Six levels of robot: Arm brings 80 companies into a shared language for physical AI

Arm has pulled more than 80 companies into a programme called Arm Total Design for Physical AI, and alongside it has published the first sketch of a Robotics Capability Framework: six levels that grade what a robot can do, from machines that follow fixed rules to systems that learn and improve themselves. The taxonomy is the part with consequence. Whoever writes the vocabulary an industry quotes in its buying decisions holds a quiet hand on the steering wheel.

Arm Watch analysis

What happened

The programme gathers expertise across software stacks, AI models, sensors, compute hardware, virtual platforms and digital twins, so that member companies can build and validate interoperable systems earlier. The roster already spans AWS, ECARX, Hugging Face, Liquid AI, NXP, PlusAI, Psyonic, QNX, Qwen, Siemens and Unitree Robotics, among others.

Dermot O'Driscoll, Arm's vice president of go-to-market and customer solutions for physical AI, told The Robot Report the idea was born of frustration. Companies wanted to build on Arm architecture but struggled to get partnerships moving, and when he moved over from Arm's cloud ecosystem work in March he found the same gap: nowhere for companies to share technology, ideas or simply collaborate. The ecosystem, he says, "could really use a point of coherence".

The framework is his answer to the vocabulary problem. Arm proposes six levels of capability, from zero to five. The bottom of the scale covers robots that react to stimuli and execute commands under fixed rules. The top covers systems with evolving behaviour that learn and self-optimise. The inspiration is SAE International's autonomous-driving levels, though O'Driscoll argues robotics is more complex and needs a multi-dimensional framework rather than a single ladder.

Two details soften any land-grab suspicion. Arm describes the original six levels as a starting point and is openly inviting the industry to rework them. It also plans member collaboration sessions in the coming months, because O'Driscoll's view is that members should see real technical and commercial benefits sooner rather than later.

There is an automotive subplot. O'Driscoll claims every vehicle on the road does at least some of its computing on an Arm chip, and that carmakers have noticed their safety-critical, real-time, vision-heavy technology maps neatly onto robots. Some are already circling the programme.

Why it matters

Robotics has never had a shared scale for capability. A warehouse buyer comparing two systems, a safety engineer writing a case for a robot working near people, or a journalist describing the field all improvise, because every finished robot is a mashup of parts from a dozen suppliers. A common vocabulary, if adopted, makes those conversations faster and comparable.

The catch is the convenor's day job. Arm sells the intellectual property underneath much of this hardware, so it is both the host of the ecosystem and a vendor with a stake in what gets built on it. That does not make the framework wrong, but it makes adoption by outsiders, not members invited by Arm, the real test.

Our read

This is Arm acting as an industry convenor rather than merely a chip licensor, and the open invitation to rewrite the levels is the credible part: a fixed vendor taxonomy would have been dead on arrival. Treat the six levels as a draft for argument rather than scripture, and notice that the flattering labels, cognitive and self-optimising, describe robots that barely exist yet. The framework's value will be decided by whether anyone quotes it back.

What to watch

• Adoption in the wild
  Whether any member publishes a product or datasheet described in the framework's own levels, the first honest test.
• Amendments
  How the six levels change once people outside Arm's building get to edit them.
• The member sessions
  What Arm's planned collaboration meetings in the coming months actually produce.
• The automotive crossover
  Whether carmakers formally join, which would confirm O'Driscoll's claim that the two industries' technology maps onto each other.

Discussion spark: Arm sells the silicon and now proposes the scale: would you quote a vendor-run capability level in a buying decision, or does a robot taxonomy only count once a neutral body runs it?

Independent WittyWires tracker for public updates about Arm. Not affiliated with or endorsed by Arm; this is not an official account.