Who Is Building the Training Data Infrastructure for Robotics?

Robotics models are improving quickly, but unlike language models, they cannot rely on vast amounts of existing internet data. To learn how to act in the physical world, they need demonstrations of people and robots performing real tasks, often captured with cameras, sensors and teleoperation systems. This has created a new market for companies building the data infrastructure required to train, evaluate and improve physical AI systems.

Category #1: Human Centric Demonstration Data

What is this category about?

• Teaching robots by recording how people work. These companies capture people performing physical tasks so that robotics models can learn from the way humans interact with objects and move through their environment.

• Creating data without needing a robot. Robot demonstrations are expensive because each recording requires a functioning robot and an operator who knows how to control it. Human demonstrations can be collected with cameras and wearable sensors while a person performs the task naturally.

• Capturing physical knowledge that is difficult to describe. Many everyday actions depend on subtle movements that would be difficult to document as instructions. A person adjusting their grip or recovering from a small mistake produces useful information that a written procedure would miss.

• Building the broad data layer for robotics models. Human demonstrations can expose a model to far more tasks and environments than a robotics company could reproduce with its own machines. The model can then be adapted with a smaller amount of data collected directly from the target robot.

• Bridging the difference between humans and robots. A human hand and a robot gripper do not move in the same way. The data therefore needs to be processed so the model can identify the underlying action rather than simply copying the exact human movement.

What do products in this category do?

• Most products begin with a person wearing a camera. The camera is usually mounted on the head so it records the task from the viewpoint of the person performing it. Some systems add cameras on the wrists to capture hand movements that may be hidden from the head camera.

• The recording is converted into structured training data. The provider identifies the actions taking place and divides a long activity into individual steps. The dataset may also show where the hands and objects are located within the scene.

• Companies such as Nferent AI, Roborecs and Renlei Labs organize custom collection projects. A robotics company describes the task or environment it needs. The provider recruits participants, records the demonstrations and delivers a dataset prepared for model training.

• Unidata and Claru combine collection with data enrichment. Their role is not limited to recording video. They process the footage so that a robotics team receives structured examples rather than a folder of raw recordings.

• Luel, Hub and Instawork use distributed contributor networks. They can ask people in different locations to perform specific tasks and upload the recordings. Instawork has an additional advantage because it already operates a large network of workers across industries where robots may eventually be deployed.

• Human Archive is building its own large collection network and capture infrastructure. The company says it works with more than 100,000 contributors and 500 industry partners. It collects human behavior across environments such as homes, hotels and industrial sites.

• OpenGraph Labs adds richer physical signals to the recording. Its capture systems can connect cameras with tactile gloves and other sensors. This is particularly useful for manipulation tasks where the model needs to understand contact rather than vision alone.

• FPV Labs gives robotics teams the tools to collect the data themselves. Its Stera product uses an iPhone to record video together with depth and motion information. The software processes the recording and exports it into formats that can be used for robotics research.

• Mecka AI connects the data layer with deployment support. Robotics teams can access existing data or commission a collection project. Mecka also helps enterprises evaluate workflows before introducing robots into a real operating environment.

• 6thSense focuses on the information that cameras cannot capture. Its wearable system records how the human hand touches and grips an object. The company then sells this tactile data to teams training robots for precise manipulation tasks.

• Cortex AI combines human demonstrations with robot specific data. It collects workplace footage for model pretraining and robot trajectories for later adaptation. The company also helps customers evaluate models during real deployments.

• The final product is usually a licensed dataset or a custom data collection project. Some companies are beginning to add software that lets customers browse samples or request new data. However, most commercial relationships still require the provider to understand the exact task the customer wants its robot to learn.

What types of customers are using these products?

• Robotics companies building general models for physical tasks. These customers need a broad base of human activity before they can adapt their model to a particular robot. Human demonstrations give the model exposure to a much wider range of behavior than robot data alone.

• Humanoid robot developers. Humanoid systems are expected to work in environments designed for people. First person recordings are therefore useful because they show how humans use the tools and objects already present in those environments.

• Teams working on dexterous manipulation. These customers need detailed information about hand movement and physical contact. Products from 6thSense and Roborecs are specifically designed to capture signals that ordinary video does not provide.

• Research labs developing vision language action models and world models. FPV Labs makes its Stera infrastructure available to researchers, while several other providers release public dataset samples to demonstrate their data quality.

• Frontier AI labs that do not want to build their own collection operations. Running data projects across many locations requires participant recruitment and quality control. Companies such as Luel and Hub allow the model developer to outsource this operational layer.

• Existing workplaces can be data suppliers rather than customers. Factories and warehouses may allow workers to record tasks inside their facilities. The data provider can compensate the company or the workers before licensing the resulting dataset to a robotics lab.

• Public customer evidence remains limited. Most companies are still selling through private samples and controlled access programs. Cortex AI provides one of the clearer proof points because it supported the data curation and independent evaluation of Ai2's MolmoAct 2 robotics model.

What does the landscape look like?

• Most of the category was created during the past two years. A large share of the companies were founded in 2025 or 2026. Their public activity often begins with a product launch, a research paper or an announcement that the company is leaving stealth.

• The market includes both new data specialists and established platforms. Human Archive, FPV Labs and 6thSense were created specifically for physical AI data. Instawork entered the market from staffing, while Unidata expanded from broader AI data services.

• The founding teams often combine robotics knowledge with operational experience. FPV Labs was founded by researchers working on egocentric vision and embodied AI. The 6thSense team includes experience from tactile data collection at Mecka AI and vision guided robotics at Tesla.

• Marketplace experience is also relevant. Cortex AI founder previously helped build Carousell into a large consumer marketplace. Instawork already knows how to recruit and coordinate workers across many locations. These capabilities become useful when the main challenge is collecting demonstrations rather than developing the model itself.

• Most teams remain small. Neocambrian AI and Human Archive currently appear to have fewer than ten employees, although both have been actively recruiting. Hub, Luel and Claru have begun building larger teams around contributor operations and data delivery.

What are the major trends shaping this category?

• Human demonstrations are becoming the pretraining layer for robotics models. The emerging approach is to expose a model to a broad base of human activity before adapting it with robot specific data. NVIDIA says its GR00T 1.7 model was pretrained on approximately 32,000 hours of real human demonstrations and another 8,000 hours of simulation. This suggests that human data and simulation will sit below the smaller set of robot demonstrations needed for deployment.

• The gap between human movement and robot action is becoming easier to bridge. Human videos do not directly contain commands that a robot can execute. Recent research is showing that models can extract the important hand and object relationships before translating them into robot movements. HumanEgo reported strong results using only minutes of human video, while EgoEngine converts egocentric recordings into robot observations and executable trajectories. These results remain experimental, but they strengthen the commercial case for collecting human demonstrations at scale.

• Data collection is moving beyond ordinary video. Cameras can show where a hand moves but not how much pressure it applies. Research such as MimicTouch demonstrates that tactile human demonstrations can improve learning for tasks involving insertion and assembly. This creates room for companies that combine vision with touch rather than competing only on the number of video hours collected.

• Consumer hardware is lowering the cost of collection. A specialized motion capture studio is no longer required for every dataset. FPV Labs uses the depth camera and motion sensors already available inside an iPhone Pro. This makes it possible to record longer sessions in real environments where a traditional laboratory setup would be impractical.

• Standardized dataset formats are beginning to emerge. Robotics data has historically been difficult to combine because every team records different signals and stores them differently. Hugging Face's LeRobotDataset provides a common structure for video, sensor signals and episode metadata. Its latest version is designed to stream and manage datasets containing millions of demonstrations. This should make data from independent providers easier to integrate into existing training pipelines.

• Simulation is becoming a complement to human data rather than a replacement. Synthetic environments can generate large amounts of variation, but the simulator still needs an accurate representation of how people interact with the physical world. NVIDIA's robotics stack combines real demonstrations with simulated rollouts instead of choosing one source exclusively. Real recordings may therefore become the material used to build and validate better simulated environments.

• Data quality will increasingly matter more than the number of recorded hours. A long recording can contain waiting time and repeated actions that provide little useful supervision. Providers are beginning to divide tasks into meaningful steps and attach information about success or failure. Standardized metadata also makes it easier for model developers to select the demonstrations that are relevant to a particular training run.

• Rights and provenance are becoming commercial differentiators. Robotics datasets can contain faces and recordings from private workplaces. Buyers need to know who agreed to the recording and what uses are permitted. Luel presents rights clearance as a central part of its marketplace, while Roborecs emphasizes consent and traceability under European jurisdiction.

• European regulation will reinforce the demand for documented datasets. The AI Act requires providers of general purpose AI models to maintain information about their training process and publish a summary of the training content. Rules for higher risk systems also place importance on dataset quality and traceability. Not every robotics model will fall under the same obligations, but European customers are likely to place greater value on datasets with clear documentation and usage rights.

How does the funding environment look?

• Funding is highly concentrated in a small number of companies. Most businesses in the category have no disclosed financing recorded on Axomap. Several describe themselves as seed or pre seed companies even though the amount of capital raised has not been made public.

• Luel has raised the largest clearly category specific round. Lightspeed led a $31.2 million financing in May 2026, with participation from General Catalyst and SV Angel. The size of the round reflects the company's ambition to build a large global marketplace rather than operate as a small data collection agency.

• Human Archive has also attracted meaningful early funding. The company raised an $8.2 million seed round while still operating with a small team. Its current hiring suggests that a significant part of the capital will be used to build capture hardware and the operational infrastructure needed to coordinate its contributor network.

• Cortex AI raised $6 million to develop its real world data operations. The company is using a marketplace model that connects robotics labs with workplaces and robot operators. This gives it a broader offering than the sale of prerecorded video alone.

• The category has therefore raised approximately $52 million in directly attributable dollar financing. This calculation includes the disclosed rounds for the specialist human data companies.

• Investors appear most interested in companies that can build a repeatable supply network. Luel and Human Archive are not simply selling a collection service. They are building systems that can continuously recruit contributors and deliver new data across many environments.

• Hardware is becoming part of the funding thesis. Human Archive and 6thSense are developing their own capture devices, while FPV Labs is trying to achieve a similar outcome through consumer hardware. Better capture infrastructure can improve data quality while reducing the cost of each additional demonstration.

• A large share of spending will go toward operations rather than model research. These companies need people who can recruit contributors and distribute equipment. They also need to verify recordings before delivering them to customers. The recent hiring activity at Human Archive, Hub, Claru and Neocambrian AI reflects this operational burden.

Category #2: Robot Centric Demonstration and Teleoperation Data

What is this category about?

• Collecting training data through the robot itself. These companies record a robot while a human operates it. The resulting dataset shows what the robot observed and exactly how its motors were controlled during the task.

• Translating human intent into robot actions. The operator may use a headset or a physical controller that resembles the robot. When the person moves, the system converts those movements into commands that the robot can execute.

• Creating data that matches a particular robot. Human video can teach a model what a task looks like, but it does not contain the exact motor commands needed to perform it. Robot demonstrations provide this missing connection between perception and action.

• Producing more valuable but more expensive demonstrations. Every recording requires access to a robot and a working teleoperation setup. Collection is therefore harder to scale than ordinary video, but each episode can be used directly to train a robot policy.

• Extending data collection into deployment. The same teleoperation system can help when an autonomous robot encounters a situation it cannot handle. A remote operator completes the task, and that intervention becomes a new example that can improve the model.

• Covering more than one part of the data workflow. Some companies provide the teleoperation infrastructure. Others operate the robots and deliver completed datasets. A smaller group combines data collection with model training and deployment support.

What do products in this category do?

• A teleoperation system connects a human operator to the robot. The operator sees the robot’s camera feed and controls its movement from another location. The platform records the observation together with the action performed at every moment.

• The recorded session becomes a structured episode. Instead of giving the customer a collection of video files, the product synchronizes the camera footage with the robot state and action commands. This creates a dataset that can be used for imitation learning.

• Humanola provides the infrastructure needed to turn teleoperation into training data. A robotics team connects its robot through the Humanola software, controls it remotely and receives a structured episode in its own cloud storage. Capture and storage are already available, while processing and dataset licensing remain less mature.

• Sensei Robotics combines teleoperation hardware with a network of human operators. A customer specifies the robot and the data it needs. Sensei then organizes the people and equipment required to collect the demonstrations.

• Aktoria Robotics focuses on remote operation of deployed robots. Its software lets a person intervene when a robot needs help. The same interaction can be recorded and reused as training data.

• XDOF operates at a much larger infrastructure level. The company builds collection systems and delivers production scale robotics datasets. Its ABC 130K release contains more than 130,000 teleoperation episodes across around 200 manipulation tasks.

• Scale AI brings its existing data operations into robotics. Its Physical AI Data Engine supports custom collection and annotation for physical systems. The partnership with Universal Robots places this infrastructure directly inside a production robot training workflow.

• Lightwheel connects data collection with simulation and evaluation. Its platform can recreate a real environment in simulation and use demonstrations to generate additional behavior data. The resulting policy can then be evaluated before it returns to the physical robot.

What types of customers are using these products?

• Robotics companies developing general purpose models. These teams need large volumes of demonstrations across different tasks before their models can perform reliably outside a controlled laboratory.

• Robot manufacturers introducing a new machine. A manufacturer may have built the hardware but still lack the demonstrations required to teach it useful skills. A data provider can create the initial training pipeline without requiring the manufacturer to build a large internal operations team.

• Developers of humanoid and dexterous robots. These systems have many joints and can perform a wide range of movements. They consequently require richer demonstrations than a traditional industrial robot performing one fixed motion.

• Industrial companies deploying robots in production. A factory can work with the data provider to record the exact process it wants to automate. Lightwheel’s work with Geely and RLWRLD’s industrial partnerships illustrate how collection is moving closer to the customer’s real environment.

• Frontier AI laboratories entering physical AI. These organizations already understand large scale model training but may not own robots or operate collection facilities. Scale AI and XDOF provide the physical infrastructure and operational capacity that these teams lack.

• Operators of deployed robot fleets. When a robot fails, a remote operator can complete the task. Humanola and Aktoria are building products that can preserve these interventions rather than treating them as temporary support actions.

What does the company base look like?

• The category includes companies at very different stages. Scale AI was founded long before the recent physical AI cycle. Aktoria and several other specialists only launched publicly in 2026.

• Technical founder experience is particularly important. The XDOF founders previously developed GELLO, a teleoperation system already used by robotics researchers. Sensei Robotics was founded by former robotics engineers from MIT and Aurora Flight Sciences. Aktoria’s founding team also comes from robotics research.

• Hiring is moving beyond pure robotics research. XDOF currently lists 18 roles that cover technical development and customer delivery. Mecka lists 36 openings across eight teams. Lightwheel has added commercial leadership and roles focused on industrial deployments.

• Field operations are becoming a central company function. Collecting demonstrations requires people who can install equipment and manage robots at customer sites. It also creates demand for employees who can verify whether each recording is suitable for training.

• Several companies sit near the boundary between the two categories. Mecka and Lightwheel collect significant amounts of human centric data alongside robot demonstrations. Shift currently records people performing tasks rather than operating robots directly. Their products may eventually span both data sources.

What are the major trends shaping this category?

• Robotics training is moving from individual datasets toward shared data across many types of robots. Google DeepMind’s Open X Embodiment project combined more than one million episodes from 22 robot embodiments. The project showed that data collected on different machines can improve the ability of a model to learn new tasks.

• Robot specific data is becoming the final layer of a broader training process. Models can first learn from human video and simulation before being adapted with a smaller amount of robot data. NVIDIA describes this as a pyramid in which data becomes more expensive as it becomes more closely connected to the target robot.

• Teleoperation is becoming part of a complete development workflow. NVIDIA’s latest GR00T platform connects simulation setup with teleoperation collection and model training. It then supports evaluation before the policy is deployed on the robot. This reduces the amount of infrastructure each robotics company must assemble itself.

• Distributed collection is replacing the single robotics laboratory. The DROID project gathered 76,000 demonstrations across 564 environments by coordinating 50 collectors in several regions. Models trained with this more varied data showed better generalization than models trained inside a small number of laboratories.

• Common data formats are lowering integration costs. The LeRobotDataset format stores robot actions together with video and sensor information in one standardized structure. Version 3 was redesigned to support datasets containing millions of episodes and to stream them without downloading the full collection.

• Real demonstrations are increasingly used as seeds for simulation. NVIDIA generated hundreds of thousands of synthetic trajectories from a much smaller collection of demonstrations. Combining this generated data with real recordings improved the performance of its robotics model compared with using real data alone.

• Robots will gradually collect more of their own training data. Google DeepMind’s AutoRT system operated more than 50 robots and collected 77,000 episodes through a combination of teleoperation and autonomous behavior. The operator remains important, but an increasing share of collection can happen while the fleet is already running.

• Data quality is becoming as important as data volume. A demonstration may contain long periods where the robot is waiting or moving without making progress. XDOF found that adding more unfiltered demonstrations could make a model perform worse. Its newer tools try to identify which parts of an episode actually contribute to completing the task.

• Failure data will become more valuable after deployment. Successful demonstrations teach the robot what to do under normal conditions. Human interventions show what the model misunderstood and create examples from situations that were not anticipated during initial training.

• Teleoperation is developing into a permanent operating layer. Early products treated remote control as a way to collect an initial dataset. Newer companies are designing it as a system that remains connected to deployed robots so that humans can intervene and continuously improve the model.

• The distinction between a data provider and a robotics model company is becoming less clear. RLWRLD collects industrial data to improve RLDX 1, while Lightwheel uses data to power its simulation and evaluation products. XDOF is also moving from dataset delivery into tools that judge and improve robot demonstrations.

• Evaluation is beginning to guide what data should be collected next. Instead of testing the model only after training, robotics teams can identify the tasks where it fails and commission new demonstrations specifically for those weaknesses. RLWRLD’s work on DexBench and Lightwheel’s RoboFinals both reflect this direction.

How does the funding environment look?

• The category has attracted significant capital despite its young age. Excluding Scale AI, Axomap records approximately $190.3 million in disclosed dollar financing across XDOF, Mecka AI, RLWRLD, Humanola, Sensei Robotics and Aktoria Robotics.

• Most of that capital is concentrated in three companies. XDOF has raised approximately $77.8 million. Mecka has recorded $68 million across its seed financing and two later investments. RLWRLD has raised $41 million across two seed rounds.

• XDOF raised the largest directly attributable round in the category. Its $70 million Series A followed a period of private work with robotics laboratories before the company came out of stealth. The capital is supporting a substantial hiring program and the expansion of its collection infrastructure.

• Mecka is financing a broader data and deployment strategy. Its funding supports more than dataset collection. The company is building capture tools and an enterprise deployment operation around the EgoVerse data ecosystem.

• Scale AI makes the category total misleading. The company has raised billions of dollars, but almost all of that financing predates its Physical AI Data Engine. Its capital provides a major advantage in data operations, but it should not be presented as investment specifically directed toward robotics training data.

• Investors appear most interested in companies that own more than the teleoperation interface. The largest rounds have gone to businesses that can operate collection programs and improve the resulting data. They also help customers move from training into deployment.

• The operations layer is absorbing a meaningful share of the capital. Robotics data companies need physical equipment and people who can install it. They also need facilities where demonstrations can be collected repeatedly and safely.