XDOF targets $1.2B: the rush for robotics data

XDOF and the new premium placed on physical-world data

XDOF is reportedly targeting a valuation of around $1.2 billion as part of a Series B round still under discussion. The information, revealed by TechCrunch, comes just three months after this young company specializing in robotics data emerged from stealth. The U.S. outlet sums up the situation in its headline: “XDOF, just 3 months out of stealth, is in talks for a Series B at a $1.2B valuation.” These are therefore reported discussions, rather than a deal announced by the company as finalized.

The figure is nevertheless significant. A $1.2 billion valuation would place XDOF in the category of companies valued at more than $1 billion, even as its name has only just publicly emerged. Above all, it would illustrate a shift in part of the financial attention around artificial intelligence: after chips, data centers and large language models, investors are looking more closely at the resources needed for AI capable of acting in the real world.

The issue is not simply that of conventional industrial robotics. So-called “general” robotics models, capable of receiving an instruction in natural language, recognizing a scene, planning an action and controlling an arm or a mobile machine, require data that are very different from those used to train conversational systems. A text model can learn from documents, books, web pages or code. A robot must also learn what it means to grasp an object, apply a measured force, avoid a collision, open a drawer, reposition its gripper after slipping or adapt to a cluttered scene.

These interactions are difficult to turn into datasets. They cannot be reduced to images: a robotic demonstration can combine video, depth measurements, joint positions, commands sent to motors, force signals, the movements of a human hand, a linguistic description of the task and a final result. The quality of execution matters as much as volume. A series of failed or poorly annotated movements may be less useful than a more limited set of varied, reproducible and properly synchronized manipulations.

In this context, the promise of a robotics data provider takes on a particular dimension. A company that manages to organize the collection of physical interactions, standardize them and make them usable for training can position itself upstream of several markets: laboratories building models, robot manufacturers seeking to deploy them, and client companies looking to automate concrete operations.

The information relayed by TechCrunch does not detail, in the available brief, the investors involved, the amount sought, the precise terms of the deal or the technical characteristics of XDOF’s offering. These elements should therefore not be extrapolated. But the mere prospect of a Series B at this valuation level provides an indication of investors’ thinking: data from the physical world are seen as a strategic asset, potentially as scarce as computing resources or teams capable of training large models.

The very name XDOF refers to a concept familiar in robotics: degrees of freedom, often referred to by the English acronym DOF, for degrees of freedom. In a mechanical system, the expression describes the independent movements an object, arm or mechanism can perform. Without presuming the exact meaning the company gives its brand, this reference naturally belongs in a world where AI no longer processes only symbolic information: it must contend with movements, geometric constraints and objects subject to gravity.

The timing also reveals a change in industry vocabulary. For several years, the debate around generative AI has been largely dominated by model size, parameter counts, benchmark performance and demand for computing accelerators. In robotics, the question quickly returns to a more material reality: what data make it possible to train a system to perform a task outside a demonstration environment? The valuation envisioned for XDOF, if confirmed, would provide a financial answer to that question.

Why robotics data are harder to produce than text data

The apparent abundance of digital data has sometimes created an illusion: since there are billions of images and texts online, AI would naturally have everything it needs to learn to act. Yet observing the world and manipulating it are two distinct problems. An image shows that a cup is sitting on a table; it does not necessarily say how much pressure to apply to pick it up, how to compensate for its weight, which trajectory avoids other objects, or what to do if the cup is wet or its handle faces the wall.

Robotics has always suffered from this gap between perception and action. Traditional systems are often highly effective when configured for a precise task in a controlled environment. An industrial cell can repeat an operation at high speed when parts arrive in the same place and in the same orientation. But automation becomes much more complex when objects vary, the scene is cluttered, materials are fragile or the workspace changes regularly.

Learning-based approaches seek precisely to broaden this capacity for adaptation. They need examples showing how an action is performed in different circumstances. For a task as seemingly simple as “put an object in a bin,” data should ideally cover objects of different sizes, shapes and surfaces; varied positions; several speeds; grasping errors; lighting variations; and the adjustments required during movement. A robot cannot entirely infer these cases from textual descriptions.

This scarcity is first due to the cost of collection. Robots, sensors, operators, equipped spaces and control procedures are needed. Unlike a corpus of texts, an hour of manipulation data generally means that a physical device is actually operating, sensors are calibrated and sequences can be retrieved. The data must also be cleaned: a sensor may lose synchronization, a movement may be interrupted, a gripper may miss its target, or a camera may be obscured.

Diversity is a second challenge. Thousands of repetitions of the same gesture in the same laboratory kitchen are not equivalent to thousands of new situations. For a model to be useful, the system must be exposed to differences in configurations, materials, sizes and scenarios. This requirement makes data production longer and more costly, because it does not merely involve accumulating hours of video. It requires choosing what robots are made to do, under what conditions and with which measurement instruments.

Data can also be collected in several ways. One method relies on teleoperation, in which a person remotely controls a robot and thereby provides demonstrations. Another uses motion-capture systems or devices worn by operators to record human gestures. Robots can also produce traces through successive trials, in real or simulated environments. Each of these approaches has its advantages and limitations: fidelity to human movement, cost, collection speed, ability to measure forces, or differences between the human body and the final machine.

The issue of hardware compatibility is central. Data generated with a given robotic arm do not automatically transfer to another device. Dimensions, grippers, number of joints, sensors and speed limits may differ. A dataset can therefore be abundant while remaining closely tied to a particular platform. For players seeking to build models usable across several robots, the challenge is to represent actions and observations in a sufficiently general way without losing the information needed for precise control of a real machine.

This is where data become a bottleneck. Computing capabilities can be bought or rented, even if they remain costly. Model architectures circulate quickly through scientific publications and open-source communities. By contrast, a collection of high-quality physical interactions that is hard to reproduce and associated with operational know-how can provide a more durable advantage. It results from infrastructure, collection processes and selection work that cannot be duplicated instantly.

The parallel with computer vision is useful, provided it is not pushed too far. The ImageNet dataset, launched in the late 2000s, played a major role in accelerating image recognition by providing a vast set of annotated images. Robotics is now seeking its own foundational resources, but the equivalent is harder to build. Recognizing a category in an image is a different task from continuously controlling a machine in three-dimensional space, with immediate physical consequences in the event of an error.

The notion of “robotics data” therefore encompasses several assets: raw recordings, annotations, metadata on the robots used, task descriptions, systems for synchronizing streams, evaluation tools and usage rights. A database that does not clearly specify the conditions under which data can be used for commercial training is worth less than a corpus with defined governance. The debate over data provenance, already major for language and image models, also arises for physical-world data.

For a company such as XDOF, if the ambition reported by TechCrunch comes to fruition, the question will therefore not only be whether it can produce a great deal of data. It will have to prove their relevance, diversity, technical quality and usability for different architectures and platforms. In a sector where spectacular demonstrations abound but reliable deployment remains difficult, the economic value of data will depend on their ability to narrow the gap between a laboratory prototype and a repeatable operation.

From laboratories to manufacturers, a value chain being reshaped

The race for physical data does not start from zero. Major technology companies and robotics research teams have long worked on learning by demonstration, reinforcement learning, simulation and control-policy transfer. Google DeepMind notably presented RT-1 in 2022, a robotics-focused model trained on data from real robots. In 2023, the team presented RT-2, which sought to connect the capabilities of vision-language models with robotic action.

This work helped popularize the idea of models capable of generalizing across instructions and tasks rather than being programmed separately for every movement. It also made a fundamental limitation visible: improvements in models cannot be separated from access to physical demonstrations. Laboratories with robot fleets and many years of collection have an advantage that is difficult to measure from publications alone.

Academic research has also tried to expand the resources available. The Open X-Embodiment project, presented in 2023, brought together 22 robotics datasets from several institutions. The initiative was specifically intended to explore training models on heterogeneous data, collected with different robots and for different tasks. Its existence highlights both the desire to pool efforts and the persistent fragmentation of the field: data are scattered, formats differ and robots do not all have the same capabilities.

In the private sector, the movement extends beyond research laboratories alone. Humanoid robot makers, collaborative arm designers, warehouse specialists and companies developing action models all need data matching their uses. Figure, a U.S. company specializing in humanoid robots, announced in February 2024 a $675 million funding round involving Microsoft, OpenAI, Nvidia and Jeff Bezos through his investment structure, among others. The deal testified to large groups’ interest in the interface between AI and physical robotics.

Nvidia, for its part, has multiplied announcements around robotics and simulation. In March 2024, the company presented Project GR00T, an initiative intended to develop foundation models for humanoid robots. Its significance does not lie only in a model or platform: Nvidia is also present in chips, simulation tools and software infrastructure. This position enables it to connect several stages of the value chain, from model training to their execution on machines.

Data companies such as XDOF would fit into another part of this chain. They are not necessarily the makers of the final robot or the publishers of the most visible model. But they can become critical partners if they hold corpora that are difficult to recreate. An imperfect but illuminating analogy can be drawn with companies that provide annotated data for computer vision: they enabled many players to develop perception systems faster. In robotics, the challenge is broader because data also include actions and physical consequences.

This situation can produce several business models. A provider can sell access to datasets; carry out custom collections for a client; provide infrastructure enabling robot fleets to record and manage their own data; or participate directly in training and evaluating models. The TechCrunch brief does not make it possible to identify XDOF’s exact model. The valuation mentioned nevertheless suggests that the market would place high value on its ability to create or manage robotics data.

Concentration is a structural risk. If a few players own the largest sets of physical interactions, new entrants could find themselves dependent on their access terms. Conversely, a market of specialized providers may also lower barriers to entry for teams that lack the means to buy and operate a large robot fleet. Everything will depend on the degree of data openness, their interoperability and the possibility for clients to use them with their own hardware and models.

Manufacturers have an interest in keeping part of their collection in-house, because data from their machines reflect their mechanical specifics and customer use cases. Model laboratories have an interest in accumulating cross-cutting corpora in order to improve generalization. Data providers, finally, seek to become neutral or compatible with several platforms. These interests can align, but they can also compete. Control over data will partly determine who captures value when a robotics model becomes more capable.

The reported negotiation around XDOF therefore comes at a stage when the sector is still seeking its boundaries. Control software, data, the model, the simulator and the robot are deeply connected. Yet investors are already funding companies positioned at each of these links. A high valuation for a data player would signal that the market no longer sees this layer as a mere execution service, but as technology and infrastructure capable of influencing the development of the entire industry.

A high valuation, between industrial scarcity and speculation on action models

The valuation level mentioned by TechCrunch calls for a cautious reading. A financing discussion guarantees neither that the deal will close nor that the initially envisioned terms will be maintained. In the technology ecosystem, announced or reported valuations reflect expectations for a company as much as its current revenues. They depend on competition among investors, the amount raised, rights attached to shares and perceptions of future potential.

In XDOF’s case, the interest lies less in the potential “unicorn” label itself than in what it reveals about the underlying reasoning. Investors appear ready to assign significant value to a player exposed to robotics data, even before general-purpose robots have demonstrated mass deployment in everyday life. It is a bet on the impending maturation of action models, but also on the idea that the best datasets will be scarce when demand accelerates.

This logic has a tangible industrial component. Building collection capacity requires equipment, sites, protocols and teams. Unlike a conventional software application, the activity may require significant hardware and operational expenditure. If a company has truly accumulated useful data and established a reproducible system for generating new data, it potentially owns infrastructure that competitors cannot replicate in a few weeks.

But the sector remains subject to major uncertainties. There is not yet a consensus on the amount and type of data required to train truly versatile models. Advances in simulation, self-supervised learning and transfer between robots could reduce certain collection needs. Conversely, they could increase demand for high-quality real data, essential to validate and adjust models against the imperfections of the physical world.

Simulation is precisely one of the points of tension. It makes it possible to generate scenes and trajectories at scale in virtual environments. It is useful for testing policies, exploring dangerous or rare situations and reducing the cost of early trials. But a simulated environment does not perfectly reproduce friction, deformation, imperfect sensors, unexpected objects or human gestures. The transition from simulation to reality, often called sim-to-real, remains a challenge. Data collected on real robots therefore retain particular value.

XDOF’s potential valuation also comes at a time when the boundary between a dataset and an AI product is becoming less clear. A collection platform can incorporate automated annotation tools. A provider can help define tasks and metrics. A dataset can be coupled with a pretrained model, a simulation environment or a deployment system. Under these conditions, valuing only “data” would be reductive; investors may be seeking to fund a complete capacity to produce actionable knowledge for robots.

The risk for the market would be to confuse the strategic importance of an asset with the certainty of a sustainable business model. Physical data are costly and difficult to acquire, but their value depends on their freshness, diversity, usage rights and fit with the needs of real customers. A corpus useful for training an arm in a laboratory may not meet the constraints of a warehouse, professional kitchen, production site or care environment.

Verification criteria will therefore be essential. Customers and investors will need to take an interest in data provenance, collection conditions, consent mechanisms when human operators take part in demonstrations, annotation quality, export arrangements and the performance achieved by models trained on these data. The number of recorded hours will not suffice as a sole indicator. Progress will have to be measured on defined tasks, specific robots and representative environments.

For young companies in the field, this phase may encourage a race for capacity: more robots, more collection sites, more operators and sensors. It may also encourage consolidation, with robot makers potentially seeking to internalize data while model laboratories could enter into exclusive agreements with providers. XDOF’s Series B discussion, barely three months after emerging from stealth according to TechCrunch, shows just how fast this race can move.

Finally, robotics must be distinguished from conversational AI on one major point: the validation timeline. A language model can be deployed online quickly, then updated through software. A robot that interacts with objects, equipment or people must be tested under stricter conditions. Design, safety and integration cycles are longer. The market can therefore value data infrastructure very early, while still waiting several years before certain robotics uses spread at scale.

What this race could change for France and Europe

For French and European players, the rise of robotics data raises a question of technological autonomy that is not limited to language models. Europe has recognized expertise in industrial robotics, computer vision, embedded systems, automation and academic research. France in particular hosts laboratories and companies working on robotics, AI and industrial applications. But the capacity to collect and exploit physical interactions at scale is a distinct layer, requiring investment, infrastructure and partnerships with end users.

European industrial environments can represent an opportunity. Factories, warehouses, laboratories and logistics sites concentrate tasks where robotics is already present or being sought. They potentially offer collection conditions relevant to concrete use cases. But they also raise strong constraints: process confidentiality, employee safety, limited availability of production lines and a legitimate reluctance to circulate data describing sensitive operations.

This tension could favor collection models closer to customers. Rather than centralizing all raw data, solutions can be designed to process or anonymize certain information locally, according to companies’ requirements. This does not automatically resolve governance challenges, but it shows that value does not lie solely in centralized accumulation. The ability to meet confidentiality constraints and demonstrate control over data flows can become a commercial advantage in Europe.

The European regulatory framework is also set to affect how these technologies are developed and deployed. The European regulation on artificial intelligence, the AI Act, introduces a risk-level approach for certain AI systems. It does not by itself transform the rules for collecting robotics data, but it belongs to an environment in which transparency, safety and governance requirements are becoming increasingly important. For companies targeting physical uses, compliance cannot be separated from engineering processes.

Liability issues are particularly sensitive when a model controls a physical device. An unsuitable result from a conversational model can be inconvenient or costly; a robotics error can damage an object, interrupt production or create a safety risk. Training data must therefore be complemented by control mechanisms, operational limits, testing procedures and appropriate oversight. No volume of data exempts a manufacturer or integrator from these obligations.

For French companies considering automation, discourse around general robotics models should therefore be read pragmatically. The most immediate use cases often remain those in which tasks are sufficiently repetitive, environments sufficiently structured and benefits measurable. The arrival of better models and more varied data may gradually broaden the scope of operations that can be automated, particularly where traditional programming reaches its limits. But this development will also depend on integration, maintenance, workforce training and safety costs.

International competition may nevertheless open opportunities for specialized European players. They can develop datasets focused on sectors where Europe has expertise: manufacturing, logistics, energy, agriculture, aerospace, healthcare or technical services. This is not to assume that XDOF or other companies will target these markets; the point is more general. Physical data tend to be contextualized. A corpus devoted to a demanding industrial task can have high value even if it is less spectacular than a general-purpose collection of robotics videos.

Research institutes and public programs also have a role in interoperability. Open initiatives, following the example of Open X-Embodiment, show the value of formats that make it possible to combine several robots and several tasks. For Europe, encouraging representation standards, evaluation protocols and secure-sharing mechanisms could prevent access to data from becoming entirely dependent on a few non-European platforms. Openness does not necessarily mean publishing sensitive industrial data; it can concern tools, methods and interfaces.

The issue of skills is equally important. AI-powered robotics requires people capable of connecting mechanics, control engineering, machine learning, vision, embedded software and safety. Data collection itself is not a secondary activity: it involves sensor specialists, operators, validation engineers and business-process experts. A strategy focused only on buying imported models or robots would risk leaving aside this operational expertise.

From this perspective, XDOF’s potential valuation is a signal for the French-speaking market, not evidence that all business models are already established. It serves as a reminder that technological sovereignty is not only at stake in ownership of algorithms or computing infrastructure. It also depends on legal, reliable and durable access to the observations and gestures that will enable machines to learn to work in the environments where European companies create value.

Toward a market where data, the robot and the model can no longer be separated

The financing discussion around XDOF comes at a turning point. Robotics is benefiting from renewed attention prompted by generative models, vision-language systems and ambitions around humanoids. But its adoption will not necessarily follow the same pace as that of software assistants. In the physical world, every advance must pass the test of reliability, integration and cost. Data are one of the most concrete means of improving that reliability, provided they are representative of situations encountered in the field.

The most likely short-term scenario is not that of a single universal model controlling all robots in all contexts. Hardware and occupational constraints will remain strong. However, more generic models could serve as a base and then be adapted to specific fleets, tasks and environments. Under this model, data become the raw material of adaptation. They make it possible to specialize a system without starting from scratch for every new use case.

This development can increase the value of companies capable of organizing a complete cycle: capturing interactions, verifying them, training or adjusting a model, measuring its results, then using the robot’s new operations to enrich the corpus. Such a data loop can become more important than the one-time ownership of a large stock of sequences. Data quickly lose relevance if they no longer correspond to the robots, software and tasks actually deployed.

The market will then have to choose between two visions. The first bets on large centralized platforms, owning massive corpora and capable of offering models to many clients. The second relies more on local, sector-specific and protected data, held by manufacturers, integrators or user companies. Both can coexist: a general model can be trained on broad data and then adjusted with confidential data from a particular site.

For XDOF, the challenge will be to demonstrate its place in this architecture. If the Series B reported by TechCrunch materializes at around a $1.2 billion valuation, the company will face high expectations. The market will want to know whether its data actually improve robot performance across varied tasks, whether they are compatible with different hardware and whether its collection capacity can keep pace with the sector’s expected acceleration. A significant valuation can fund the necessary infrastructure, but it also makes execution more visible and more demanding.

Over the longer term, the decisive factor may be less the raw volume of data than their power to generalize. The winners will not necessarily be those recording the most gestures, but those identifying the most informative demonstrations, the most instructive failures and the most useful variations. Robotics data could thus become an asset comparable to a detailed map of the world of actions: not merely an archive of past movements, but a foundation enabling machines to better anticipate the consequences of their future gestures.