchers to define scoring rubrics and test model checkpoints in simulation and on real robots, letting them compare training choices against physical outcomes.

The public portal accepts evaluation requests for robot-control, vision-language, and language models against selected benchmarks. This extends the deliverable beyond training files: customers can use an external operator to run tests and return results instead of maintaining all the equipment and evaluation staff themselves.

XDOF sells access to an operation that produces usable robotics data and tests models. Robotics developers pay for work they would otherwise have to build and staff themselves: procuring and maintaining robots, training operators, recording demonstrations, checking quality, and preparing the resulting data. Its commercial contact process supports bespoke engagements; public materials do not establish a standard subscription price or a uniform per-hour fee.

Its delivery costs include human operators, robot equipment and repairs, facilities, cameras, storage, and engineering work to make the recordings usable. The economic opportunity is to collect more accepted demonstrations with the same equipment and staff, and reuse tooling across engagements. Extra recordings only improve the economics if customers can use them; failed demonstrations or data requiring extensive rework still consume labor and machine time.

Model evaluation adds another potential paid service built on the same physical infrastructure. Test stations and evaluation staff can support repeated comparisons as customers release new models, although robot setup and scoring must remain consistent for the results to be useful. XDOF's open ABC work provides evidence of this capability, without establishing the pricing or margins of its commercial contracts.

The free ABC-130K release gives researchers a way to try XDOF's data and training setup before commissioning additional collection or evaluation. It could lower the cost of acquiring customers, but the open license also means the company cannot rely on exclusive ownership of that public dataset as its competitive advantage.

XDOF competes with specialist robotics data suppliers such as Mecka and with customers building their own collection facilities. The central contest is who can turn physical demonstrations into data that improves a customer's robot, at an acceptable cost and with less operational work for the customer.

Mecka combines human-motion capture hardware, reconstruction models, and quality systems to supply training data for robotics. Its October 2026 financing announcement also describes commercial robot integration: installing hardware, capturing data on site, adapting models, and operating deployments for enterprises.

XDOF's GELLO roots and ABC release demonstrate a particular strength in collecting actions directly on robot arms and testing trained models. Mecka's human-motion approach can cover environments and activities without putting a robot at every collection site, but human recordings must still be translated into information useful for controlling a particular robot. Mecka's enterprise deployments could also give it access to operational data that a supplier serving research laboratories would have to collect separately.

An AI laboratory can own its robots, employ teleoperators, and keep its training data private. That approach provides direct control over tasks, hardware, and sensitive experiments, while requiring the lab to manage calibration, maintenance, facilities, and staffing alongside its model research.

XDOF offers to take on that operational burden and share the cost of infrastructure across customers. Open resources such as DROID, Open X-Embodiment, and XDOF's own ABC release give internal teams a starting point, however, reducing the value of selling undifferentiated recordings. XDOF must earn additional business through task coverage, reliable collection, useful annotations, and evaluation rather than the simple existence of a dataset.

XDOF's core buyer is a team developing robot intelligence, with spending concentrated in research and training data. Its hardware, collection operations, and testing tools could support other buyers and budgets, but entering those markets would require additional products or delivery responsibilities rather than merely collecting more demonstrations for existing customers.

An industrial business buying a working robot system has a different budget and objective from a laboratory buying training data: it wants a machine that completes a production task reliably. XDOF's ability to collect demonstrations and evaluate robot behavior could support a move into adapting robots for individual customer sites, capturing local examples, and testing performance before deployment.

Mecka's announced integration business demonstrates an adjacent commercial route, rather than an announced XDOF offering. To pursue it, XDOF would need installation, safety, service, and integration capabilities beyond delivering datasets. The potential return would come from implementation and ongoing operating contracts, with substantially greater responsibility for whether the customer's physical operation works.

ABC opens another possible route into universities and smaller robotics teams that cannot afford a large collection facility. Its open hardware setup, training infrastructure, simulation pipeline, and accessible robot arms reduce the prerequisites for experimentation, while XDOF's portal supplies a place to inspect data and request evaluations.

XDOF could package managed infrastructure, equipment support, or hosted testing for these buyers, drawing on software and hardware already used in its research collaborations. That would address laboratory equipment and research software budgets alongside large AI-lab data contracts. The challenge would be charging for reliable operation and support around an open stack, since the public dataset and code are not themselves exclusive paid products; no such packaged commercial offering is established by the available evidence.

Collection costs and rework: XDOF's product requires physical equipment, maintenance, and trained operators, so scaling data production also creates operating costs. Recordings that do not improve customer models can require expensive recollection or annotation work, limiting the benefit of higher collection volumes.

Customer insourcing: XDOF serves robotics developers capable of building their own collection systems, while openly distributing part of its own data and tooling. Large customers could internalize valuable tasks and leave suppliers competing for less differentiated collection work.

Transfer to new robots: ABC demonstrates extensive manipulation data on a particular two-arm setup, rather than universal performance across every robot and environment. If customers require substantially different hardware or tasks, XDOF may need new equipment and collection campaigns before existing data and processes become useful.