About Humanoid Index

1 · A demo doesn't show autonomy

A robot folding laundry on stage may be running its own policy or may be a puppet worn by a person in a motion-capture suit in the next room. The footage looks identical. The distinction is the entire question — it separates a product from a prototype — and it is almost never in the caption.

So autonomy is the heaviest pillar here at 25%, scored from a published ladder rather than an impression, and teleoperation is capped near the floor of it: research 20, teleoperated 35, supervised 70, task-autonomous 92. No profile for any robot category is permitted to redefine those rungs.

2 · Spec sheets are marketing documents

Payloads are quoted at best case, runtimes without load, and the numbers that would embarrass a machine are simply absent. A ranking that fills the gaps with estimates quietly rewards whoever discloses least.

So every figure here is cross-checked against at least two independent sources on different registrable domains, carries a per-robot confidence grade and an as_of date, and — the rule that does the real work — a missing spec scores zero for that sub-metric and is never imputed. Absence is a finding, not a blank to fill in charitably.

3 · One score can't span every robot

The obvious move — one grand index over every robot — produces nonsense the moment you leave humanoids. Score a quadruped against a human hand's 22 degrees of freedom and manipulation returns 0.0 for lacking thumbs. Score a surgical arm on walking speed and mobility returns 0.0 for being bolted to the floor. Both are arithmetically correct and semantically worthless.

So each category carries its own index — its own weights, its own reference scales — and composites are ranked within a category, never across. The 10 categories →

4 · Rankings go stale faster than they're rebuilt

This field ships something material every month. A hand-built ranking is substantially wrong within a quarter, and the ones you find are usually a blog post with a date on it and no way to tell what changed since.

So this one is continuously re-researched, every change is checked against its sources before it lands, and every change that lands is published on a public changelog and RSS feed with those sources attached. You can see what moved, when, and on whose evidence.

Every claim on this site is traceable

• The methodology renders each scoring profile verbatim from the code that computes it — weights, ladders, reference scales and what each denominator is anchored to.
• Each robot page lists its source URLs, its confidence grade, and the date its figures were last verified.
• The full dataset is a JSON endpoint. Recompute the index yourself and tell us where we're wrong.
• The changelog records every applied change; entries applied by an agent rather than a human reviewer are badged as such, never presented as human-reviewed.
• All 156 published humanoid scores are locked by a snapshot test, so a change to the scoring code that silently moves a robot fails the build instead of shipping.

What this is not

• Not a buyer's guide. A high index score means a machine is well documented and capable on published evidence — not that it suits your process, your safety case or your budget.
• Not investment advice. The ecosystem pillar reads funding and valuation as signals of survival odds, nothing more.
• Not vendor-funded, and not affiliated with any manufacturer. No company pays to appear, to rank, or to be re-checked, and no listing is removable by request — corrections are, with sources. See the trademark notice.
• Not a measure of what a robot could do. It scores what has been demonstrated and published. Unreleased capability is real and is not evidence.