The Robot-Data Trade Just Got a Street Address

Human Archive is a San Francisco company that pays people to wear head-mounted camera rigs through ordinary shifts — cooking, assembling, stocking — and sells the resulting footage to teams training robot foundation models. Consider a factory worker clipping on a headset before a shift in a manufacturing plant, or a home cook wearing one while prepping dinner. Both feed the same pipeline that eventually teaches a robot arm how to grip a spatula or place a part on a conveyor.

That company is real, it's funded, and it just scaled past a threshold worth noting. At the same time, a separate arXiv paper on retail video just delivered an inconvenient finding for anyone assuming more footage automatically means a better model. Together, they point at the same lesson field data teams have known for years: the camera's point of view is a design decision, not an afterthought.

This matters beyond robotics labs. Market researchers running in-store audits, GIS surveys, and mystery visits are wrestling with the identical question — whose eyes should the camera represent? — and the answer changes what a capture program should actually look like.

A Startup Just Put Cameras on 1,000 Heads

Human Archive came out of Y Combinator's Winter 2026 batch and raised an $8.2 million seed round led by Wing Venture Capital and NVP Capital. Human Archive is a Y Combinator Winter 2026 company based in San Francisco that raised an $8.2M seed led by Wing Venture Capital and NVP Capital. The scale is the part worth sitting with. 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.

The company isn't just recording video. It combines egocentric video with depth, motion capture and tactile sensing into synchronized datasets, then runs its own QA, anonymization and annotation. That's a full field-operations stack — recruitment, hardware, sync, cleaning, labeling — built specifically because robot-model teams don't want raw footage. They want structured, verified, multimodal episodes they can drop straight into a training pipeline.

Why does this matter outside robotics? Because it confirms something we've argued before: the bottleneck in physical AI was never cameras, it's operations. A logistics company wearing an AI startup's branding is exactly what Human Archive looks like from the outside — a recruitment and QA operation that happens to sell training data.

Why a Supermarket Dataset Just Told Robotics Teams They're Filming Wrong

Volume is not the whole story, and a new paper on retail video proves it with hard numbers. Researchers introduced RetailSMV, a corpus of 32,105 captioned retail clips from five supermarkets with synchronized ego/exo capture from the store-staff perspective — stocking, arranging, weighing, managing supply carts, scanning at checkout — rather than the customer-centric framing of prior retail video corpora. They then trained matched model variants on egocentric-only, exocentric-only, and combined footage to see which viewpoint actually produced the strongest results.

The finding cuts against intuition. On a 200-clip held-out test set evaluated with seven complementary metrics under a strict paired statistical protocol, exocentric-only adaptation matched or exceeded combined adaptation on six of seven point estimates, despite training on only 15,985 exocentric clips versus 32,105 for the combined set. Half the data, equal or better performance. Even more telling: adding exocentric data to egocentric-only training helped, while adding egocentric data to exocentric-only training hurt.

Egocentric video refers to footage captured from a first-person, head-mounted point of view, as distinct from "exocentric" footage shot from a fixed external camera watching the scene. For robotics and for field research alike, the assumption has been that first-person always wins because it mimics the agent's own perspective. RetailSMV says: not always, and not automatically. The camera's job is to answer the specific question a model — or a client — is asking. Point it at the wrong thing and extra hours just add noise.

Market Researchers Have Been Living This Problem for Years

Swap "robot foundation model" for "brand insight" and the same tension shows up in market research. Greenbook's 2026 GRIT Insights Practice Report found mid-size research firms — 101 to 500 employees — now lead the industry in revenue growth, capability expansion, and AI governance maturity, while the largest firms are three years into a deliberate exit from fieldwork toward consulting and analytics. Fieldwork itself — the unglamorous business of collecting clean, structured, real-world observation — is becoming the differentiator precisely as the big shops walk away from it.

And the quality problem is intensifying, not shrinking. Greenbook's GRIT research found that data quality concerns have surged 40% year-over-year, driven in large part by synthetic respondents and rising survey fatigue among younger participants. At the same time, adoption of AI in research delivery has crossed a threshold: Greenbook's GRIT Business Outlook research found that 67% of research suppliers now embed generative AI directly into client deliverables, automating tasks like survey design and cross-tab analysis rather than using it as a side tool. Faster output, same trust problem. Greenbook's own coverage of emerging tools shows what that automation looks like in practice — tools like Scalafai now automate survey programming, fielding, weighting, statistical testing, and even report creation, with the goal of freeing researchers to focus on interpretation rather than production work.

None of that automation fixes a bad capture design. If your in-store audit points a camera at the wrong shelf, at the wrong angle, on the wrong day, no amount of AI cross-tabbing rescues the finding. That's the exact lesson RetailSMV just delivered to robotics, except retailers and CPG brands have been paying that tax for decades.

Where the Vendors Stack Up — And Where They Stop

It's worth being honest about limitations here, including our own.

Human Archive — best for teams that need multimodal, in-home and light-industrial footage at volunteer scale, fast. Where it stops: it's a young company. By its own reviewer's assessment, "it is early," and enterprise buyers should expect the durability and consistency questions that come with any one-year-old capture network.

Scraped internet video — best for nothing you'd want to ship in production. The constraint is rights: scraped footage is a legal liability, and commissioned capture is slow and expensive per hour. Cheap up front, expensive the moment legal gets involved.

Purpose-built field programs (our lane) — best for research questions with a defined viewpoint requirement: does the auditor need to see what the customer sees, or what the associate sees? Where it stops: we don't have a shelf of a million pre-recorded hours sitting in a warehouse. You're commissioning capture built around your question, which means timelines follow the research design — not a marketplace catalog you can browse and ship same-day.

So What Does This Mean If You're Planning a Capture Program?

Ask the viewpoint question before you ask the volume question. RetailSMV shows that egocentric footage isn't automatically superior — it depends entirely on what the model, or the analyst, needs to learn. We've made a version of this argument before in our piece on what a robot dataset teaches fieldwork: the format has to match the downstream use, not just the easiest thing to film.

Then ask the quality question, because Greenbook's numbers say that's where programs actually fail. A pile of unverified footage — robotic or human-collected — is a liability the moment someone asks how it was gathered, cleaned, and checked. That's operations work, not a camera spec.

Frequently Asked Questions