By the EVST Applications Engineering Team · Last updated 1 June 2026 · Reviewed by EVST robotics integration engineering
On 1 June 2026 China put 102 national standards into effect, including a wave that directly touches robotics, embodied AI, industrial automation and AI agents. The core shift for vendors is from claiming intelligence in demos to proving it in testable, traceable, accountable delivery — and the legal cost of getting it wrong has risen sharply. This article explains what changed, the five categories of company most affected, how delivery contracts must be rewritten, what legal risk now looks like, and the four systems every robotics and AI vendor should build immediately.
Key takeaways
- The industry is moving from feature delivery (vague claims) to metric delivery (verifiable numbers).
- Recommended standards become binding once written into a contract, tender or product manual; mandatory standards can block sale and import outright.
- Five categories of company are exposed: embodied-AI, industrial robotics, AI applications/agents, system integrators, end users.
- Build four systems now: standards mapping table, test-and-acceptance system, log-and-traceability system, safety-and-compliance system.
- The valuable hire is now hybrid: process + AI + acceptance + standards literacy.
This article is written for founders, technical leads, procurement and compliance officers at robotics, embodied-AI and AI-agent companies operating in or selling into China. It interprets the 1 June 2026 standards wave at a strategic level; it is not legal advice for any specific product or contract.
What actually changed on 1 June 2026
On 1 June 2026 China brought 102 national standards into force. Several of them touch robotics, industrial automation, AI applications and embodied AI directly — covering testing methods, safety requirements, evaluation metrics and data handling. They sit alongside, and in places explicitly reference, established international standards such as ISO 10218 (industrial-robot safety), ISO/TS 15066 (collaborative-robot operation), ISO 9001 (quality management), ISO 3834 (welding quality) and IATF 16949 (automotive quality). The practical effect is that the language a buyer can use in a tender, in an acceptance test, or in a lawsuit, just got far more specific.
According to industry observation, this is the largest single-day batch of standards touching robotics and AI to date in China, and it lands squarely in the year that “embodied AI” became a procurement category rather than a research label. The era of selling intelligence on a demo video is closing.
From “claim” to “prove”: the metric-first shift
EVST scopes every robotics and AI deployment with what our engineers call the Prove-Don’t-Demo principle: any capability that is named in a contract must come with an agreed test, an agreed metric, and an agreed log. Anything that cannot be tested, measured or logged should not be in the contract. The principle exists because contractual claims that can’t be measured become arguments, and arguments become liability.
| Old “feature delivery” claim | New “metric delivery” specification |
|---|---|
| “supports intelligent recognition” | recognition accuracy %, dataset, lighting conditions, miss rate |
| “supports autonomous picking” | pick success rate, cycle time, mis-pick rate, collision rate |
| “AI decision-making” | human-takeover count, decision audit log, model version |
| “high reliability” | continuous runtime, MTBF, recovery mechanism, log retention |
| “data secure” | data permission scheme, retention period, access audit |
Customers are no longer buying a sense of intelligence — they are buying a capability that can be verified, reproduced and held accountable. In our integration work, the projects that survive customer acceptance are the ones where the metric grid above was written into the contract from day one, not the ones where intelligence was promised and arguments came later.
Five categories of company most affected
- Embodied-AI companies — must explain how perception, decision and execution are tested; what the test environment is; where the failure boundaries are.
- Industrial robotics and automation companies — must move past payload/reach/accuracy/price to prove safety mechanisms, recovery capability, task-success rates and operation logs.
- AI applications and agent companies entering factory scenarios — must prove output traceability, audit trails, human confirmation for high-risk actions, and model/data version management.
- System integrators — must upgrade from “experience-based delivery” to standardized delivery: documented procedures, acceptance criteria, and standards mapping.
- End users (factories) — can no longer accept a sales demo; must look at standards conformance, test reports, data scope and responsibility boundaries.
The integrator and end-user categories are the most under-served in current discourse. Buyers in particular need to learn to ask the seven procurement questions in the FAQ below.
The legal picture: recommended ≠ optional
Standards in China fall into two classes — mandatory (强制性, GB) and recommended (推荐性, GB/T) — and the second is where many companies misjudge their exposure. A recommended standard is non-binding until it is written into a contract, a tender, an acceptance clause or a product manual; once it is, it becomes an enforceable commitment. Mandatory standards carry higher risk still: products or services that fail to meet them may be barred from production, sale, import or provision.
According to publicly summarised practice under the standardization and product-quality laws, exposures for robotics and AI vendors include product removal, sales suspension, administrative fines, confiscation of illegal gains, contract-breach compensation, product-liability claims, false-advertising risk, data-compliance risk, credit-record damage and tender disqualification. Where mandatory standards are breached and significant personal or property harm results, criminal liability is also possible. None of this is academic — it is now ordinary procurement.
The four systems every robotics and AI vendor should build now
EVST recommends building these in parallel, not in sequence:
- Standards mapping table. For each delivered system, map every part — robot body, controller, vision, fixtures, safety modules, software platform, AI model, data interface — to the applicable national, industry, customer and contractual standards. This is the document a buyer’s compliance officer will ask for first.
- Test and acceptance system. Build a task library, a failure-sample library and a boundary-scenario library. Test scenarios must include lighting variation, object displacement, occlusion, network interruption, human approach, and abnormal shutdown — not just nominal happy-path runs.
- Log and traceability system. Record sensor data, model versions, control commands, abnormal events and human-takeover records. ISO 9001 and IATF 16949 already require this for adjacent processes; the new standards extend the discipline to AI decisions.
- Safety and compliance system. Define emergency stop, speed limits, permission scopes, human-confirmation triggers, risk boundaries, data security and responsibility division. For collaborative-robot tasks the ISO/TS 15066 risk assessment is non-negotiable.
In practice the failure we see most often is treating these as a documentation exercise to satisfy an audit. They are operational systems — if they aren’t generating logs, catching mis-loads and triggering re-assessments in production, they will not stand up in a dispute either.
The new valuable hire
This wave shifts the hiring market too. The valuable engineer is no longer the one who only teaches robot points or only calls a large-model API. It is the hybrid professional who reads process, AI, acceptance and standards documentation fluently. Roles that will harden over the next year: embodied-AI test engineer, robot acceptance engineer, industrial AI evaluation engineer, AI safety compliance engineer, simulation test engineer, data-loop engineer.
Where this lands across industries
- Automotive powertrain and EV — IATF 16949 already enforced; the new standards add traceability obligations for AI-mediated decisions on the line.
- Steel structures, heavy machinery, pressure vessels — ISO 3834 weld-quality discipline now extends to robot-and-vision-controlled welding cells.
- 3C, appliance and consumer-product assembly — collaborative-robot deployments must carry task-level ISO/TS 15066 assessments documented in delivery.
The cross-industry pattern is the same: the burden of proof has shifted from the buyer (asking “is this real?”) to the seller (showing “here is the test”).
FAQ
What is the difference between mandatory and recommended standards in China? Mandatory standards (GB) must be met for the product to be sold; recommended standards (GB/T) are voluntary unless written into a contract, tender, acceptance clause or product manual — at which point they become binding.
Do the new standards apply to foreign companies selling into China? Yes — they apply to products and services placed on the Chinese market, regardless of the vendor’s origin. Tender language and acceptance criteria will routinely cite them.
What should an end user ask a robotics or AI vendor at procurement? At least seven questions: which standards apply, what test scenarios are used, what success and failure rates are guaranteed, can abnormalities be traced, how is safety responsibility divided, how are data and model versions managed, who maintains and upgrades after delivery.
Are demos still useful? Yes — to win attention. They are no longer enough to win acceptance. Demos must be backed by a testable product, an acceptable solution, and a repeatable delivery system.
What is the fastest first step for a small robotics company? Build the standards mapping table for one flagship product, link each line to a test and a log, and rewrite one contract template with metric-based delivery clauses. Three weeks’ work; pays back the first time a buyer asks.
Bringing it into your delivery
The 1 June 2026 standards wave is not a paperwork tax — it is the moment the robotics and AI industry crosses from intelligence theater to standardized delivery. The companies that win the next cycle are the ones that can prove what they ship: testable products, acceptable solutions, repeatable delivery, and risk control a buyer can sign off on. EVST designs robotics and automation cells around metric-first delivery and the four systems described above — see our application guides to robot machine tending, robot ground rails for one-track-multi-machine layouts and welding cells with positioners, or talk to EVST about scoping a standards-aligned project.
About the author — The EVST Applications Engineering Team designs and integrates industrial-robot, collaborative-robot, ground-rail and welding-positioner cells across automotive powertrain, steel-structure, appliance and consumer-product industries, using the Prove-Don’t-Demo principle described above. Reviewed by EVST robotics integration engineering for technical accuracy. This article is general industry commentary on the 1 June 2026 standards wave and is not legal advice; specific contracts and certifications should be reviewed by qualified legal and standards counsel for the relevant jurisdiction. Corrections and updates: see the Last Updated date.