Partnering with Arm on a shared language for robots

Anaxi Labs is an ecosystem partner in Arm's new Robotics Capability Framework — a common way to describe what robots can actually do. Here's what it is, why we contributed, and how it connects to what we build.

Last week Arm launched the Robotics Capability Framework, alongside Arm Total Design for Physical AI. Anaxi Labs is one of the ecosystem partners who contributed ahead of launch, and we joined Arm and Lenovo on the launch panel to talk about why the industry needs it.

The problem: the same words mean different things

Ask five robotics companies what "autonomous" means and you'll get five answers. One company's autonomous robot still needs continuous human supervision; another uses the same word for a system operating independently in a bounded environment.

"Collaborative" is no better. It can mean a safety-rated industrial arm, a mobile robot working near people, or a service robot dealing with the public — three very different machines, three very different risks.

That isn't only a semantics problem; it's structural. When capability claims aren't comparable, systems are hard to compare, deployments hard to plan, and — for the buyers, regulators and insurers who decide whether a robot goes to work — hard to trust.

What the framework is

An architecture-agnostic way to describe what a robot can do, in what context, with how much supervision, and with what assurance. It has two views:

• Levels of intelligence — from reactive, through deliberative, adaptive and contextual, to cognitive and self-improving.

• Capability areas — where that intelligence shows up: perception, manipulation, decision-making, learning, interaction, safety, security, interoperability and more.

Together they describe a robot as a capability profile tied to a task, an environment and a supervision model — not a single marketing label.

Two things we like about it. It describes capability, not implementation, so it doesn't tell anyone how to build a robot. And it's explicitly a starting point — Arm is inviting the wider ecosystem to shape it.

Why we contributed

Anaxi Labs works on the two inputs physical AI runs on:

• Real-world training data. A global contributor network across more than a dozen countries films real everyday and occupational tasks, hands in frame, in real environments. We prioritise by economic value and by the objects robots will actually meet — diverse, not merely large.

• Independent evaluation. We're building an evaluation arena where robot foundation models are compared head-to-head on standardised tasks — first in simulation, then in a physical facility — so a capability claim can be checked by someone other than the company making it.

A shared vocabulary is what makes both useful to anyone beyond ourselves.

The framework's higher levels — contextual, cognitive — are really statements about generalisation: can a system handle objects, rooms and tasks it has never seen? That's a data-coverage question, and it's the one we organise our collection around. A common description of tasks and environments is what lets data from many companies and countries be pooled into the heterogeneous corpus that generalisation actually needs.

And a level only means something if someone can test for it. Today every company grades its own homework, and a success rate measured in one lab often doesn't survive replication in another, on different hardware. The framework gives the industry a shared rubric. We're building the exam.

"The robotics industry remains fragmented, with companies often using different language to define capabilities and requirements. Through a common framework, Arm is rallying the ecosystem to bring greater consistency to those efforts, making it easier for organizations to design, compare and integrate systems. We look forward to working with the broader industry as the framework evolves to support the next generation of robotics." — Kate Shen, Co-founder, Anaxi Labs

What's next

The framework is a first version, and it will evolve. We'll keep contributing — particularly on how each level should be evidenced, and what data it takes to reach the next one.

That's also where we help directly. If you're building robot foundation models, we supply real-world training data covering the tasks, objects and environments your own fleet can't reach — and independent, head-to-head evaluation that shows what your model can actually do. The framework defines the levels; we can help you reach them, and prove it.

For more on the framework, visit Arm's Robotics Capability Framework page.