ISO/WD 26264-1 is a formal international standard in development that defines a data architecture for humanoid robots: kinematics, coordinate systems, and data synchronization. The goal is to make robot data reusable across brands. Its current status is working draft.

Technology

The protocol stack of Physical AI

Q: What protocol do humanoid robots use?

Virtually every modern humanoid robot runs ROS 2, which is built on DDS (Data Distribution Service). ROS 2 defines how software nodes on the robot publish and subscribe to data streams; DDS handles the real-time, peer-to-peer transport underneath. MQTT is typically added for edge-to-cloud fleet telemetry, and gRPC for connecting heavy AI models to fast control loops.

Q: What is a VLA format?

VLA stands for Vision-Language-Action. A VLA format stores robot interaction data as three streams synchronized to the millisecond: RGB-D vision, language commands, and action vectors or forces. It is the core data standard of modern robot learning, used by projects like Google RT-X / Open X-Embodiment and Physical Intelligence pi-zero to train foundation models on demonstrations.

Q: What is OpenUSD used for in robotics?

OpenUSD translates the physical world — geometry, physics, materials, lighting — into a standardized 3D file. In robotics it is pushed by NVIDIA (Isaac Sim) and the Alliance for OpenUSD. It lets one robot scan an environment and share that spatial model with simulators and other robots.

Q: Is there a standard for transferring what a robot has learned?

No. VLA and LeRobot are training formats used before a robot ships. No open standard exists for the knowledge a deployed robot accumulates in the field, or for moving that knowledge between devices and vendors. ISO/WD 26264-1 is the nearest effort, but it covers dataset architecture only, not runtime knowledge or ownership, and it is still a working draft.

Four layers of protocols make a robot work: transport, skills, space, and safety. A fifth layer is missing. This page maps all of them.

Last updated: July 2026

The landscape

Four layers, one stack

Physical AI is harder to standardize than software AI. A robot processes massive real-time sensor streams, runs heavy AI models, and carries physical safety risk. One protocol cannot cover all of that. So the landscape is organized in layers, each with its own standards.

In 2026 the stack looks like this. Four layers exist. Each is covered in detail below — and then the layer that nobody has built yet.

Layer	Standards	Function
Infrastructure	ROS 2 / DDS, MQTT, gRPC	Real-time internal communication between sensors, motors, brain
AI & behavior	VLA formats (Open X-Embodiment), LeRobot, ISO/WD 26264-1	Storing and sharing physical skills
Environment	OpenUSD, URDF/SDF	Digitizing and sharing 3D space and robot bodies
Safety	NVIDIA Halos, OPC UA Safety	Runtime guardrails against physical risk

Layer 1

Communication & middleware — the nervous system

Everything inside a robot has to talk in real time: cameras, joint encoders, motor controllers, the AI model deciding what to do next. Four protocols carry that traffic.

DDS

Data Distribution Service is the industry standard for real-time, decentralized data exchange. It is peer-to-peer: there is no central broker, so one failing component does not silence the rest of the robot.

ROS 2

The Robot Operating System 2 is built on top of DDS. It defines how robot software nodes talk to each other through a publish/subscribe model. Virtually every modern humanoid runs ROS 2.

MQTT

Originally an IoT protocol. Now the standard for edge-to-cloud fleet telemetry: robots reporting status, position, and diagnostics to a fleet backend, through platforms like EMQX.

gRPC / Protobuf

How heavy AI models (usually Python) talk to fast robot control loops (usually C++). Low latency, efficient binary serialization. The bridge between the thinking half and the moving half.

Layer 2

Skills & trajectory data — the brain

This layer defines how physical skills are stored and shared. It is where robot learning happens.

VLA formats (Vision-Language-Action)

The core standard of modern robot learning. Google RT-X / Open X-Embodiment and Physical Intelligence pi-zero store interaction data as three streams synchronized to the millisecond: RGB-D vision, language commands, and action vectors or forces. Foundation models for robots are trained on exactly this data shape.

LeRobot dataset format

A fast-growing open-source standard from Hugging Face for storing and sharing robot manipulation data. Built for training on human demonstrations, and increasingly the default way research labs publish manipulation datasets.

ISO/WD 26264-1

A formal international standard in development. It defines a data architecture for humanoid robots — kinematics, coordinate systems, synchronization — so that data becomes reusable across brands. Status: working draft. It is the first sign that standards bodies see the cross-vendor data problem.

Layer 3

Spatial awareness & simulation — the eyes

OpenUSD

Pushed by NVIDIA (Isaac Sim) and the Alliance for OpenUSD. It translates the physical world — geometry, physics, materials, lighting — into a standardized 3D file. One robot scans a room; others know it. The same format drives simulation, so skills trained in Isaac Sim map onto real spaces.

URDF / SDF

XML formats that describe exactly how a robot body is built: joints, link lengths, torque limits. AI models read this description to control a specific body. Without it, a policy trained on one arm cannot drive another.

Layer 4

Safety & runtime — the emergency brake

When a chatbot hallucinates, you get a wrong answer. When Physical AI hallucinates, something or someone gets hit. That difference is why safety is a protocol layer of its own.

NVIDIA Halos

A full-stack runtime guardrail. Before motors execute, Halos checks whether the commands coming out of the AI model are physically safe in the current environment. The model proposes; the guardrail disposes.

OPC UA Safety

The industrial functional-safety standard. It guarantees that robots and production lines stop when a safety limit is crossed, independent of what the AI stack is doing.

Layer 5

The missing fifth layer

Look at what each layer does not do, and a gap appears.

VLA and LeRobot are training formats. They exist to build foundation models from demonstrations, before the robot ships. No standard exists for what a deployed robot learns afterward: this site, these exceptions, these people. Every vendor solves it proprietarily — fleet learning stays locked inside each manufacturer.

OpenUSD stores geometry, not meaning. It knows where the wall is. It does not know that the third floor is natural marble that only takes pH-neutral cleaner.

Nothing in the stack defines a persistent knowledge entity — "this person", "this site", "this fleet" — that survives across devices and vendors. Nothing regulates who owns learned knowledge or how it moves when you switch brands. ISO/WD 26264-1 comes closest, but it covers dataset architecture, not runtime knowledge, not ownership — and it is still a draft.

There is also no provenance layer. Once knowledge transfers between machines, which learned behavior is verified safe? Where did it come from? Halos checks commands at runtime; nothing certifies transferred knowledge.

The stack has transport, training, space and safety. It has no memory.

There is a historical parallel: this is the internet before an identity and data layer existed. TCP/IP and HTTP were there; "who are you and what do we know about you" had to be invented on top.

This fifth layer — a persistent, owned, transferable knowledge layer — is what benned builds with Kin. See One AI. Every machine., field-learned knowledge for fleets and the developer platform.

FAQ

Frequently asked questions

What protocol do humanoid robots use?

Virtually every modern humanoid runs ROS 2, built on DDS for real-time, peer-to-peer transport. MQTT handles fleet telemetry to the cloud; gRPC connects AI models to fast control loops.

What is a VLA format?

Vision-Language-Action: robot interaction data stored as three streams synchronized to the millisecond — RGB-D vision, language commands, and action vectors or forces. It is the data shape robot foundation models are trained on, used by Open X-Embodiment and Physical Intelligence pi-zero.

What is OpenUSD used for in robotics?

OpenUSD digitizes the physical world — geometry, physics, materials, lighting — into a standardized 3D file, backed by NVIDIA (Isaac Sim) and the Alliance for OpenUSD. One robot scans a room; simulators and other robots can use it.

Is there a standard for transferring what a robot has learned?

No — that layer does not exist yet. VLA and LeRobot cover training data before deployment. ISO/WD 26264-1 is the nearest draft, and it covers dataset architecture only, not runtime knowledge or ownership.

What is ISO 26264?

ISO/WD 26264-1 is an international standard in development that defines a data architecture for humanoid robots: kinematics, coordinate systems, and synchronization, so data becomes reusable across brands. Current status: working draft.

Keep reading