ohmlot-collector
The collector behind ohmlot.com — a daily record of what personal
robots (humanoids, robot dogs, desktop robots) cost in the US, whether they are actually in stock,
and what their FCC authorization status is.
Every snapshot is committed to this repository the day it is taken, by GitHub Actions. That is the
point of the repo being public: the site alone cannot prove that a price history is real, because a
JSON file can be written at any time with any date in it. Commit timestamps cannot be backdated, so
the git log is the evidence.
Collection started 2026-09-22.
What gets collected
| What | Where it lands | How often |
|---|
| Manufacturer list prices, per variant, pinned to the US market | data/prices/YYYY-MM-DD.jsonl | daily |
| Raw store responses, kept verbatim | data/prices/raw/<date>/ | daily |
| A 6-hourly sample of three high-churn sources | data/prices/intraday/ | every 6h |
| FCC conditional-approval records | data/fcc/ | weekly |
| Changes derived by diffing consecutive snapshots | data/events.jsonl | daily |
| Catalog of 32 robots, 15 of them price-tracked | data/models.json | manual |
The 17 robots that are not price-tracked are real, currently-sold products whose price is
quote-only or behind a contact-sales form. They are listed with the reason rather than left blank or
dropped.
How it works
Prices (src/prices.ts) read each store's own product endpoint — Shopify .js/.json for most
sellers, an HTML parse for 1X — with ?country=US so the market is pinned. Collection is per
variant, so a product with 15 variants produces 15 rows, each with its own price, availability and
SKU. Sources are configured in config/price-sources.json, not hardcoded.
FCC (src/fcc.ts) reads the PSHSB public notices listed in config/fcc-notices.json as PDFs
from docs.fcc.gov, keeps the PDF bytes verbatim under data/fcc/raw/, and extracts all three
columns of Appendix B using a coordinate-based reader (src/lib/pdf.ts) that works off each text
item's x/y position. The flattened .txt rendering that FCC also publishes merges the columns
irrecoverably, which is why the PDF is parsed directly.
apps.fcc.gov — the host that serves the equipment-authorization API — is Akamai-blocked from CI
runners and from every other origin tried. That is recorded in context/fcc-api-notes.md and no
workaround was used: no proxy service, no scraping-bypass tool. The result is that this repo has
less FCC data than a bypass would produce, and says so.
Events (src/events.ts) are generated by comparing each day's snapshot to the previous one.
Nothing is hand-entered.
Rules this repo follows
These are not aspirations; they are why several things here are missing or marked unknown.
- Never invent a value. If something cannot be read, it is
null or unknown. There is no
"reasonable estimate" anywhere in data/.
- Separate instrumentation from the world. When a field is observed for the first time, that is
recorded as
observation_started, not as a change. Those events carry observation_change: true
and are shown separately on the site. The first 20 events in this repo are all of that kind.
- Separate observation from reconstruction. Prices this collector saw are
primary_observation. Three changes found in the Internet Archive from before collection started
are kept in data/historical-findings.jsonl and deliberately never merged into events.jsonl.
- Don't route around a block. An official source that refuses is logged and skipped.
- Prices with weak provenance are not displayed as numbers. A price taken from a press release
or an unverified vendor page is excluded from the site's filters, sorting and comparisons.
What this is not
- These are manufacturer list prices, not transaction prices. This records what a store is
asking, not what anyone paid.
- The history is short. It starts 2026-09-22 and that is stated on the site rather than hidden.
- There is no marketplace, no listings and no affiliate links.
- FCC coverage is limited to what the public notices contain. New IDs need manual verification on
apps.fcc.gov.
Why it exists
A Wayback Machine study (context/recoverability.md, run 2026-09-25) measured whether this data
could just be recovered later instead of collected. The finding was the opposite of the starting
assumption: where the Internet Archive has a capture, most fields come back fine — price recovered
in 26 of 26 sampled snapshots, stock in 23 of 26.
The gap is coverage, not readability. DEEP Robotics' Lite3 product page has never been crawled
by the Internet Archive, not once, while two other products from the same store were. KEYi's Loona
product URL has never been captured either; its store appears in the archive only inside a five-week
window in May–June 2026 — and inside that narrow window its price changed three times. For some of
these products there is no "look it up later."
Running it
npm install
npm test
npm run collect:prices # same as the daily workflow
npm run collect:fcc # same as the weekly workflow
npm run collect:intraday # the 6-hourly sample
Failed sources are written to data/logs/<date>.log; a failure never silently produces a row.
That directory does not exist in a clean checkout, because it is only created when something
fails — which is also how a CI bug once went unnoticed for three days (git add on a path that
only exists on error exits non-zero, and silently skipped the commit step after it).
Scheduling lives in .github/workflows/ (daily.yml, fcc.yml, intraday.yml). Note that the
cron times in those files are requests, not start times. GitHub's scheduler delays these runs,
and the delay is neither small nor predictable: measured starts have landed between roughly two and
six hours after the requested time, varying run to run and workflow to workflow. Changing the cron
value does not help, because the delay is queue-side. This is why the site displays the collection
time read from the data itself rather than the time the workflow asked for.
The site in web/ is a Next.js app that reads data/ at build time.
Sources and license
Model metadata draws on Robot Price Index and humanoidrecord.com, both CC BY 4.0; attribution is in
data/SOURCES.md. Prices come from manufacturer and official-reseller stores. FCC material is US
government public record.
Ohmlot is independent and not affiliated with any manufacturer or store.
Working notes in context/ are written in Korean. They are the actual instructions the work was
done from, including the ones that record mistakes and corrections.
