European Humanoid Reference Platform · in formation
Shared specifications for humanoid robots. Before every maker sets its own.
Humanoid robotics will not be won in isolated silos. EHRP unites Europe's component makers, robot OEMs, research institutions, and test bodies around open reference architectures, shared safety evidence, and standard interfaces - before closed monopolies lock down the market.
STATUSIn formation
INITIATIVEPre-competitive European Reference
WORKING DOMAINHumanoid Robotics & Safety Standards
While closed platforms in the US and China pursue vertical integration, Europe's competitive edge lies in precision engineering, specialized sensors, and rigorous safety. Yet without shared reference interfaces, every team spends valuable development cycles recertifying the same baseline from scratch.
No shared target description.
Which components, with which properties, at which interfaces, are needed is not defined in European or international standardisation. Suppliers who want to enter the market do not know what to develop for.
The end product is assessed.
The legal framework assesses the robot for its intended use, not the part. Each robot maker therefore translates its suppliers' evidence into its own safety case.
Every customer asks differently.
Own questionnaires, own audits, own assumptions about the application. With N robot makers and M components, up to N × M translations of the same kind arise.
We have all seen this before
Humanoid robotics is not the first industry to face this. Three patterns everyone knows, and one warning.
The charger drawer
For years, every phone maker had its own charging socket, and drawers filled up with adapters that fit nothing else. Today one connector, USB-C, works across manufacturers: specified by an industry consortium, and mandatory for new phones sold in the EU since 28 December 2024.[1]
For humanoid robots: Joints, hands and sensors need shared interfaces before every robot maker fixes its own.
Tested once, accepted everywhere
Since the 1990s, chip makers qualify a part once against AEC-Q100, and car makers accept it without repeating component-level tests.[2]
For humanoid robots: That is what EHRP wants for components of humanoid robots: evidence provided once, read by every integrator.
Red means stop
A traffic light means the same everywhere, no matter who built it. Nobody would accept a city where green means "go" on one crossing and "wait" on the next.[3]
For humanoid robots: A humanoid robot's signals must mean the same on every robot, regardless of the maker.
The warning: too late, everything goes in
In the 1990s, makers of industrial field buses could not agree. The international fieldbus standard IEC 61158 therefore contains numerous incompatible protocol types side by side.[4]
For humanoid robots: A standard that comes after the market has settled can only collect the variants.
In North America, Tesla's proprietary charging connector became the industry standard SAE J3400 after Tesla opened it to others in 2022. Whoever ships first sets the interface.[5]
Evidence routes are being set now
In 1980 the technology scholar David Collingridge described a dilemma: while a technology is young, it is easy to shape, but its effects are hard to foresee; once they are clear, it is entrenched and costly to change.[6] For the interfaces and signals of humanoid robots, that window is open now.
Status: October 2026
The Machinery Regulation (EU) 2023/1230 applies from 20 January 2027. Humanoid robots placed on the market as machinery fall under it. Manufacturers are deciding their evidence routes now.[7]
AI requirements come through the Machinery Regulation. Under Regulation (EU) 2026/1744, requirements for AI systems in machinery are to be set by delegated acts under the Machinery Regulation, to apply by 2 August 2028 at the latest.[8]
The safety standard is under way. With ISO 25785-1, ISO/TC 299 is developing a safety standard for dynamically stable industrial mobile robots; it is at Committee Draft stage.[9]
Others are setting interfaces. In China, the standardisation committee for humanoid robots of the Ministry of Industry and Information Technology (MIIT) presented a national standards system for humanoid robots in February 2026 that explicitly covers limbs and components.[10] In the US, proprietary systems are creating facts.
A shared interface only helps if it is in place before the first large deployments fix their own.
Specification, test suite, reference
Every EHRP specification has three parts. The model is the Java Community Process; in industry, the EtherCAT Technology Group and AUTOSAR work in a similar way.
Part
Content
Access
Specification
Interfaces, data formats, behaviour
open and free of charge, including for non-members
Conformance test suite (CTS)
Digital test specifications, test vectors and validation procedures (informational; no physical hardware provided)
open and free for self-testing
Reference implementation
Working implementation showing that the specification can be implemented
open
The conformance test suite (CTS) is normative, not the reference implementation. This way no single implementation becomes the hidden benchmark.
Three building blocks exist as drafts. They are a basis for discussion, not final decisions.
Example 1 – A designation every integrator can read Draft
EHRPCore+Motion+Manipulation/G2/E-HYG/PL d
Core+Motion+Manipulation
Profile: which parts of the specification the component implements. Core is mandatory.
G2
Grade: load class, here industrial multi-shift operation. The permitted temperature range is part of the grade.
E-HYG
Environment class: hygiene and washdown areas. Environment classes can be stacked and refer to existing regulations.
PL d
Integrity level: the safety-related part is designed as a subsystem for safety functions up to Performance Level d under ISO 13849-1. The safety function in the robot is still verified by the robot maker.
Read as: a joint actuator for multi-shift operation in washdown areas. Suppliers know the pattern from IP69K or AEC-Q100. Because each axis is tested separately, the evidence adds up instead of multiplying.
Example 2 – What a robot shows the people around it Draft for comment
A humanoid robot has no driver. Its signals address people who know nothing about the system: a visitor in a warehouse, a nurse on a night shift, a resident in a care home. People accept a robot more readily when they can see what it is about to do. The Signal Convention therefore builds on patterns people already know from traffic and everyday devices, and defines, across manufacturers, which state is shown and how. If green means "autonomous" on one maker's robot and "safe to approach" on another's, a mixed fleet becomes a safety problem and loses people's trust.
B1 · Start-up warning
Yellow · blink
The robot starts moving from standstill. Head ring and torso bands blink yellow; a start-up tone sounds if a person
is nearby. Safety-related (I1), mandatory.
A4 · Teleoperated
Blue · pulse
A remote operator takes control of motion. Head and torso bands pulse blue (Z1), signalling to bystanders that a person is in active control, complemented by privacy indicator D1 (I2-D). Safety-related (I1), mandatory.
C1 · Imminent fall
Red · flash
The balance controller reports an instability it can no longer correct. Red flashing and an unmistakable warning
tone within 200 ms. This is a supplementary warning: it does not give people right next to the robot enough time to
step away. Whether it can become a safety function is one of the draft's open questions.
The draft covers 36 states in six groups, 16 of them mandatory, and a small vocabulary of four safety colours under IEC 60073 plus white, four patterns, nine tones and five privacy symbols. Every meaning is coded twice so that it stays readable with colour vision deficiency.
A lidar unit qualified for vehicles comes with evidence on EMC, vibration, temperature, ingress protection and functional safety under ISO 26262. In a humanoid robot, mounting position, motion profile, proximity to people and the safety regime all change. Today each customer works out on its own which part of the evidence still holds.
Today
Up to N × M translations
With EHRP
Essentially M, plus a small delta (Δ) per integrator
EHRP aims to do this translation once per application profile: for each component category it records which existing evidence holds unchanged, which holds with a named additional test, and which does not hold. Up to N × M analyses become essentially M, plus a small delta per integrator.
This is only proven once a pilot case has measured the effort saved. We are looking for that pilot case.
the owner of shared specifications, application profiles and delta rules
a notified body; CE conformity stays with the manufacturer
the operator of a voluntary conformity test against EHRP specifications (planned)
a standards organisation
a place to pool case data and position papers
a lobby for lower protection levels
EHRP supplies input to standardisation and does not compete with it. EtherCAT (today IEC 61158 and IEC 61784) and OPC UA (IEC 62541) took this route.[11][12]
plannedPilot case: one component, one supplier, two integrators, one test body. The result is the measured share of the evidence that can be reused.
plannedFirst quarter of 2027: fix the axis model and designation as part of the core specification.
plannedSignal Convention for comment: feed the safety-related part into the standardisation work on ISO 25785-1.
plannedExpert workshop on component interfaces: users mark demand, suppliers mark capability.
plannedAim for end of 2027: core specification version 0.9, a working conformance test suite (CTS) and two demonstrably conformant implementations.
A Message from the Initiator
"We did not launch EHRP to form another committee. We launched it because open reference architectures always defeat closed monopolies."
Initiated by Dr. Shota Okujava (CEO of isento GmbH) and isento GmbH as a pre-competitive foundation for Europe's humanoid robotics ecosystem. All intellectual property, specifications, and test suites pass unconditionally to the independent association upon formal registration.
Taking part is reserved for members. The results are open to all.
We are looking for organisations that want to work, not just support: experts who lead a working group, co-own a pilot case or contribute test methodology.
And we are looking for disagreement. Is starting with a single pilot case right? Does the axis model hold up in testing practice? What would keep you from taking part?
We use essential storage for member session access and optional analytics (HubSpot) to evaluate document reach and improve pre-standardisation drafts. For details, see our Privacy Policy.