Methodology

What this product is

Dubsar is a cited research database of humanoid robots and the companies that build them. It is a diligence desk — not a news feed, not a scoring product, and not investment advice. Marketing lives on the public site; the research UI is at /app.

Humanoid scope

We track bipedal or wheeled-humanoid platforms intended for general-purpose or industrial work, plus the companies behind them.

• In scope: full-body humanoids (bipedal or humanoid-form wheeled bases) from the locked research cohort.
• Out of scope: quadrupeds, toy pets, isolated hands or arms sold as components, vacuum or consumer appliances, and “robot” directory padding.
• Roster size: the cited set is a locked cohort of about 30 companies. We do not inflate coverage with watchlist stubs.

Empty is better than fake

Missing height, payload, price, valuation, customer, or patent number stays blank. We do not midpoint ranges, invent FX conversions, or copy aggregator “total raised” widgets. A dash means “not cited here,” not zero.

Source tiers

• Primary: company newsroom / PR, product or datasheet pages, IR and SEC / HKEX filings, patent office bibliographic pages, court dockets, Federal Register.
• Quality press: Reuters, Bloomberg, Caixin, and similar outlets when they quote the vendor or a filing. Labeled reported when they restate a briefing.
• Not used as sole cite: PitchBook / Tracxn / Crunchbase scrapes, unofficial spec blogs, Wikipedia timelines, and competitor AI blurbs.

What a citation contains

Every non-obvious fact should carry:

• Source URL
• Access date
• Short snippet (quotation, not a republication)
• Confidence: confirmed (primary), reported (quality press), or rumored (named as rumor)

Click the small citation chip next to a number for that payload. Software/stack rows use a separate claim vs confirmed label: vendor-stated capability versus a source that discloses the mode or an independent report that it happened.

Capital

One announcement is one round. We do not triplicate a single raise as seed + Series A + strategic. “Exceeded $X” is stored as a floor with a note. Public companies get a public-markets note — not an invented humanoid program budget. Homepage KPIs count rows we have, never a synthetic “market total.”

Commercial

Named counterparties only. A trial is not a deployment. An LOI or offtake target is paper, not a paid pilot. Deal size and unit counts stay empty unless the source states a number.

Specs and prices

Dated spec sheets live as versions. Numeric fields need a single stated number — ranges are not midpoints. List / EDU / deposit / rumored prices are stored only when a vendor page or quality press states them. We do not publish capability scores or editorial radars without a cited method.

Legal / IP

Patent rows are number, title, dates, assignee, and a public page URL. We do not republish claims or specification text. Legal events need a docket, statute, or Federal Register cite. This is not legal advice or a freedom-to-operate opinion.

What we do not track

• Capability or “readiness” scores
• Inflated market-size totals
• Live-agent or scrape-theater status badges
• Unnamed “select customers”
• Non-humanoid catalog filler

Machine-readable access

A structured Data API (REST or similar) is planned so analysts — and later, agents — can read the same cited records without scraping HTML. It is not live. Subscriber CSV export is available today. See /llms.txt for a plain-text product summary.

Update cadence

Seed changes and corrections appear on the changelog. There is no claim of continuous machine aggregation. If a cell is wrong or stale, report an issue.