Physical AI training data providers: how to tell them apart
Every published ranking of physical AI data providers is written by one of the providers. That is not a complaint about anyone's integrity, it is a description of the available information: the category is roughly two years old, no analyst covers it, and the comparison pages that exist were built by vendors with a position in the ranking. The result is that a team scoping its first collection program has no neutral way to tell a company with a capture facility from a company with a landing page about capture facilities. This piece organizes the field by the thing you actually need, names who does what, and flags where marketing and output have come apart.
Key takeaways
- Sort providers by the capture mechanism you need, not by a general ranking. Egocentric video, teleoperation, in-facility capture, and sensor annotation are four different operations.
- Open datasets are the strongest available quality signal in this category, because you can inspect distribution and labeling yourself before committing. Several well-funded providers publish none.
- Check published output against marketed positioning. A provider's Hugging Face and GitHub organizations show what it has actually shipped, which does not always match the website.
- The segment is young and consolidating, so verify what a provider operates now rather than what it announced.
Egocentric video is the most contested segment
First-person video of people performing tasks is the foundational dataset type for imitation learning, and it has become the most competitive segment.
Build AI has published the largest open corpus. Egocentric-100K, released in December 2025, comprises 100,405 hours from 14,228 workers and 10.8 billion frames, captured through a company-built head-mounted camera worn by workers during shifts. It followed Egocentric-10K in November 2025. Human Archive collects through wearables using a gig-worker model in India, published the open ego500 dataset in August 2026, and raised an 8.2 million dollar seed round in May 2026. Mecka AI hosts the Egoverse consortium, a ten-lab collaboration whose aggregate totals it reports. Encord runs egocentric pilots alongside teleoperation at its San Leandro facility.
What to check: the activity distribution, not the hour count. A hundred thousand hours of warehouse picking is a narrow corpus despite the number, and the open releases let you verify this before you buy. Also check camera and calibration consistency, because mixed rigs across a corpus limit what can be learned from it.
Teleoperation data is limited by the rig that captured it
Contact-rich manipulation data has to be performed, and the rig determines what is recoverable from it.
XDOF operates teleoperation stations for manipulation capture, launched publicly in June 2026 with ABC-130k released under Apache 2.0, a companion arXiv paper with UC Berkeley faculty co-authors, and reported funding of roughly 70 million dollars across about 20 customers including several frontier labs. Adamo offers managed teleoperation with sub-40-millisecond latency, which matters for dexterous tasks where operator feedback delay degrades demonstration quality. Encord runs leader-follower arm teleoperation at San Leandro, and a July 2026 TechCrunch profile described pilots extending to brain-wave sensors with Zander Labs and forearm EMG. RoboStream and Cortex AI both run capture-and-annotate operations in this space.
What to check: whether force and tactile channels are recorded, the control latency, and whether failure cases are captured deliberately or discarded. Teleoperation corpora skew heavily toward successful demonstrations, which is a known limitation for policies that need to recover from error states.
Facility capture is an operations problem before it is a data problem
Collecting inside an operating factory, warehouse, or clinic is an operations problem before it is a data problem, involving site access, worker consent, union and privacy constraints, and capture that does not interrupt production.
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 and one preparing a factory deployment at announcement. Build AI's head-mounted camera model runs on factory floors. HumanSignal Services runs collection in our own studios, on site at customer facilities, and in the field, with synchronized motion, vision, audio, and sensor capture.
What to check: who holds the consent and release paperwork, and whether it survives a later change in how you use the data. Facility capture generates the most legally entangled data in this category, and the entanglement surfaces during diligence rather than during collection.
Sensor annotation quality depends on your exact configuration
If you are already collecting, the constraint is usually annotation throughput on LiDAR, point clouds, multi-camera video, and fused sensor streams.
Encord provides native annotation for video, LiDAR, audio, text, and sensor fusion, with embedding-based curation for finding distribution gaps, and holds a 4.8 out of 5 rating on G2 across 65 reviews as of September 2026, which is the deepest third-party review base in this set. Scale AI annotates 3D and LiDAR within its Data Engine and holds DoD IL4 and FedRAMP High authorizations, which is decisive for defense-adjacent robotics work. Appen added a physical AI pillar covering LiDAR annotation, robotics trajectories, and sensor fusion. iMerit, now part of EXL Service following an August 2026 acquisition, brings in-house subject-matter experts to autonomous vehicle and geospatial annotation. Sama annotates computer vision and sensor data for automotive programs.
What to check: whether the vendor annotates your sensor configuration specifically. Point-cloud annotation quality varies sharply by sensor type and mounting geometry, and general 3D experience does not transfer automatically.
A short decision path
| If you need | Start with | Verify |
| Egocentric video at volume | Build AI, Human Archive, Mecka AI | Activity distribution, rig consistency |
| Manipulation and contact-rich data | XDOF, Adamo, Encord | Force and tactile channels, failure coverage |
| Capture inside a working facility | MicroAGI, HumanSignal, Build AI | Consent scope and durability |
| Sensor annotation at throughput | Encord, Scale AI, iMerit, Appen | Your exact sensor configuration |
| Defense or government deployment | Scale AI | FedRAMP and IL4 scope for your program |
| A bespoke protocol end to end | HumanSignal, Shaip, Encord | Prior delivery of your specific capture type |
The column that matters most is the third one. This segment is young enough that funding announcements and facility photographs are widely available while evidence of delivered datasets is not, and the gap between the two is where a procurement goes wrong. Every verification in that column can be done before a first call, from open dataset releases, published papers, consent documentation, and the provider's own description of its sensor stack. Weight that published evidence more heavily than funding totals or facility tours, because it is the part of the record that reflects delivery rather than intent.
Working with HumanSignal
We run collection in our own studios, on site, and in the field for physical AI and robotics programs, with quality workflows running in Label Studio Enterprise so reviewer decisions and agreement metrics stay visible to you. We launched the services line in November 2025 through the acquisition of Erud AI, which means we have less public track record on services engagements than the incumbents in this piece, and we would rather you weigh that openly than discover it later. If you are scoping a capture program, we can walk through protocol design and acceptance criteria before any commitment.
How much egocentric video does a robot policy actually need
There is no established threshold, and hour counts are a weak proxy. What governs transfer is coverage of the task distribution, consistency of calibration, and whether failure and recovery states appear. A smaller corpus matched to your deployment conditions generally outperforms a larger mismatched one, which is why inspecting an open release is more informative than comparing totals.
Why do so few physical AI data vendors publish pricing
Collection cost is driven by protocol design, site access, rig configuration, and session count, which vary enough per engagement that a rate card would be misleading. Expect quote-based pricing across this category and ask for the cost drivers rather than a unit rate.
Is an open dataset from a vendor a reliable quality signal
It is the best one available in this category, with a caveat. An open release lets you inspect distribution, calibration, and labeling yourself rather than taking a claim on trust. The caveat is that a vendor's open release may be its best work rather than its typical work, so treat it as an upper bound on quality and confirm with a paid pilot on your own task.
Should I license existing robotics data or commission collection
License when an existing corpus genuinely covers your deployment distribution, because it is faster and cheaper. Commission when your model's advantage depends on data nobody else has, since a licensed corpus is available to every competitor willing to pay for it. Most programs do both, licensing a base and commissioning for the gap.