Robot Training Data Companies: The 2026 Landscape

Two years ago, there was no market for robot training data. If you were training a robot foundation model, you collected the data yourself. Today there is a real market, and it filled up fast. Most teams evaluating it are working from a picture that is six months out of date.

This post lists the companies supplying real-world data to robot foundation model and world model teams as of mid-2026. For each one, it covers what they capture, how they capture it, and which industries they have actually worked in.

Disclosure. This survey is published by DreamVu, which is one of the companies listed in it.

Four questions that separate these companies

Most vendor conversations skip these and go straight to price per hour. Price per hour tells you the least.

Do they capture real data or simulate it?

Simulation gives you unlimited volume, perfect labels, and physics you control. It does not give you the real world. Real capture gives you the real world and none of the other three. Most serious programs use both. The question is where the line sits.

Do they capture egocentric only, or egocentric and exocentric?

Egocentric means first-person, from a head-mounted camera. It is the most common capture mode because it is cheap, and because the first-person view looks like what a robot camera sees. Exocentric means third-person. It shows the body, the room, and the objects the person is not looking at. Most vendors sell ego only.

Is the 3D geometry measured or estimated?

Some vendors measure geometry directly using stereo cameras, LiDAR, or optical motion capture. Others record ordinary video and estimate depth afterwards. Both are valid. They are not the same product and should not cost the same.

Do they cover manipulation only, or the whole body?

Most available data is hands and tabletops. Grasping, bimanual work, fine motor skills. That is the right data for a robot arm. It is not enough for a robot that has to walk into a room, look around, move past a person, and then pick something up. Walking, balance, and whole-body coordination are a different capture problem. Few vendors solve it.

The comparison table

The companies

1. Scale AI

Scale is the incumbent. It is the only company here with ten years of history in related work. It took the data operation it built for autonomous vehicles and extended it to robotics. That means a global network of collectors, dedicated data collection robots, human demonstration capture, and the annotation tooling it already owned. It has named Physical Intelligence, Generalist and Cobot as robotics customers.

The strength is scale and the ability to run a whole program for you. The trade-off is that Scale is a general data operation applied to robotics, not a company built around one capture method. If you know exactly what you want collected, that helps. If you want an opinion on what to collect, it helps less.

2. XDOF

XDOF came out of stealth in mid-2026 with $70M from Thrive Capital, Spark Capital, a16z, Lux and WndrCo. It is a UC Berkeley spinout. The founders built GELLO, a low-cost teleoperation rig that many academic labs adopted. At launch, it released ABC-130K, which it calls the largest open-source bimanual robot manipulation dataset.

XDOF describes its work as a three-tier data pyramid. The top tier is teleoperation on the exact robot being deployed. The middle tier is generalized teleoperation. The bottom tier is egocentric capture from humans doing everyday tasks. The tiers run from most faithful and most expensive down to least.

XDOF has been unusually direct that hardware choices upstream decide data quality downstream. That is correct, and most vendors avoid saying it. The focus is manipulation. The flagship dataset is bimanual. That is a deliberate choice, not an oversight.

3. Lightwheel

Lightwheel has the most complete simulation offering. It builds SimReady assets in OpenUSD, which are physics-accurate objects and scenes for Isaac Sim. It also sells EgoSuite for egocentric human data and RoboFinals as a simulation evaluation platform. It works closely with NVIDIA. It has named Google DeepMind, Figure, AgiBot, ByteDance, Geely and BYD among the teams using its assets.

Lightwheel is also the most serious about the sim-to-real gap, because it has to be. Its recent work measures real physical properties such as friction, bending stiffness and torsional rigidity, then uses those measurements to build simulation assets that behave correctly. That is a real engineering answer to the usual criticism of synthetic data.

The limit is the one synthetic data always has. When a robot fails in the field for a reason nobody modelled, more simulation does not fix it.

4. Mecka AI

Mecka has raised roughly $68M in total. An $8M seed led by Neo closed in August 2025, followed by a $25M Series A and a $35M follow-on, both led by Framework Ventures and announced together in June 2026. Menlo Ventures, SV Angel and Kindred Ventures also invested. The company says it projects a $100M annual run rate from contracts already signed. That is the strongest commercial claim anyone in this category has made in public.

The company launched from Toronto in 2025 and is now headquartered in New York.

Mecka collects human motion using custom body-worn sensors and iPhones. It captures hand gestures, walking gaits and full-body kinematics.

Mecka named 1X Technologies as an early partner when it launched, and has said it holds contracts with several Fortune 100 companies, though it has not named them. It has not published a dataset, which makes the technical offering hard to evaluate from outside. What is clear is the modality. Consumer cameras plus motion sensors produce very good kinematic data and ordinary video. They do not produce measured scene geometry.

5. Config

Config is based in Seoul and San Jose. It raised a $27M seed at a reported $200M valuation, backed by the venture arms of Korea’s largest manufacturers including Samsung Venture Investment. It describes itself as the data infrastructure for general-purpose bimanual robotics. It has stated public targets of one million hours of data operations and $10M ARR by the end of 2026.

Two things stand out. Config closes the loop. It trains and tests policies on its own data rather than shipping a dataset and leaving. It also ships a first-generation bimanual foundation model. And its investors are its customers, which gives it a defined industrial deployment context that most vendors do not have.

The scope is stated on its own website. Bimanual manipulation. Config is not claiming to solve walking.

6. DreamVu

DreamVu captures egocentric and 360-degree exocentric video at the same time, synchronized frame by frame, with native stereo depth from its own camera.

The ego stream is head-mounted and matches what a robot camera sees. The exo stream comes from a single 360-degree stereo rig that sees the whole scene. It captures the person’s full body, the objects behind them, and the layout of the space, with measured geometry rather than depth estimated from flat video.

The reason for the design is that the two views answer different questions. Pairing them frame by frame lets a model learn how one relates to the other. DreamVu’s published work on retail environments found that training on exocentric data alone matched or beat training on ego and exo combined across most measured metrics. That was not the expected result. It suggests the third-person view carries more signal than the field currently assumes.

In practice, this covers whole-body and navigation tasks as well as manipulation, because the exo rig records the body and the room, not just the hands. DreamVu has run capture programs in retail, warehousing and logistics, industrial manufacturing, and commercial services, in both the United States and India.

The trade-off is real. Proprietary hardware means capture is tied to rigs. DreamVu cannot hand a thousand people a phone and start collecting tomorrow, and its programs take longer to set up. It is the wrong choice for a buyer who wants a lot of undifferentiated volume quickly. It is the right choice for a buyer who has hit the ceiling on what first-person video can teach a model.

7. Bones Studio

Bones Studio is the most technically unusual company on this list, and the one people miss most often, because it came from entertainment motion capture rather than robotics.

It runs optical motion capture at 120 frames per second with sub-millimetre accuracy. It records 3D motion, multi-view video, audio, face capture and 3D scene reconstruction together in one take, synchronized frame by frame.

Its RP01 data was used to build SONIC, a whole-body control model for humanoid robots. It released BONES-SEED, a set of over 142,000 annotated motion sequences, in NVIDIA SOMA and Unitree G1 formats.

Bones is the clearest whole-body and locomotion specialist in the category. The limit is the stage. This is studio capture with performers, not people doing their real jobs in real workplaces. For locomotion policy and motion generation, that is fine and probably better. For a robot that has to handle a cluttered warehouse aisle, it is a different kind of data.

8. Encord

Encord is not a capture company and does not claim to be. It is the data infrastructure layer. It raised a $60M Series C led by Wellington Management in February 2026, taking total funding to $110M, and reported that its physical AI revenue grew tenfold over the previous year. The platform handles LiDAR, radar, point clouds and synchronized video, with curation, annotation, versioning and evaluation.

It belongs on this list because many teams shopping for a robot data vendor actually need tooling, not collection. And whoever you buy capture from, something has to manage the result. Encord works alongside every other company here.

9. Human Archive

Human Archive is a Y Combinator Winter 2026 company based in San Francisco. It raised an $8.2M seed led by Wing Venture Capital and NVP Capital. It runs custom head-mounted rigs at scale, with more than a thousand active headsets across homes and manufacturing sites, and a large share of its collection network operates in India. It combines egocentric video with depth, motion capture and tactile sensing into synchronized datasets, then runs its own QA, anonymization and annotation.

It is early. The multimodal synchronization thesis is right, and the deployment footprint is real.

10. Luel

Luel is a rights-cleared marketplace rather than a capture operation. It sells curated subsets of Ego4D and Ego-Exo4D, and runs custom collection through a large contributor network, with consent recorded from the moment of capture. Lightspeed backs it.

The pitch is legal cleanliness and speed. If you are under GDPR or facing procurement questions about where your data came from, that matters. The trade-off comes with the model. Quality varies between contributors, and curated academic subsets are by definition data your competitors also have.

11. Build AI

Build AI is the volume leader in open egocentric data, and it did not exist eighteen months ago. Founded by Eddy Xu and backed by roughly $15M from Abstract Ventures, Pear VC and HF0, it works with factory networks across Southeast Asia and equips workers with its own camera glasses.

The release history is the story. Egocentric-10K in November 2025, then Egocentric-100K in December 2025 at over 100,000 hours from 14,228 factory workers, then Egocentric-1M in April 2026 at roughly one million hours. All released openly on Hugging Face under a permissive license that allows free commercial use.

It is narrow on purpose. Every hour is a worker doing skilled industrial labor, not walking around or sitting idle, so the density of active manipulation is high. If industrial is where you are deploying, this is the first place to look, and it costs nothing.

It is also the clearest evidence of what has happened to the price of first-person video.

How to choose

Match the vendor to the problem you are trying to fix.

• Your policy fails on the target robot, on a task you already know. You need teleoperation on that robot. Look at XDOF’s top tier, Config, or a custom Scale program.
• Your policy fails on task variety. You need volume and breadth. Look at Mecka, Luel, Build AI, Human Archive.
• Your policy fails on physics you cannot safely stage. You need simulation. Look at Lightwheel.
• Your policy fails on walking or whole-body coordination. Look at Bones Studio for motion accuracy and DreamVu for whole-body activity in real workplaces.
• Your policy fails on spatial reasoning and does not understand the room. You need exocentric views and measured geometry. This is the gap the market currently serves worst.
• You have the data and cannot manage it. Look at Encord.

The most common mistake in this market is buying more of the data you already have. If your model is failing on scene understanding, another ten thousand hours of first-person footage of hands will not fix it. It will feel like progress right up to the point where you evaluate.