Read the website of any robot data company, and you will see the same list. Warehouse. Retail. Healthcare. Manufacturing. Logistics.
Very few of them can point to a completed capture program in any of those.
This post is about how to check. It is the fifth post in this series on the robot training data market. The first post covered the companies in the market.
Why vertical experience does not transfer
It is tempting to assume that a team that can record a kitchen can record a warehouse. They cannot, and the reasons have almost nothing to do with cameras.
Access. Getting permission to record inside an operating hospital, a food processing plant or a distribution centre is a months-long process involving legal, operations, safety and sometimes a union. Getting into someone’s kitchen takes an afternoon.
Consent. Recording workers means consent from every person in frame, plus a process for people who decline, plus handling of anyone who walks through unexpectedly. In healthcare, add patients. This has to be documented before you arrive, not after.
Safety and compliance. Sites have rules about what equipment can be installed, where, and how it is powered. A camera rig in a food facility has to meet requirements a camera rig in an office does not.
Environment. Warehouse lighting, refrigerated storage, outdoor agriculture, sterile clinical spaces. Each one breaks a different part of a capture setup.
Knowing what matters. This is the one people underestimate. If you do not know the work, you do not know which activities are worth recording. You come back with a hundred hours of the wrong thing.
None of this is transferable from one vertical to another. It has to be built each time.
How to check a claim
Grade every vendor’s vertical claim on this ladder. Strongest evidence first.
- A published dataset from that vertical, that you can inspect.
- A named customer in that vertical who has said so publicly.
- A named site or partner where capture actually took place.
- A described program with detail about scale, duration and method.
- A word on a website.
Most vertical claims in this market are level 5. Ask which level any specific claim is at. It is a fair question and a vendor with real programs will answer it easily.
The same test is applied to DreamVu below.
Coverage across the market
Manufacturing and industrial. The best-served vertical. Config is backed by the venture arms of Korea’s largest manufacturers, which gives it a defined industrial context. Lightwheel names manufacturing among its customer applications and has named AgiBot, Geely and BYD among the companies using its assets. Build AI has released roughly a million hours of open first-person footage recorded by factory workers across Southeast Asian industrial sites.
Automotive. Well served, mostly as a legacy of autonomous vehicle work. Scale AI’s robotics operation grew out of its AV data business. Lightwheel has named automotive partners and has an announced partnership to bring its simulation technology into the Japanese automotive market.
Warehouse and logistics. XDOF describes teleoperation capture in warehouse-scale environments. Scale AI works across logistics. Reasonably served.
Domestic and residential. Crowded, because access is easy. Human Archive deploys rigs in residential settings. Luel’s contributor network covers domestic environments. Most crowdsourced egocentric footage is domestic.
Retail. Thinner than it looks. Retail requires store access, customer consent and cooperation from an operator whose priority is trading, not recording.
Healthcare. Poorly served, because it is the hardest. Patient consent, clinical protocol and infection control all apply before anyone discusses cameras. Lightwheel has announced a healthcare partnership focused on perioperative workflows, which is a serious move into the vertical.
Agriculture. Poorly served. Outdoor, seasonal, geographically dispersed, and the equipment has to survive conditions most rigs are not built for.
Food service. Some coverage through domestic and commercial kitchen capture. Variable.
Where the gaps are
The pattern is clear. Verticals with easy access are crowded. Verticals with regulatory friction, consent complexity or difficult physical conditions are thin.
Healthcare and agriculture are the two clearest gaps. Retail is thinner than the marketing suggests.
This is not because those verticals are less valuable. It is because they are harder to get into, and a company built around handing out consumer devices cannot get into them at all.
Where DreamVu sits
Vertical range is the part of DreamVu’s offering that needs the least argument, so this section is short.
DreamVu has run capture programs in retail, warehousing and logistics, industrial manufacturing, and commercial services. It operates in both the United States and India, which required consent frameworks, PII removal processes and liability structures that work in both jurisdictions.
DreamVu has published a retail dataset and a paper benchmarking against a world model, which puts its retail claim at level 1 on the ladder above. Its manufacturing and commercial services programs are level 3, run at named partner sites.
The trade-off is depth against breadth. A vendor who does nothing but factory floors will have more factory floor hours. For a deployment confined to one vertical where a specialist exists, the specialist is the better call. Range matters when the model has to work in more than one kind of place.
DreamVu does not currently claim healthcare or agriculture coverage. Both are identified above as market gaps, and neither is a gap DreamVu has filled.
Frequently asked questions
Why does industry experience matter when buying robot training data?
Because access, consent, safety compliance and environmental conditions differ completely between industries, and none of it transfers. A vendor who has recorded in homes has not solved the problem of getting equipment into an operating theatre or a refrigerated warehouse. Knowing which activities are worth recording also requires familiarity with the work.
How do I verify a data vendor’s industry claims?
Ask which level of evidence the claim sits at. A published dataset you can inspect is strongest, then a named public customer, then a named capture site or partner, then a described program with real detail, then an unsupported website claim. Most vertical claims in this market are unsupported website claims.
Which industries are best served by existing robot data vendors?
Manufacturing and industrial are the best covered, followed by automotive, warehouse and logistics, and domestic settings. Domestic is crowded because access is simple.
Which industries are hardest to get robot training data for?
Healthcare and agriculture. Healthcare requires patient consent, clinical protocol compliance and infection control before capture can begin. Agriculture is outdoor, seasonal and geographically spread, with conditions most equipment is not built for. Retail is also thinner than vendor marketing suggests.
Which vendors have worked across multiple industries?
Lightwheel names manufacturing, automotive, healthcare, logistics and agriculture among its applications and has named enterprise customers. Scale AI covers autonomous vehicles and general robotics. DreamVu has run capture programs in retail, warehousing and logistics, industrial manufacturing, and commercial services across the United States and India.
