The Associated Press moved a wire on Tuesday that ran through Wednesday and Thursday on AP-syndicated U.S. local papers and into Europe via Euronews on May 14. The lede is a hospitality worker named David Park at the five-star Lotte Hotel Seoul, in his ninth year on the food-and-beverages team, folding banquet napkins — head-cam, chest-cam, two glove-cams. Each motion is fed into a database. The database is the AI brain of a humanoid robot. The humanoid is intended for industrial deployment by 2028.
「I’ve been doing this about once a month,」 Park told Fortune on May 12, apparently without visible distress. The startup running the pipeline is RLWRLD (pronounced 「real world」), Seoul-based, $41 million in total seed funding, closed a $26 million Seed 2 in February — and the cap table is the part of this story that explains why the napkin matters.
The cap table tells you the deployment plan
The Seed 2 round, led by Headline Asia and Z Venture Capital, includes CJ Logistics, Lotte Ventures, Kakao Investment, Hanwha Asset Management, Hyosung Ventures, Smilegate Investment, and T Investment. The four Korean conglomerate-affiliated funds are not strangers — they are the strategic owners of, respectively, Korea’s largest logistics network (CJ), the hotel and retail brand in whose ballroom Park is currently miked up (Lotte), the convenience-store and last-mile fulfilment operator (Hanwha-adjacent), and the industrial systems group (Hyosung). The training sites in the AP piece — Lotte Hotel Seoul, CJ Group warehouses, Lawson convenience stores — are not independent customers. They are the investors’ own facilities, providing the training data RLWRLD needs to ship the foundation model that will then automate the investors’ own labour.
This is the closed loop:
| Stage | Site | Subject | Cameras | What is captured |
|---|---|---|---|---|
| Hospitality | Lotte Hotel Seoul (F&B) | ~10 workers incl. David Park | Head + chest + 2x hand | Napkin folds, glass polishing, table setting, finger joint angles, grip force |
| Logistics | CJ Group warehouses | Warehouse workers | Head + chest + hand-cams | Lifting, gripping, sorting packages |
| Retail | Lawson stores (Japan) | Convenience-store staff | Head + hand-cams | Stocking shelves, organising food displays, handling goods |
| Government | National 「Master Technician」 program | Senior factory technicians | Per Euronews | 「Instinctive know-how and skills」 of master technicians, ₩45B (~$33M USD) project |
The hospitality session captures the precision motions a parallel-jaw gripper cannot do. The logistics session captures the load-handling vocabulary. The retail session captures the small-package dexterity. The government project captures the manufacturing tacit knowledge that nobody writes down — and underwrites the whole stack with public money.
What RLWRLD’s executive said about the gripper problem
The most-quoted line in the AP wire and the Euronews follow-up is from Hyemin Cho, RLWRLD’s business-and-strategy executive:
「If you were to have a robot fold napkins, a gripper wouldn’t be able to achieve the precise and crisp folds expected of hotel service quality.」
That sentence is the one-paragraph thesis for the physical-AI category. The five-fingered hand is the part of the humanoid form factor that is most expensive, most fragile, most under-developed — and the part the previous generation of warehouse robots (Kiva, Locus, Symbotic) deliberately avoided by redesigning the warehouse around suction grippers and conveyors. RLWRLD is doing the opposite: training the hand to fit the world, not the world to fit the gripper. The training input is human hands.
A few notes on the recording rig per the wire:
- Three or four body cameras per session — head-mounted POV, chest-mounted reference frame, two wrist-mounted hand cameras.
- Data captured includes: finger positioning, joint angles, applied force at each contact point — not just video, but instrumented motion capture at hand-resolution.
- Cadence: 「about once a month」 per worker (Park), running for at least the past several months.
- Goal: industrial deployment of the foundation model around 2028, expanding to home robots after that.
There is no opt-out language in the AP wire. There is no compensation figure in the AP wire. There is no union representation referenced in the AP wire.
The Korean government’s $33M parallel database
The smaller and quieter half of the story is that the Korean government — per the Manila Times re-run of the AP wire on May 13 — recently announced a $33 million national project aimed at recording the 「instinctive know-how and skills」 of experienced technicians, to help train AI-powered manufacturing robots.
This is the Korean industrial-policy answer to the same problem Hyundai is bringing Boston Dynamics’ Atlas humanoid into its Georgia plant for in 2028, the same problem Schaeffler signed a four-digit-unit Humanoid deal for last week, the same problem STMicroelectronics signed Oversonic’s RoBee for in its European fabs in March, and the same problem Texas Instruments signed UBTECH’s Walker S2 for at its Sherman, Texas SM1 mega-fab in January — namely, that the last generation of industrial workers carry tacit physical knowledge that has not been encoded in any manual, and that the next generation of industrial robots cannot do the job without it. The Korean government has decided this know-how is national-strategic enough to fund a public database. The startup running its private version, against the same physical-AI roadmap, is RLWRLD.
The political-economy framing is direct: the worker’s hand motions become the public-and-private property of the system trying to replace the worker.
The 2028 number is what to argue about
Several reference points on whether 2028 is real, optimistic, or marketing:
- Boston Dynamics’ Atlas is targeting 30,000 units/year by 2028 from its Savannah, Georgia plant.
- Figure AI is targeting 100,000 Figure 03 units over four years from its BotQ facility, with mass-production beginning end-2026.
- 1X is shipping NEO consumer units in late-2026.
- Schaeffler-Humanoid targets a four-digit deployment by 2032, with initial production from December 2026.
- RLWRLD’s 2028 target sits in the middle of that cohort, and is the software-foundation-model side of the bet — not the unit-shipment side. The 2028 deadline is when the model is supposed to be deployable across multiple OEMs’ robot bodies, not when a single OEM ships its own robot.
The bet is reasonable on paper. The data-collection methodology is reasonable on paper. The five-fingered-hand-instead-of-redesigning-the-warehouse-around-grippers thesis is the most credible physical-AI thesis of 2026 because it is the one no incumbent has industrially solved.
The thing that is being deliberately not said in the AP wire is what David Park — who has folded a banquet napkin for nine years and is on the recording schedule about once a month — is supposed to do for work in 2029.
What to watch
- RLWRLD’s robotics foundation model release. The company said in its Seed 2 press release it plans to officially unveil the model in H1 2026 — i.e., the next six weeks. The benchmark question is whether it ships against generalist physical-AI scores like the Roboflow physical-AI benchmark or remains a CJ-Lotte-Lawson-specific stack.
- Compensation disclosure for the recorded workers. The AP wire does not say what David Park, the CJ warehouse workers, or the Lawson staff are paid for the recording sessions, on top of their hourly wage. If the answer becomes public, the labour-economics story becomes the much louder one.
- The MOTIE 「master technician」 database tender. Whether the $33M Korean government program is run by RLWRLD, an open consortium, or a Hyundai-Samsung joint venture is the cleanest single tell on whether RLWRLD has won the home market or is just the highest-profile entrant.
- Hyundai’s Boston Dynamics Atlas Georgia rollout. Plant target: 2028. If Hyundai’s Korea-grown competitor RLWRLD ships its model first, the most-watched 2028 datapoint in physical AI is the model layer, not the hardware layer.
The four-camera napkin-folding session at Lotte Hotel Seoul is the cleanest image yet for what the physical-AI training cycle of 2026 looks like in practice: a hospitality worker, instrumented at hand-resolution, doing his job correctly enough that the foundation model can do it without him. The same pipeline at CJ Logistics. The same pipeline at Lawson. And a government cheque underwriting the version where it works on a factory floor.
Per the wire, the next recording session is scheduled for next month.