---
title: "Five Standards Landed. No Humanoid Has Entered the Factory Yet."
date: 2026-11-21
category: Policy & Governance
site: NeuroAI
canonical: https://neuroai.site/a/na-policy-humanoid-standards-1121
language: en
---

# Five Standards Landed. No Humanoid Has Entered the Factory Yet.

> Five documents, four layers. One layer holds two.

Five documents, four layers. One layer holds two.

That is the shape of the humanoid robot standards released by TC8, a technical committee under China's Ministry of Industry and Information Technology. The five cover capability, components, data and application.

Five divided by four is 1.25. At least one layer carries two documents; the rest carry one. Which layer holds the extra one is not in the public account.

So the confirmed structure is five files, four layers, one remainder.

## Four words, no expansion

The published scope is four words, side by side, with no order of precedence and no list of indicators inside each.

Capability. Components. Data. Application.

No standard numbers, no titles, no effective dates are public. What can be written is which four layers are covered. What each layer specifies in detail sits in the documents themselves.

Within that boundary, one observation holds: the layers describe parts of a robot, and the complete machine is not one of the four. Capability and application sit close to the product; components and data sit close to the base. Two layers govern how machines are compared, two govern whether a result can be reproduced.

## Comparability before measurement

Take a claim that a robot lifts 20 kilograms. A second maker claims 25.

If the weighing position, the speed and the fixture differ, the two figures are not comparable. Once components and data have shared definitions, the numbers acquire shared coordinates. Only then does measurement mean anything: the same capability tested under the same conditions produces results that can be placed side by side.

Coordinates first, scores second.

The unit of comparison also changes. A joint module, a dataset, a work sequence each sit on the same table once defined the same way. Comparing whole machines shows the final gap; comparing layer by layer shows which segment produced it.

## Chengdu counts by batch

On November 19, Chengdu Rail Transit published a second list of 10 "rail plus" scenarios, following a first list of 16.

Ten is 62.5% of 16. How much time passed between the two lists is not stated. What is confirmed is the sequence and the counts: 16 first, then 10. A later batch is not automatically a replacement.

The lists write needs down as items; suppliers match against them. How many batches appear in a year is not public.

Of the first 16, 8 were signed and 4 more had a partner named. That is 12 of 16, or 75%. Four items, a quarter of the batch, have no counterpart yet.

Twelve have a landing point. Twelve are not necessarily live: signing and going operational are two different states, and the distance between them is not given.

A ratio belongs to its own denominator. 75% of 16 is 12; 75% of 10 is 7.5. The two batches cannot be merged into a 26-item pool, because whether the second list adds to or replaces the first is not stated.

## 85% is an average

The same operator's AI hazard-recognition model covers 12 major construction scenarios with overall accuracy above 85%.

Read that as at most 15 errors in every 100 recognitions.

Whether one model serves all 12 scenarios or each has its own is not stated, and the two structures read differently. One shared model spreads its errors across every class. Twelve models keep the errors separated.

Above 85% is an overall figure, not a floor for each class. Which scenarios score high and which low is not broken out. Whether the aggregate is weighted or a simple mean is not stated either, and the two methods produce two different 85s.

An average cannot stand in for a single class.

The scenario list comes before the product. It records where people are needed, where the work is dangerous, where it repeats. A product then answers which device, how many eyes, how often.

Questions first, answers later.

## Booth, test track, line

This week the humanoid appeared at a booth. GRG Banking showed a humanoid teller.

A robot dog is testing on the Chengdu-Meishan line.

And at Aiboge's plant in Changzhou, Tashi Zhihang's robots are working on harness assembly. The public account does not say whether those machines are humanoid.

Three locations, three kinds of evidence. A booth shows a prototype. A test track shows a route. A line shows units leaving on takt.

The standards describe humanoids. The line report does not use the word.

## The dog is still testing

Testing and scheduling are different things. Testing asks whether a device works at all. Scheduling asks whether it works every shift.

The five standards govern capability, components, data and application. The scenario lists govern which job comes first and who takes it. Both arrive before the product does.

The dog on the Chengdu-Meishan line inspects. It replaces eyes, not hands: fasteners, rail surfaces, transponders. A humanoid bound for a factory is graded on hands.

A device that replaces eyes fails by missing something. A device that replaces hands fails by stopping the line.

What can be counted is the three positions: exhibited, tested, installed. And in the items reported this week, the installed machines are not described as humanoid.

That is what the title says. It does not claim no humanoid will enter a factory. It records that in the confirmed reports so far, none has.

Four layers of definitions now exist for a category whose production evidence is still a booth and a test track. That is the week's actual arrangement: the vocabulary arrived before the throughput.

The machine that replaces hands is still the one being waited for.

*Disclaimer: This article is based on publicly available information and is for reference only. It does not constitute investment advice.*

---

Published by NeuroAI (https://neuroai.site/) — https://neuroai.site/a/na-policy-humanoid-standards-1121
Free to quote with attribution and a link to the original.
