Why a $22M Robot-Data Startup Is Really a Logistics Company

Egocentric data collection means recording video and sensor data from a first-person point of view, usually through a head-mounted or chest-mounted camera, so a robot or AI model can learn how a human hand actually reaches, grips, and adjusts mid-task. Picture a gig worker in Ho Chi Minh City strapping on a small headset before her shift, walking a crowded market stall, chopping vegetables, restacking boxes. Every motion gets streamed off to a training pipeline on the other side of the world.

That's roughly the pitch behind Ropedia, a startup that just closed a funding round to scale exactly that kind of capture. Its device, called HOMIE, is a head-mounted device, equipped with a 360º camera view, that captures first-person data for training robots , according to The Robot Report. The company says it plans to keep expanding, and continue expanding data collection in Southeast Asia and North America, growing its U.S. team, and increasing manufacturing of HOMIE .

Here's the part worth sitting with. The money is chasing hardware. But the industry's real constraint keeps showing up somewhere else entirely: finding enough different people, in enough different places, doing the task correctly, with consent documented and quality checked. That's a field operations problem. It's one market researchers have been solving since long before anyone called it "physical AI."

The Money Chasing First-Person Footage

Ropedia isn't alone in betting that first-person capture hardware is where the next round of robotics infrastructure gets built. The company's device straps a wide-angle camera to a person's head and records what they see and do, frame by frame, hands included. According to The Robot Report, the company also plans to collaborate with upstream and downstream partners and is interested in working with industry, academia, and even robotics hobbyists . It's also not stopping at one device. Looking ahead, Ropedia plans to create more wearable data capture hardware , per the company's founder.

That's a sensible bet on paper. Robots learn manipulation faster from footage that matches the camera angle they'll actually use in the field — which is exactly why first-person view has become the default format for training data, a shift we broke down in How Egocentric Video Is Collected. But a camera is a commodity. What isn't a commodity is getting a market vendor in Hanoi, a warehouse picker in Ohio, and a home cook in São Paulo all wearing one correctly, consistently, with signed consent, on the same day.

What "Egocentric Data Collection" Actually Means

Egocentric data collection refers to capturing video, audio, and motion signals from the wearer's own viewpoint rather than an external, third-person camera. The distinction matters more than it sounds. A robot trained on footage of a room from a fixed overhead angle never sees what a hand looks like as it closes around a mug. A first-person feed does. As one industry roundup put it plainly: Robots fail in the real world because they train on the wrong view. Third-person cameras show the scene. They miss the hands. They miss the contact. They miss the exact pixel a robot will see when it reaches for a tool.

That gap is the entire reason wearable capture devices exist. It's also why the dataset scaling race has gotten so aggressive over the past year — and why hardware alone hasn't settled it.

The Scale Race: Bigger Numbers, Same Old Bottleneck

Scale has become the headline metric. Build AI's Egocentric-1M is the loudest example: founder Eddy Xu called it "the internet for physical AI." The growth curve behind it is genuinely wild — 10K hours in November 2025, 100K in December, 1M in April 2026 . That's five months to go from a respectable dataset to the largest one anyone has ever released.

But raw hours aren't the whole story. A newer research dataset, EgoVerse, takes a different approach, prioritizing task and demonstrator diversity over sheer volume. Per a 2026 technical report, EgoVerse is a large-scale collaborative egocentric manipulation dataset spanning 1,362 hours across 1,965 tasks, 240 scenes, and 2,087 demonstrators . Fewer hours than Egocentric-1M, but a far wider spread of tasks and people per hour recorded. That's the actual tension in the market right now: bulk footage versus curated variety, and neither one is free to produce at scale.

Some robotics companies are choosing to build their own capture pipelines in-house rather than buy footage or hardware. Apptronik, for instance, opened its newly expanded Robot Park, its flagship data collection and training facility for humanoid robots in Austin, Texas , this past July — a decision we examined in more depth in our piece on the Robot Park boom. Building your own facility solves consistency. It doesn't solve geographic diversity, which is exactly what a wearable-hardware company like Ropedia is chasing instead.

Hardware Is the Easy Part

Here's the uncomfortable truth nobody selling a headset wants to lead with: manufacturing a camera rig is the tractable half of this problem. Recruiting a representative, well-distributed pool of humans, training them to record cleanly, verifying consent across jurisdictions, and catching bad footage before it pollutes a training set — that's the expensive half, and it doesn't scale with a factory line.

It's not unique to robotics, either. Field data collection in agriculture is running into the same wall. The Robot Report noted this month that Agtonomy has added autonomous multi-point turning and enhanced field data collection to its commercial off-road platform — a sign that even outdoor, vehicle-based robotics is now treating structured field capture as a core product feature, not an afterthought. Meanwhile, the sheer volume of footage flowing in has created its own downstream mess, something we've covered in detail in The Egocentric Video Data Glut.

Greenbook's newest industry report backs this up from the market research side. In its 16th edition, the 16th edition of the GRIT Insights Practice Report is here , and it finds the industry has converged on three tasks where agentic AI is already embedded: analyzing data, updating reports, and preparing and integrating data. But AI governance remains a critical gap, and confidence in AI risk management correlates directly with exceeding business goals. Swap "insights teams" for "robotics teams" and the finding holds just as well: adoption is outrunning operational discipline.

What Market Research Already Learned About Scaling Field Work

This isn't a new problem. It's the same one survey and qualitative research shops have wrestled with for decades, just with a smaller budget and less venture funding attached. AI has already rewired how fast that industry can move. Greenbook's 2025 GRIT report found that 72% of insights buyers now use generative AI in at least one stage of a research project, up from 23% in 2023. Sample sizes moved just as fast: the n=200 qualitative ceiling is breaking, with 2026 studies routinely running n=500 to n=2,000 conversational interviews per project.

What made that jump possible wasn't a better microphone or a slicker interface. It was operational infrastructure — recruitment pipelines, quality gates, incentive logistics — built to run at higher volume without falling apart. Robotics data collection is heading down the same road, just a few years behind, and with cameras instead of clipboards.

Picking Your Approach: A Field Guide

So which path actually fits your program? It depends on what you're optimizing for.

• Public datasets (Ego4D, Egocentric-1M, EgoVerse): best for pretraining and ablation studies where public scenes are close enough to your task. Where it stops: you can't control demographic mix, task variety, or embodiment match, and licensing often blocks commercial use.
• Consumer wearable startups (Ropedia's HOMIE and similar): best for fast, low-cost bulk capture across a distributed contributor base. Where it stops: quality control and task-specificity are inconsistent when contributors aren't trained to a shared spec.
• In-house facilities (Robot Park-style builds): best for tight consistency and repeatable protocols under one roof. Where it stops: geographic and demographic diversity are structurally limited by who can physically show up.
• Custom third-party field programs (what TeamPL runs): best for a specific research question — a defined task set, a defined population, verified consent, built-in QA. Where it stops: it's slower to spin up than downloading a public dataset, and it costs more per hour than scraping existing footage. You're paying for precision, not volume.

None of these is universally right. The mistake is picking one because it's trending, rather than because it matches what your model actually needs to learn.

Start a project

TeamPL runs field data collection programs — egocentric video, in-store audits, mystery visits, GIS surveys — designed around a specific research question, not a generic panel. If this topic touches a program you're planning, tell us what you're trying to learn.

Frequently Asked Questions