---
title: "China's first embodied-AI industry standard is a test method, not a safety rule"
date: 2026-11-17
category: Policy & Governance
site: NeuroAI
canonical: https://neuroai.site/a/na-policy-embodied-ai-benchmark-standard
language: en
---

# China's first embodied-AI industry standard is a test method, not a safety rule

> A recommended MIIT standard in force since June standardizes how embodied-intelligence systems are benchmarked in simulation and in the real world.

A robot vendor claims its machine completes 90% of tasks. A rival claims 70%. Neither number means anything until both sides agree how the test is run. China now has a published standard for exactly that — and it took effect on 1 June 2026.

It is the first industry standard in embodied intelligence (具身智能). It is also a measurement method, not a safety rule, and that distinction is being lost in the summaries.

## Key takeaways

- **Stage: in effect since 1 June 2026.** 《YD/T 6770-2026 人工智能 关键基础技术 具身智能基准测试方法》 (Artificial Intelligence — Key Foundational Technologies — Embodied Intelligence Benchmark Testing Methods) was approved and published by the Ministry of Industry and Information Technology (MIIT, 工业和信息化部) in March 2026, under its Announcement No. 3 of 2026, and took effect on 1 June. It is not a draft for comment (征求意见稿) or a draft standard (草案).

- It was proposed by the MIIT Artificial Intelligence Standardization Technical Committee (MIIT/TC1, 工业和信息化部人工智能标准化技术委员会) and released as one of a batch of eight AI industry standards with the same effective date.

- MIIT's own committee, the People's Daily and Xinhua all describe it as the first industry standard in embodied intelligence.

- The "/T" suffix means recommended (推荐性), not compulsory (强制性). It becomes binding only if cited in a contract, tender or regulation.

- It covers the full "perception–decision–execution" (感知—决策—执行) chain of a single embodied-intelligence system, in both simulated and real environments.

## What the standard actually fixes

The problem the document addresses is mundane and expensive: without a shared benchmark, every vendor measures a different thing. MIIT/TC1's explanation says different technical routes and products had no accepted common basis for comparison, so companies were "each speaking their own language."

The standard builds a single benchmark-testing framework with four parts:

- **Environment setup (环境设置)** — requirements for building simulation and real-world test environments, intended to reflect the complexity of actual settings.

- **Task-library construction (任务库构建)** — a method for assembling tasks, plus a generalization evaluation benchmark (泛化评估基准) that tests performance across different task complexities and capability dimensions.

- **Testing process (测试过程)** — covering both models and whole machines running a deployed model. The named methods are static simulation testing (静态仿真测试), dynamic simulation testing (动态仿真测试), real-environment testing (真实环境测试) and combined testing (组合式测试).

- **Indicator calculation (指标计算方法)** — how the results are turned into numbers.

That structure is the whole point. A claim like "our robot picks up objects reliably" becomes a statement about a defined task library, a defined environment and a defined calculation.

## The five numbers it produces

The standard specifies five indicators:

- **Task execution efficiency (任务执行效率)**

- **Task success rate (任务成功率)**

- **Human intervention rate (人工干预率)**

- **Scene disturbance attenuation rate (场景扰动衰减率)** — how much performance drops when the setting changes

- **Average task energy consumption (平均任务能耗)**

Read that list carefully. Two of the five measure what happens when things go wrong or cost more — intervention and energy — rather than simply whether the robot succeeded. A vendor quoting only a success rate is reporting one-fifth of the framework.

For a global reader, the useful comparison is to benchmark suites in language models: the value is not any single score but the existence of a shared procedure that makes scores comparable. This standard performs that role for embodied systems, with the added complication that the physical environment is part of the test.

## Recommended, not mandatory

This is where precision matters. YD/T numbers are industry standards in the communications and information sector; the "/T" denotes a recommended document. Nothing in the standard makes compliance a condition of selling a robot in China.

That does not make it inert. A recommended standard becomes a de facto requirement the moment a procurement officer writes it into a tender, a regulator cites it in a filing rule, or a court treats it as the ordinary standard of care. The same dynamic applies to China's national GB/T and GB/Z documents. The legal weight comes from the place the document is cited, not from its prefix.

So the correct sentence in a bid document is "tested in accordance with YD/T 6770-2026," not "certified compliant with a mandatory national requirement."

## The stack around it

The benchmark method does not stand alone. MIIT established a humanoid-robot and embodied-intelligence standardization technical committee in December 2025; its work draws in more than seventy leading companies and research institutes, according to the ministry's equipment-industry development centre.

A broader framework, 《人形机器人与具身智能标准体系(2026版)》 (Humanoid Robot and Embodied Intelligence Standard System, 2026 Edition), was published in 2026 with six parts: common foundations, brain and intelligent computing, limbs and components, whole machines and systems, applications, and safety and ethics. The benchmark standard sits in the testing and evaluation layer of that structure.

There is an international dimension too. MIIT/TC1 says the standard has been taken forward as an international-standard project at ITU Study Group 21. That is a proposal stage, not an adopted ITU standard — a distinction worth keeping.

## Honest limitations

I am working from MIIT/TC1's official interpretation, People's Daily and Xinhua reports, not the full text of the standard, which sits behind the standards platform's access controls. I therefore cannot verify clause numbering or the precise tolerances in the test environments. The claim that this is the first industry standard in the field comes from MIIT's own committee and state media; I have not independently surveyed every earlier sectoral standard. The standard also covers a single system's benchmark test, not multi-robot coordination or long-term safety, and I found no data on how often it is already cited in tenders.

## What readers can do now

- **Get the stage right.** It was published in March 2026 and has been in effect since 1 June 2026; it is neither a consultation draft nor a compulsory rule.

- **Demand the method, not the headline score.** Ask any supplier quoting success rates which five indicators they measured, in which environment, and under whose test procedure.

- **Watch the procurement layer.** The practical force of YD/T 6770-2026 will come from tenders and internal acceptance criteria; track whether your sector's buyers begin naming it.

---

Published by NeuroAI (https://neuroai.site/) — https://neuroai.site/a/na-policy-embodied-ai-benchmark-standard
Free to quote with attribution and a link to the original.
