Arm has proposed six levels for describing robot abilities and brought more than 80 companies into its physical AI programme. The framework is a starting point, not an independent safety standard.
New to this? Read it in simple words
- Arm proposed six levels to describe what a robot can do. The levels go from RL0 to RL5.
- More than 80 companies are taking part in Arm’s physical AI programme. They include AWS, Hugging Face, NXP, Siemens, and Unitree.
- Shared words may help buyers compare what different robots can do. But each level needs tests that independent labs can repeat.
- Arm says the framework is a starting point. It is not an independent safety standard.
- Physical AI
- AI that controls machines, such as robots, that sense and act in the real world.
- Framework
- A basic set of ideas or rules that people can use as a guide.
The framework runs from reaction to learning
Arm’s Robotics Capability Framework starts at RL0 and ends at RL5. A low-level system mainly reacts to direct input. Higher levels add more context, planning, learning, and independent improvement.
The six levels are meant to describe ability, not one type of machine. A warehouse arm, delivery robot, or humanoid robot could use the same language. This may make product claims easier to compare.
Arm says the framework is a starting point. It does not set one required chip, model, or robot design. It also does not mean that every company in the programme has adopted the levels for its products.
Arm’s RL0 to RL5 framework is a proposed shared language. Each level still needs repeatable tests and clear safety limits.
Physical AI needs a full system
A robot must sense, decide, and act in the real world. That requires cameras or other sensors, computing hardware, software, motors, and a safe control system. A strong model cannot solve every problem alone.
Arm expanded its Total Design programme to physical AI. More than 80 companies are taking part across cloud services, chips, models, software, sensors, and machines. Named members include AWS, Hugging Face, NXP, Siemens, and Unitree.
The group may reduce repeated engineering work when companies use common ideas and interfaces. However, membership is not proof that two products work together. It is also not an independent safety mark.
A level should answer a practical question
Buyers need to know what a robot can do in their own space. A level should explain the environment, task, time, failure rate, and kind of human help. Without those details, the label may be too broad.
Safety tests must also look at recovery. What happens when an object moves, a sensor fails, or a person enters the work area? A robot that succeeds in a clean demo may still fail in a busy factory or home.
Independent labs should be able to repeat the test. Results should show both success and serious failure cases. Different products can then use the same words without hiding important differences.
Arm has opened a useful discussion about shared robot language. The framework becomes valuable when clear tests follow it. Until then, RL0 to RL5 is a map for future work, not a simple score of safety or quality.
Sources
Every fact in this story comes from the sources below. Open them to check our work.
- 1Primary source · September 8, 2026Arm brings the ecosystem together to define the next phase of physical AI Arm
- 2
- 3
We used Arm’s newsroom and framework pages for the six levels, programme, and member claims. We used independent coverage to check the main announcement. We separate joining the programme from adopting or passing a future standard.