Robot and Physical AI data has gone from a relatively hidden infrastructure function to a crowded capital target over the past year. In an interview with PANews, Axis Robotics founder Chris Feng outlined why investors are moving into the segment and how Axis is trying to build around that demand. The company closed a $12 million seed round in late July led by Hack VC, while startups built specifically around robot training data have begun to surface in both China and overseas markets.
The acceleration has been especially visible in China. PANews reported that in the first half of 2026 alone, 25 Chinese embodied AI data startups raised more than 17 billion yuan in total. Guanglun Intelligence completed multiple large fundraising rounds during the year, and Mihive Technology, a data company incubated by AgiBot, also raised several hundred million yuan soon after it was founded. Outside China, XDOF announced a $70 million financing round in June, and Axis completed its seed round at the end of July.
Axis describes itself as a “Physical AI data engine.” Its current entry point is a distributed contributor network paired with web-based simulation data collection, and it is extending that stack into first-person data and post-training data products. Feng said the company has already collected more than 2.2 million robot trajectories through browser-based teleoperation, representing 28,000 cumulative hours of data.
Even with capital already flowing in, Feng said the robot data market is still unsettled. Questions remain on the ultimate size of demand, how the industry will balance real-robot, simulated, and first-person data, and whether independent data companies can build durable business models. Those questions framed the PANews interview with Feng, which focused on Axis’ production model, technical path, and plans for the next stage.
The data shortage behind Physical AI
Feng did not start in robot hardware or core robotics algorithms. He came into the field through consulting, investing, and data infrastructure, and said the move into robotics became clear once the scale of the data gap in Physical AI came into view.
He used GPT-3 as a reference point, saying it required about 15 million hours of human internet data for training. Robots, by contrast, have to deal with a physical world that is much more complex than the digital world. They do not just need recognition and reasoning. They also need to understand space, motion, object states, and what changes after an action takes place.
On that basis, Feng said robotics may need 100 million hours of data or more if the field is going to reach its own “GPT moment.” The difference is not only one of scale. The underlying data base is also far weaker. Before large language models arrived, the internet had already developed for decades. Web pages, books, forums, code repositories, and other digital content created a naturally accumulated body of trainable data.
Human activity in the physical world never went through a comparable process of systematic recording. People cook, organize objects, and use tools every day, but those actions are rarely captured continuously and almost never converted at scale into standardized training data for robots. As VLA models and world models advance, robot model companies have become direct buyers of this kind of data, while some robot makers are building model and data capabilities in parallel.
Demand is also fragmenting as robot form factors diverge. Humanoid robots, robotic arms, and specialized systems built for industrial or service settings do not all need the same data. Once demand starts to expand at that level, the limits of traditional collection methods become harder to ignore.
Early robot companies typically relied on real-robot teleoperation, where human operators directly control a robot performing tasks and the resulting trajectories are recorded. That data is the closest match to deployment conditions, but it is slow to produce, expensive, and constrained by the number of robots, physical sites, and operators available.
Feng said that if robot foundation models ultimately need data on the order of tens of millions of hours or more, simply adding more robots, more facilities, and more teleoperators will not produce an economically efficient linear scale-up. The real world is too varied, and a company’s internal fleet and collection environments cannot easily cover what a model may face later on. For that reason, he sees one of the central business opportunities in Physical AI as producing sufficiently diverse data at controllable cost and at a scale that can still improve model performance.
A data pyramid: precision, scale, and generalization
Feng’s view is that robot data should not be framed as a single winning route. Different data sources vary sharply in realism, cost, scalability, and cross-embodiment reusability. He grouped the main paths into real-robot teleoperation, simulation, and first-person human data, and described them as a rough pyramid based on how directly they match a target robot body.
At the top: real-robot teleoperation
At the top of the pyramid is real-robot teleoperation. Operators directly control a physical robot through tasks such as grasping, placing, and moving objects, while sensors record joint states, end-effector signals, visual input, and actions. The appeal is straightforward: the training data comes from the real machine, so physical information is as complete as possible and the mismatch with the target embodiment is minimal.
The drawbacks are just as clear. Feng cited industry figures showing that one hour of high-quality real-robot teleoperation data can cost as much as $200 in total. The data is often tightly tied to a specific robot body, which limits reuse across different systems. A trajectory collected on a Unitree G1, he said, cannot be directly adapted to robots with different morphologies.
In the middle: simulation data
Simulation sits in the middle of the pyramid. It attempts to recreate a robot’s size, joints, motion range, and task environment inside a virtual space, where either human operators or software agents can perform tasks. Feng pointed to Isaac Sim and MuJoCo as examples of simulation platforms widely used in robot research and development.
Its strongest advantages are scalability and control. In the physical world, companies cannot rebuild thousands of kitchens, warehouses, or factories just to collect data. In simulation, comparable scenes can be duplicated quickly and run in parallel. Existing trajectories can also be expanded into more environments by changing parameters such as lighting, materials, camera positions, object types, and spatial layouts.
Task states in simulation are easier to read precisely as well. Whether a task was completed, whether a robotic arm collided, and where an object ended up can all be pulled directly from the environment. That makes validation and automated filtering easier.
Still, simulation does not fully reproduce the real world. Friction, object deformation, contact states, sensor noise, and many uncontrolled factors all contribute to the sim-to-real gap. A model that performs well in simulation may lose performance once moved onto a real robot. Feng said closing that gap remains one of the field’s central challenges.
At the base: first-person human data
Another route that has heated up over the past two years is first-person human data. The idea is to reduce dependence on expensive robot hardware by asking contributors to wear head-mounted cameras, GoPros, or simply use phones to record real actions such as cooking, folding clothes, cleaning rooms, or using tools.
The premise is that if future robot visual and action logic ends up sufficiently similar to human behavior, then large-scale recordings of how humans interact with the world could become a rich source of pretraining data for robot foundation models. The attraction lies in potential scale and environmental diversity. Compared with deploying large fleets of robots for teleoperation, collecting behavior data from ordinary people in homes, factories, and shopping areas can lower both cost and participation barriers while covering many more real environments.
Feng said the uncertainty is substantial. The training value of first-person human data depends heavily on what future robot bodies look like. If robots are built around wheeled bases, robotic arms, or specialized grippers rather than humanoid forms, the direct transfer value of human action data at the motion level drops sharply. Even if humanoid robots become mainstream, human movements still cannot be copied directly. Hands and body motion have to be retargeted into trajectories a robot’s joints and hands can execute.
He also highlighted a point he believes is often overlooked: camera perspective. In principle, first-person data should align as closely as possible with human visual position, but collection devices may sit on the head, chest, or elsewhere. Changes in viewpoint materially alter the spatial relationship among hands, objects, and the camera. For models that need to learn 3D spatial and manipulation relationships, that becomes a real challenge.
That is why Feng does not see real-robot data, simulation, and first-person data as simple substitutes. Real-robot data delivers precision. Simulation amplifies scale. Human video contributes real-world scene diversity. In his view, the future robot data stack is more likely to be a combination of sources than a winner-take-all model.
He also pointed to research that is starting to step away from the idea of a single “best” kind of data. Dyna Robotics’ recently released Dyna-2 reportedly observed scaling effects in the transfer from human behavior data to robot capability using more than 1 million hours of first-person human video for pretraining. Axis’ own research, he said, suggests that filtered and processed simulation trajectories can also keep improving model performance as scale increases.
From annotation to data production
For Feng, the more important benchmark is no longer whether each individual data point matches a target embodiment as closely as possible. Scale, diversity, and whether the data actually turns into model capability are becoming the deciding factors. That creates a more basic question for robot data firms: are they only there to collect and process according to client instructions, or should they play a deeper role in designing and producing the data a model needs?
Axis is aiming for the second role and uses the term “Physical AI data engine” to make that distinction clear. Feng said “data annotation” and “data production” are not the same business layer. Annotation starts with existing data. The customer decides what to label and how, and the vendor executes. Data production goes further. A data company has to judge what a model is missing, design tasks around training goals, organize collection, and adjust the next round of data according to model performance.
He used kitchen manipulation as an example. If a robot model needs to learn kitchen tasks, a data company has to decide which tasks to design, how basic actions such as grasp, push, pull, open, and close should be combined, whether depth information is necessary, how cameras should be placed, and which failure cases should be collected on purpose.
Feng said Axis’ commercial work with clients including Booster Robotics and Qingyu Technology has already moved beyond pure data delivery into “customized data production plans built around model training goals.” In its work with Booster Robotics, he said Axis developed dedicated foundation models tailored to the Booster T1’s visual inputs and action space. He said that setup supports rapid transfer and iteration for specific tasks with very few samples, cutting the barrier to embodied AI development and reducing engineering adaptation costs.
That is also why Axis stresses the term “data engine” rather than “data factory.” In Feng’s description, commercial clients can first define their target scenarios and required capabilities, then Axis designs the tasks and uses its distributed contributor network to complete the collection work.
Axis is also running its own dataset and model validation research. Feng said the team released the AXIS Franka dataset in the first half of the year. He described it as one of the largest open-source robot manipulation datasets for the Franka Research 3 robotic arm, covering 207 diverse manipulation tasks, more than 50,000 human demonstration trajectories, and over 60,000 task-scene variants.
To evaluate training outcomes, the research team used the dataset for continued pretraining on the state-of-the-art VLA model π 0.5. According to Feng, the model trained on the AXIS-100% data snapshot reached an overall success rate of 88.8% on the LIBERO-Plus robustness benchmark.
He also said the AXIS Franka dataset produced clear gains under real-world disturbance conditions. Model robustness improved by 13.7% under sensor noise and by 11.3% under camera viewpoint shift. As the dataset snapshot increased from 25% to 100%, the model’s manipulation generalization showed a stable scaling trend, which Axis takes as evidence of a positive relationship between data scale and downstream task generalization.
Not tied to one data type
Feng said robot data is still a fast-changing market without a settled template. Industry attention has broadened from real-robot teleoperation to simulation and first-person data, and as foundation models move deeper into post-training, new demand is emerging around failure states, correction processes, and recovery trajectories.
He put the concern plainly: “If the market mainly wants first-person data today, I can just do first-person data. What happens if six months later nobody needs it?” Because of that, he does not want Axis to be defined as a company that only does one kind of data.
His approach is to separate relatively stable infrastructure from constantly changing collection methods. At the bottom are the distributed contributor network, task management, data validation, cleaning, augmentation, and processing pipelines. At the top are the different data products and collection methods added according to model and customer needs.
Axis started from web-based simulation data collection and has since expanded into first-person data and model post-training. Feng said its latest product already includes correction-data collection for model bias and failure states. The team is also working on post-training data collection and processing for approaches such as DAgger, and plans to release its first large-scale Human-Gated DAgger post-training dataset before year-end.
He argues that as robot models keep iterating, the value of a data partner will also be judged differently. The question is not just how much data can be delivered in one pass. It is whether the company can reliably produce high-quality data, operate a contributor network over time, maintain the right data-processing tools, and adjust the content and structure of the next round based on model behavior.
Feng said this thinking is partly informed by the paths taken by Scale AI and Surge AI, with particular attention on Surge AI. His goal is for Axis to become “the Surge AI of robotics.” In his view, Surge AI’s edge is not only its scale in data annotation. It is that the business continues to track shifts in model demand, moving from large-model post-training into newer steps such as agent training environments. That ability to keep changing the data and tools offered as Physical AI training changes is what Axis wants to replicate.
Where blockchain fits
Axis carries a label that many peers in robot data do not: blockchain.
Feng was explicit that blockchain is not needed for the core act of producing robot data. Simulation still runs in conventional robotics environments, and model training is no different from other AI companies. The role of blockchain comes after data is created. Axis assigns a data ID to verified trajectories and records the relationships among tasks, data, and contributors on Base.
He said the technology addresses two main issues.
- First is contributor incentives. When data is produced by a globally distributed user base, users differ in both quantity and quality. The platform needs to know who contributed what and whether the resulting data actually met the required standard. For contributors, the key is not simply volume, but who provided data that was genuinely useful.
- Second is provenance tracking. Feng said AI training today often resembles a black box. Outside observers can usually see only the final model, not which data produced which capability or who generated that data in the first place. If the relationship among tasks, trajectories, and contributors is recorded continuously, a relatively complete source record can remain when a client eventually purchases a given dataset.
Axis has already launched a points system to record and measure user contributions. Feng added that if the company issues a token in the future, contribution points could become one basis for incentives. If a contributor’s data is ultimately commercialized, Axis is also considering whether that person could participate in revenue sharing.
Those mechanisms are still being designed and refined. Rather than building a token economy first and then trying to find a use case, Feng said the priority is to establish what concrete problems blockchain can solve. “I have never thought of myself as a crypto robotics company,” he said. In his definition, Axis is a robot data company first. Blockchain is used only where it helps with contributor incentives and provenance tracking, and if it does not create practical value, there is no need to add it just for a “Crypto + AI” story.
The next six to 12 months
Axis’ roadmap for the next six to 12 months is concentrated around products, community growth, and commercialization.
On products, the company plans to continue expanding first-person data collection and add UMI data as well as teleoperation data for mobile manipulation tasks, broadening the range of data types and tasks supported on the platform. Beyond data collection portals for general users, Axis also plans to offer task-generation APIs and data-processing tools to enterprise customers and developers, turning more of its underlying production capabilities into products.
Community growth is another focus. Feng said Axis wants to expand its community to three times its current size over the next half year and deepen coverage across different regional markets. The company is not limiting that effort to ordinary data contributors. It also wants more developers, researchers, and industry participants on the platform, with the longer-term aim of building an open ecosystem around Physical AI data production.
On commercialization, Axis says it has already won benchmark clients in some vertical segments, and that both simulation data and first-person data have produced paid orders. Over the next six months, the company’s goal is to raise annual recurring revenue to between $500,000 and $1 million and to make it onto supplier lists at leading robot model companies.
For Feng, that second target may matter more than the revenue number itself. Whether Axis can enter the supply systems of top model companies, he said, may be the stronger test of the company’s current approach.