Arm Robotics Capability Framework: RL0 to RL5

The Arm Robotics Capability Framework describes six levels of robot intelligence, from RL0 Reactive to RL5 Self-improving. Released on 8 September 2026, it is a draft vocabulary for describing capability, operating context and supervision. It does not certify robots or require every robot to pursue RL5.[1][2]

Arm introduced the proposal alongside Arm Total Design for Physical AI, an ecosystem program with more than 80 participating companies across AI, software, sensors, compute hardware, virtual platforms and robotics.[1]

How do Arm's robotics capability levels work?

Arm's Robotics Capability Framework uses six labels, RL0 to RL5, but it does not reduce every robot to one score. The proposal pairs those labels with a capability profile covering perception, localization, physical action, decision making, learning, interaction and policy control. One robot can occupy different levels across those areas.[2]

RL0 systems react to immediate inputs, while RL1 systems execute predefined plans.[2] RL2 systems use feedback to adjust how they perform a task. RL3 systems interpret context and intent.[2]

RL4 systems reason across goals and tradeoffs. RL5 systems can refine their behavior over time, with controls for review, rollback, monitoring and human oversight.[2]

Arm's six labels describe system sophistication, but a robot may sit at different levels across different capabilities.[2]

Arm's document gives a warehouse logistics assistant as an example.[2] It puts contextual perception and decision making at RL3. Adaptive mapping and physical action sit at RL2, while scripted interaction sits at RL1.[2]

The example is a use case profile, not a rating for a named commercial robot.[2]

For applied context, our survey of humanoid robots in warehouses and report on the CJ Logistics packing task separate documented work from broader capability claims.

Is Arm's framework a robotics standard?

The Arm Robotics Capability Framework is a draft vocabulary, not an industry standard, product architecture or certification scheme. Arm says it is independent of processor architecture, software stack, AI model, compute location and robot body. It also records operating context, human supervision and assurance assumptions.[1][2][3]

Arm designed the framework to describe what a system can do and the conditions attached to that claim.[2] It does not require Arm processors or tell anyone how to construct a robot.[2]

Arm calls the proposal a starting point and is asking robot makers, customers, integrators and standards bodies to help shape it.[1][3]

The company has not published evidence that buyers or engineers have adopted the terms.[1][2][3]

Arm Robotics Capability Framework questions

What is the Arm Robotics Capability Framework?

It is Arm's draft vocabulary for describing robot capability, operating context, supervision and assurance. It combines six levels with a profile of individual capabilities, so it does not assign one overall intelligence score.[1][2]

What are the RL0 to RL5 robotics levels?

Arm names the levels RL0 Reactive, RL1 Planned, RL2 Adaptive, RL3 Contextual, RL4 Cognitive and RL5 Self-improving. The labels describe increasing system sophistication.[2]

Does RL5 mean a robot is fully autonomous?

No. At RL5, a capability can improve its behavior under defined controls. The framework also records context, human supervision and assurance, and a robot may sit at different levels across its capabilities.[2]

Is Arm's Robotics Capability Framework a standard or certification?

No. Arm presents it as an architecture-independent starting point for discussion. It does not certify products, prescribe a hardware or software design, or require Arm processors.[1][2][3]

Sources

[1] Arm, "Arm brings the ecosystem together to build and define the next phase of physical AI," 8 September 2026. https://newsroom.arm.com/news/arm-total-design-and-robotics-capability-framework-for-physical-ai

[2] Arm, "Robotics Capability Framework," version 1.0, September 2026. https://www.arm.com/-/media/files/pdf/physical-ai/robotics-capability-framework.pdf?rev=b820e79720074955baa5a77b79630844&revision=b820e797-2007-4955-baa5-a77b79630844

[3] Richard Grisenthwaite, Arm, "Robotics Capability Framework: A common foundation for the next era of intelligent robotics," 2026. https://armkeil.blob.core.windows.net/developer/files/pdf/manifesto/robotics-capability-framework-manifesto.pdf