RESEARCHYou Can't Buy Taste. We Tested It.

Best human data providers for robotics and embodied AI

Robotics is the youngest segment in human data and the one where provider claims and provider output diverge most. Nine companies in the current landscape are tagged for robotics and embodied data, and they range from a two-year-old startup that has open-sourced a hundred thousand hours of footage to established annotation vendors adding a physical AI line. This guide is a shortlist organized by provider, with what each has actually shipped. If you are earlier than shortlisting and still working out which capture mechanism your model needs, start with how to tell physical AI data providers apart instead and come back to this.

Key takeaways

  • Open datasets are the strongest credential available here, because you can inspect distribution and calibration yourself. Build AI, XDOF, Human Archive, and Mecka AI all publish; most others do not.
  • Teleoperation, egocentric wearables, and in-facility capture are different operations with different failure modes. Few providers run more than one well.
  • Most of the field is seed-stage. Weigh that against program length, because a two-year collection program outlasts a lot of seed-stage companies.
  • The incumbents annotate robotics data competently and do not produce it. If your constraint is capture, an annotation vendor is the wrong shortlist.
  • Verify sensor configuration specifically. General 3D experience does not transfer across rigs.
  • HumanSignal designs the capture around the model rather than around a single rig, which is the binding constraint on most specialists in this field.

What makes robotics data different

Three properties separate this from every other modality.

It has to be performed. There is no scraped equivalent of a robot completing a contact-rich manipulation task, so every example costs someone's time in a physical space. That makes collection economics, not annotation economics, the thing to model.

The rig determines what is recoverable. Whether force and tactile channels exist, how many cameras are synchronized, what the control latency was during teleoperation: these are decided at capture and cannot be added later. A corpus recorded on the wrong rig is not fixable downstream.

Failure coverage is usually missing. Teleoperation corpora skew heavily toward successful demonstrations because operators retry until they succeed and the failures get discarded. Policies that need to recover from error states need the failures, which has to be specified deliberately.

The providers

Protocol-designed capture across mechanisms

Most of the field below is built around one capture mechanism each. HumanSignal Services is built around the protocol instead: we work out which distribution the model is missing, write the collection protocol to produce it, and run the capture in our own studios, on site at a customer facility, or in the field, with motion, vision, audio, and sensor streams synchronized from the start. That matters when a model needs streams a single-rig specialist does not record, or a distribution their standing operation does not produce. Quality workflows run in Label Studio Enterprise, so agreement metrics and reviewer decisions are reportable during the capture rather than after it.

Open-data specialists

Build AI captures egocentric factory-floor video through a company-built head-mounted camera worn by workers during shifts. It open-sourced Egocentric-100K in December 2025: 100,405 hours from 14,228 workers and 10.8 billion frames, following Egocentric-10K a month earlier. For scale of inspectable egocentric data, nothing else in this set is close.

XDOF runs teleoperation stations for manipulation capture, launched publicly in June 2026 with ABC-130k under Apache 2.0 and a companion arXiv paper with UC Berkeley faculty co-authors. Press reports around 70 million dollars raised and about 20 customers including frontier labs. The open release plus peer-reviewed methodology is the strongest research credential here.

Human Archive collects through wearables, published the open ego500 dataset in August 2026, and raised an 8.2 million dollar seed in May 2026. Mecka AI hosts the ten-lab Egoverse consortium and publishes its aggregate totals, raising a reported 60 million dollar Series A.

In-facility capture

MicroAGI exited stealth in July 2026 with a 55 million dollar seed round, reported as Germany's largest, running its Atlas system to capture manual work inside client factories, with five companies collecting at announcement. Build AI's head-mounted model also operates on factory floors.

Capture inside a working facility is an operations problem before it is a data problem: site access, worker consent, privacy constraints, and recording that does not interrupt production. Ask who holds the consent paperwork and whether it survives a later change in how you use the data, because facility capture produces the most legally entangled data in this category.

Teleoperation specialists

Adamo provides managed teleoperation with sub-40-millisecond latency, which matters for dexterous tasks where operator feedback delay degrades demonstration quality. RoboStream runs end-to-end robotics capture and annotation. Cortex AI collects workplace robot and egocentric datasets using field collectors and human-in-the-loop operators.

All three are seed-stage and narrow. For the specific capture they run, that focus is an advantage. For a multi-year program, ask about runway.

Broad vendors with a robotics line

micro1 lists robotics alongside its text and code expert network. Encord, Scale AI, iMerit, Sama, and Appen all annotate sensor and 3D data competently, and Encord now operates a physical AI collection facility in San Leandro, California, running leader-follower teleoperation and egocentric pilots. For annotation throughput on data you have already collected, these are the deeper benches. For producing data that does not exist, they are mostly the wrong shortlist.

Comparing the options

ProviderCapture mechanismOpen dataStage
HumanSignalStudio, on site, and field, synchronized multimodalNone publishedGrowth
Build AIEgocentric wearable, factory floorEgocentric-100K, 100,405 hoursSeed
XDOFTeleoperation stationsABC-130k, Apache 2.0, with paperSeed
Human ArchiveWearablesego500Seed
Mecka AIConsortium aggregationEgoverse totalsSeed
MicroAGIIn-facility, Atlas systemNone publishedSeed
AdamoManaged teleoperation, sub-40msNone publishedSeed
RoboStreamEnd-to-end capture and annotationNone publishedSeed
Cortex AIField and human-in-the-loopNone publishedSeed

How to choose

Ask what the provider has actually shipped, not what it describes. A Hugging Face organization, a GitHub repository, or a paper takes ten minutes to check and distinguishes companies that have run the operation from companies that have announced it.

Ask whether the capture mechanism matches your model. Egocentric video, teleoperated manipulation, and in-facility observation produce different data, and a provider strong at one is not automatically competent at another.

Ask about consent scope and durability, especially for facility capture. The question is not whether consent was obtained but whether it covers the uses you will want in eighteen months.

Why HumanSignal is the better choice for protocol-designed capture

Most of the field above is organized around one capture mechanism each: a wearable, a teleoperation station, a factory deployment. That focus works when your need matches the mechanism. It becomes a constraint when your model needs synchronized streams the provider does not run, or a distribution their standing operation does not produce.

HumanSignal Services starts from the gap rather than from the rig. We write the collection protocol, recruit for the physical or professional characteristics it requires, and run the capture in our studios, on site at your facility, or in the field with motion, vision, audio, and sensor streams synchronized where the model needs them together. Control over the distribution is the point: which tasks, which conditions, which failure modes, how many repetitions, and which edge cases are deliberately included rather than hoped for.

Review workflows run in Label Studio Enterprise, the commercial edition of the open source labeling tool we build, so agreement metrics and reviewer decisions are reportable to you while the capture runs. We publish contributor pay ranges. Our customer agreement gives you exclusive ownership of customer data and of insights produced specifically for you, which matters more in robotics than in most modalities because the data is expensive to reproduce. We hold SOC 2 Type 2 and HIPAA compliance, documented on our security page.

The capability came in through the acquisition of Erud AI in November 2025. The physical AI data services page sets out how a program gets scoped.

How much egocentric video does a robot policy need

Hour counts are a weak proxy. What governs transfer is coverage of the task distribution, consistency of calibration across the corpus, and whether failure and recovery states appear. A smaller corpus matched to your deployment conditions generally beats a larger mismatched one, which is why inspecting an open release tells you more than comparing totals.

Why are so many robotics data providers seed-stage

The segment is roughly two years old and formed around the shift from scraped data to performed data. The practical consequence for a buyer is vendor-continuity risk on long programs: ask about runway and about what happens to your captured data and consent records if the company is acquired.

Can annotation vendors handle robotics data

For annotation, yes. Encord, Scale AI, iMerit, and Sama all handle LiDAR, point clouds, and multi-camera video, and several have added physical AI lines. The distinction is production: most annotate data you collected rather than producing data that does not exist. Encord is the clearest exception, operating its own collection facility since 2026.

Should I license an existing robotics dataset or commission capture

License when an open or commercial corpus covers your deployment distribution, since it is faster and cheaper, and several providers here publish open releases you can evaluate at no cost. Commission when your model's advantage depends on data nobody else has. Most programs do both.

Related Content