Arm Proposes RL0–RL5 Robot Capability Levels With 80+ Physical AI Partners

What Arm announced

Arm expanded Total Design into physical AI with more than 80 companies across the technology stack. Participants include AWS, Hugging Face, NXP, Siemens, Unitree Robotics, QNX, Qwen, PSYONIC and others spanning AI models, software, sensors, compute and digital twins.

The second component is a Robotics Capability Framework. Arm wants a shared vocabulary for describing and comparing robot capabilities, similar in spirit to how SAE levels organized discussion around driving automation. The framework links robot behavior to system requirements such as latency, compute placement, memory, power, determinism and safety.

RL0–RL5: what the levels mean

The initial structure runs from RL0 reactive systems to RL5 self-improving systems, with increasingly context-aware, autonomous and cognitive behavior between those endpoints. Arm does not present this as an already adopted standard; it explicitly calls it a starting point to be shaped with the ecosystem.

That distinction matters because terms such as autonomous, physical AI and general-purpose are currently used very loosely. Two robots described as autonomous may rely on very different levels of teleoperation, exception handling and environmental generalization.

Why 80+ partners matter

A capability framework would be easy to dismiss as another manifesto without broad participation. Arm is bringing together companies from semiconductors and real-time operating systems through robot manufacturers and model providers. That breadth increases the chance of common metrics or reference implementations becoming useful beyond a single vendor.

However, participation in Total Design does not mean every company has formally adopted RL0–RL5 for its own products. Collaboration and ecosystem membership are not equivalent to certification.

The business logic for Arm

Arm also sees physical AI as a compute market. The company estimates an approximately $200 billion annual compute opportunity in the 2030s. That is Arm’s own estimate, not an independent market forecast.

Strategically, Arm wants to provide the underlying compute platform for systems that combine perception, AI inference, planning, real-time control and safety. If robotics moves toward more standardized compute architectures, Arm can benefit regardless of which humanoid brand leads final hardware sales.

What is confirmed and what is not

• Confirmed: more than 80 companies are joining Total Design for Physical AI.
• Confirmed: Arm is introducing a Robotics Capability Framework as a starting point for common language.
• Confirmed: the framework includes behavior plus compute, latency, memory, power, determinism and safety requirements.
• Not confirmed: RL0–RL5 is not currently an ISO, IEC or SAE standard.
• Unknown: which vendors will label products using these levels and whether an external testing process will emerge.

Why Europe and RoboMorrow should care

If the framework gains adoption, it could become editorially useful. Instead of repeating a manufacturer’s claim that a robot is autonomous, publications could use more structured capability language. RoboMorrow should not assign RL0–RL5 labels independently until criteria are stable and verifiable.

For Europe, participation from companies such as Siemens and NXP matters because it creates a path to influence physical-AI evaluation language early. For Polish integrators, a common framework could eventually improve procurement comparisons, but it is too early to use RL0–RL5 as a formal purchasing specification.

What to watch next

The key next steps are a detailed methodology, first vendors labeling products according to the framework and any involvement from standards bodies. Without those developments, RL0–RL5 remains a useful discussion tool rather than a standard.

RoboMorrow treats this as a physical-AI infrastructure and standardization story, not a robot ranking.

The hardest problem is measurement, not naming

A level system becomes useful only when every label is backed by a test procedure. Criteria need to address what share of a task is completed without human help, what happens during exceptions, how often teleoperation is required, whether a skill transfers to a new object and how the system behaves outside familiar data distributions. Without such rules, two vendors could claim the same level for fundamentally different capabilities.

Domain scope is another challenge. A warehouse robot can be highly autonomous within one narrow workflow and useless when moved into a kitchen or workshop. A single global number can therefore hide differences between task autonomy, mobility, manipulation and adaptation. A mature framework will need either separate dimensions or a very precise definition of the operating domain covered by the level.

Potential impact on RoboMorrow’s Robot Database

If Arm and its partners publish stable criteria and independent testing emerges, the framework could become a useful Robot Database field. For now, RoboMorrow will not convert existing profiles into RL0–RL5 scores based on Marketing claims. Without an official classification or measurable evidence, doing so would create false precision.

In the short term, the better approach remains reporting what can actually be verified: whether a robot works autonomously or with teleoperation, its environmental limits, how it handles exceptions and where human responsibility remains. If RL0–RL5 eventually organizes those facts into a comparable system, it could materially improve specifications and procurement.