BLOGS

On September 28, 2026, China's market regulator formally published the country's first six national standards for humanoid robots — five parts of GB/T 48317, *Humanoid Robots — Technical Requirements*, plus GB/T 48316, a technical specification for simulation testing platforms. They take effect April 1, 2027. The headline numbers are striking: recover from a push within one second, grasp rigid objects with at least 90% success, cap speed near humans at a reference 0.5 m/s. But the deeper story is who has to produce those numbers. Almost every quantified requirement lands, physically, on the joint. This article walks through what the standards actually measure and what they demand from joint module engineering.

Six Standards, One Closed Loop

The first thing to understand is the architecture. The six documents are not six unrelated rulebooks; they form a deliberate loop that mirrors how a humanoid actually works:

• GB/T 48317.1-2026 — Part 1: General principles. A shared technical vocabulary: body composition, system architecture, interface boundaries, and baselines for dimensions, performance parameters, battery voltage and communication protocols. Every other part depends on it.
• GB/T 48317.2-2026 — Part 2: Environmental perception. Splits perception into five capability domains — scene understanding, object recognition, multimodal sensing, behavior prediction, 3D reconstruction — each with testable error ranges and recognition-accuracy metrics.
• GB/T 48317.3-2026 — Part 3: Decision and planning. Defines information adaptation, decision-making, task planning and knowledge-base management, with acceptance figures such as command accuracy above 95% for typical cross-scenario tasks and single-action completion within one second.
• GB/T 48317.4-2026 — Part 4: Motion control. The densest part of the set: a hardware layer plus three software layers with explicit boundaries between them, and performance defined across accuracy, real-time behavior and stability.
• GB/T 48317.5-2026 — Part 5: Manipulation. Single-arm, dual-arm and mobile operation, with the most concrete numeric thresholds in the whole series.
• GB/T 48316-2026 — Simulation testing platform. A common, reproducible virtual test bench, pushing the industry toward "physical test plus simulation verification" as a combined evaluation model.

All six are recommended national standards (GB/T, not mandatory GB), administered under the national robotics standardization technical committee, with domestic humanoid manufacturers participating in drafting. The drafting principle that matters most has been summarized in one line: set the ruler, don't lock the route. The standards prescribe what capability a robot must demonstrate and at what performance — they do not mandate which algorithm, actuator topology or hardware realization gets you there. Direct drive and geared, electric and otherwise, all compete on the same test bench.

The Numbers That Become Contract Language

The second thing to understand is how these numbers will be used. Once a recommended standard is cited in procurement documents and acceptance tests — which is exactly what a unified metric set is for — its figures stop being marketing comparisons and become contract language. Based on the official standard listings and industry summaries of the texts, the quantified highlights include:

A note on sourcing: the six-document list, publication date and implementation date are confirmed on China's official standards registry. The 0.5 m/s near-human limit and the 90% rigid-grasp threshold appear across multiple independent industry summaries. The remaining figures in the table above are drawn from a detailed industry write-up of the standard texts and should be read as widely reported summaries rather than quotations from our own copy of the documents — treat them as directionally firm, pending the purchasable full texts.

Two of these rows deserve their own discussion, because they reshape joint engineering rather than merely grading it.

Disturbance Recovery Is a Joint Problem

"Push the robot with a rated external force and it must return to a stable working state within one second" is the requirement industry summaries highlight from Part 4, and it is the one that most directly exposes what is inside a joint.

Recovering from a push is a chain of events measured in milliseconds. The joint must first *feel* the disturbance — which is where torque transparency and low-friction transmission decide whether the disturbance is visible to the current loop or hidden behind stiction. It must then reject the impulse without saturating, which is a function of peak torque margin and, critically, of how much error the transmission injects before the controller even sees the deflection. Backlash in the tens of arc-minutes means the output moves before the motor-side encoder notices anything; as we argued in our earlier piece on the loaded error budget of a robot joint, the encoder's resolution says nothing about what the load actually experiences. A joint that cannot see the disturbance cannot reject it inside one second, no matter how good the balance controller is.

This is also why the standards' insistence on a "hardware layer plus three software layers" architecture with explicit interfaces is quietly significant for joint vendors. A joint module that exposes clean, low-latency state — position, velocity, torque estimates — at a defined interface makes the recovery problem tractable for whoever writes the balance layer. A joint that buries its state behind an opaque protocol forces every integrator to rediscover it. The standard does not mandate a torque sensor; it mandates an outcome. But the outcome is only reachable when the joint's electromechanical design cooperates with the software stack instead of fighting it.

The same logic runs through the manipulation thresholds. A 90% success rate on rigid grasps, and 85% on deformable ones, cannot be achieved by a hand alone — the wrist and arm joints must deliver smooth, low-ripple torque at very low speeds, and they must be compliant enough that misestimated contact forces become gentle transients instead of crushed objects. Force control entering mass production, as we covered earlier this year, is precisely the industry learning to hit numbers like these consistently, unit after unit.

Why Buyers Should Read the Test Method, Not the Headline

For anyone specifying humanoid components in the next eighteen months, the practical effect of a unified metric set is that the test conditions become the product. Until now, a datasheet could quote peak torque at the motor shaft, accuracy measured unloaded, and cycle life at a friendly duty point — three different robots' spec sheets were three different grading schemes. We have made this argument before from the vendor side: peak torque sells, continuous torque ships. The standards push the same argument from the regulator side.

What changes for a procurement engineer evaluating joint modules:

• Ask for the metric under the standard's conditions. If a recovery or repeatability figure is quoted, ask which test produced it — loading, posture, temperature. A number without a test method is a marketing sentence.
• Treat interface boundaries as a hard requirement. The general-principles part standardizes vocabulary and interfaces precisely because mismatched interfaces are a leading integration cost. A joint that already exposes standardized state and control interfaces is cheaper to adopt, not just more compliant.
• Check the limiting modes. The 0.5 m/s near-human cap implies joints must support speed and power limiting as a first-class operating mode, not a firmware afterthought. Ask how the limit behaves when the joint is hot or near its torque ceiling.
• Score vendors the way the standard will. Grasp success, push recovery, repeatability — build your acceptance tests around the same quantities now, and 2027 acceptance becomes a formality instead of a redesign.

There is a second-order effect worth naming: unified metrics compress the value of vague datasheet superlatives and raise the value of measured, load-honest performance. That is a structural advantage for integrated joint makers, whose products are tested as a unit, over component sellers whose parts get graded only after someone else's integration work.

What Changes Before April 2027

The standards are recommended, take effect April 1, 2027, and the industry read of their intent is consistent: from "each speaks its own language" to one shared technical ruler, ending the era in which every vendor graded its own homework. Between now and then, the work splits into three lanes.

For joint and actuator vendors: inventory your product against the quantified thresholds now — including our own lightweight joint modules, which are built and factory-tested as single units precisely so that figures like repeatability and thermal behavior can be quoted for the joint, not for its parts. Disturbance recovery, repeatability under load, limiting modes, interface documentation — every gap discovered today is a spec-sheet fix; discovered in an acceptance test next year, it is a redesign. Build or rent access to the standardized test conditions, and start quoting test-condition-normalized figures.

For robot builders: the standards do not lock your technical route, but they do define the acceptance language your customers will use. Aligning your internal test suite with the standard's quantities — push recovery, grasp success, command accuracy — shortens every future customer conversation.

For the industry's supply chain: a common ruler favors modules over loose parts. When the graded unit is "the joint as tested," the integrated joint module — motor, reducer, encoder, driver, factory-calibrated as one unit — is the artifact that can be certified once and reused everywhere. That has been the direction of travel for two years; the standards give it an official destination.

The six documents are a ruler, not a cage. But a public ruler changes behavior in a private datasheet's shadow: the vendors who can show their numbers under the standard's conditions will be the ones whose claims survive contact with a test bench — and with a customer.

Source: Open Standards Information Platform (openstd.samr.gov.cn), GB/T 48316-2026 and GB/T 48317.1–.5-2026, published 2026-09-28, effective 2027-04-01 (official registry).

Source: 同花顺财经, "首批六项人形机器人国家标准核心要点梳理", 2026-10-23 (industry summary of the six standards' quantified requirements).

Source: Multiple independent industry and financial-press summaries of the first humanoid robot national standards, October 2026 (corroborating the 0.5 m/s near-human limit, ≥90% rigid-grasp threshold, and April 2027 implementation).