---
title: "China's New Robot Grading System Runs From L0 to L5 — and It Isn't a Law"
date: 2026-11-20
category: Policy & Governance
site: NeuroAI
canonical: https://neuroai.site/a/na-policy-humanoid-grading-standards
language: en
---

# China's New Robot Grading System Runs From L0 to L5 — and It Isn't a Law

> Five approved industry standards give Chinese humanoid-robot makers a shared vocabulary for intelligence levels, training data and parts — but they are recommended, not compulsory.

A buyer at a factory wants to compare two humanoid robots. One vendor says its machine is "intelligent." The other says the same thing. Neither sentence can go into a purchase order, a tender document or an insurance clause, because nobody has agreed what the word means.

That is the gap a small group of newly approved Chinese standards is meant to close.

On 16 November 2026, the Ministry of Industry and Information Technology (MIIT, 工业和信息化部) published Announcement No. 30 of 2026, dated 10 November. It approved 274 industry standards (行业标准) across eleven sectors. Buried in the attachment, at serial numbers 118 to 122, sit five standards for humanoid robots (人形机器人). One of them gives the industry something it has been arguing about all year: a ladder from L0 to L5.

## What was actually approved

The announcement is an approval notice, not a draft for comment (征求意见稿). The five standards carry the "SJ/T" prefix, marking them as recommended industry standards (推荐性行业标准) in the electronics sector:

- **SJ/T 12353—2026** — technical specification for building a humanoid-robot training ground (人形机器人训练场构建技术规范).

- **SJ/T 12354—2026** — requirements for managing training data (人形机器人训练数据管理要求).

- **SJ/T 12355—2026** — grading of humanoid-robot intelligence capability (人形机器人智能化能力分级).

- **SJ/T 12356—2026** — technical specification for integrated electric-drive joints (人形机器人电驱动一体化关节技术规范).

- **SJ/T 12357—2026** — technical specification for dexterous hands (人形机器人与具身智能灵巧手技术规范) used in humanoid robots and embodied intelligence (具身智能).

MIIT's humanoid-robot and embodied-intelligence standardisation committee, known as MIIT/TC8, told the Beijing News that it manages the five standards. The secretary's office described three clusters: evaluating intelligence capability, defining key components and organising robot training.

The grading standard's official summary is deliberately spare. It sets out an architecture, a grading framework, level definitions and assessment methods, and says it applies to research and development, testing, evaluation and product selection. According to TC8, the framework scores four dimensions — perception and cognition (感知认知), decision and learning (决策学习), execution performance (执行表现), and collaboration and interaction (协作交互) — across levels L0 to L5.

That L0–L5 shape will look familiar from autonomous driving, but the two systems answer different questions. Driving levels allocate responsibility between a system and a human driver. These levels describe how capable a machine is at perceiving, deciding, acting and working alongside people. Nothing in the published summary allocates legal liability.

## Why "recommended" is the important word

The "/T" suffix is easy to miss and changes the meaning. A recommended standard in China is not a compulsory national standard (强制性国家标准). No robot maker is required by this notice to certify to it, and no product becomes illegal for lacking a grade.

It becomes consequential in a more indirect way. Procurement documents can cite it. A tender can ask for a particular level on a particular dimension. A buyer can write the classification into an acceptance test, and an insurer can refer to it when describing a machine's intended use. Standards often acquire force not because a regulator commands it but because contracts adopt them as shorthand.

That is why the release matters more than the label. For most of 2026, Chinese humanoid-robot marketing has been organised around demonstrations — backflips, dancing, pouring a cup. A common grading language lets a buyer ask a different question: not "what can it do?" but "on which dimension, at which level?" The standard does not settle who wins. It makes the comparison possible in writing.

The other four standards point in the same direction. A training-ground specification addresses how a site is built; a data-management requirement addresses how the records produced there are kept; joint and dexterous-hand specifications address the parts that previously forced every maker to invent its own interfaces. Together they describe a supply chain trying to become legible to itself.

## What the announcement does not settle

Three things are worth keeping separate.

First, the 274 standards are a batch, not a robotics package. Electronics accounts for 92 of them and telecommunications for 65 — nearly six in ten. The humanoid set is five items inside a much broader approval, and the headline standard on embodied-intelligence data generation sits in the artificial-intelligence section rather than under robotics.

Second, approving a standard is not the same as implementing it. The notice records approval and publication. It does not, in the material reviewed here, establish a compliance deadline or a certification body for the five robot standards.

Third, a grading framework is not a safety regime. Intelligence capability and functional safety are different properties. A machine can score well on task execution while still being unsafe near people, and the published summary does not claim otherwise.

## Key takeaways

- MIIT Announcement No. 30 of 2026, published on 16 November, approved 274 industry standards, including five for humanoid robots.

- SJ/T 12355—2026 grades humanoid-robot intelligence from L0 to L5 across perception and cognition, decision and learning, execution performance, and collaboration and interaction.

- The other four cover training grounds, training data, electric-drive joints and dexterous hands — the unglamorous inputs behind a robot demo.

- These are recommended standards, not compulsory rules; their influence will come from being written into tenders, tests and contracts.

## Honest limitations

The account is drawn from MIIT's announcement page and reporting by the Beijing News, which quoted the TC8 secretariat. The full texts and attachments of the five standards were not examined line by line here, so the summary of the four grading dimensions relies on the committee's public explanation rather than the normative document. No effective or mandatory date has been asserted because none was established in the sources reviewed. The distinction between a recommended industry standard and a compulsory national standard follows the SJ/T designation; readers should verify the status of any specific standard before citing it in a contract.

## What readers can do now

If your organisation buys robots, start putting the four dimensions into your next request for quotation — even a rough "which level, on what evidence" question will separate serious vendors from demo shops. If you build components, check whether SJ/T 12356 or 12357 matches a part you make, because interface standards tend to decide which supply chains get invited. And if you read a headline saying China has "regulated humanoid robots," check the verb: approving a recommended standard is a real step, but it is not the same as a rule taking effect.

---

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