Arm has expanded its Total Design ecosystem into physical AI and introduced a proposed Robotics Capability Framework intended to give the industry a more consistent way to describe what robots can actually do.
The initiative, announced on September 8, brings together more than 80 companies from across the robotics and computing supply chain. Participants named by Arm include AWS, Hugging Face, NXP, QNX, Siemens, Unitree Robotics and a range of robot makers, chip companies and software specialists.
The framework is not a finished international standard, certification scheme or safety rule. Arm describes it as a starting point and an invitation for broader industry collaboration. That qualification is important: robotics still lacks a universally accepted vocabulary for comparing systems that may look similar but differ sharply in autonomy, adaptability and human supervision.
Why physical AI needs a common language
Robots are increasingly being marketed as “intelligent,” “autonomous” or “general purpose,” yet those terms often conceal very different capabilities. One machine may repeat a programmed movement in a fixed workcell. Another may recognize objects, plan around obstacles and adapt its actions to a changing environment. Both can be described as autonomous even though their technical and operational requirements are far apart.
Arm’s proposal aims to connect a robot’s intended use with observable behavior, outputs and system requirements. The company’s official announcement says the framework spans a progression from reactive machines to context-aware, cognitive and eventually self-improving systems.
A common vocabulary could help buyers ask better questions and help developers design hardware around clearer workloads. It could also make it harder to compare a tightly scripted demonstration with a system that has been tested under variable real-world conditions.
What the framework measures
The proposed model looks beyond a single headline metric such as AI processing speed. Arm says capability definitions should connect behavior to factors including latency, the location of computing, memory, power consumption, determinism and safety.
Those dimensions matter because physical AI operates under constraints that cloud software can often avoid. A robot may have milliseconds to react to a moving person, limited energy to complete a shift and strict requirements for predictable behavior. Moving more computation onto the robot can reduce dependence on connectivity, but it also increases power, cooling and cost challenges.
Determinism is particularly important in industrial and safety-sensitive settings. A useful robot must do more than produce a plausible answer: it must execute actions within known limits and respond reliably when conditions change. Capability labels that ignore those requirements can make systems appear more mature than they are.
An ecosystem around shared hardware foundations
Arm Total Design began as a collaborative model for companies building custom silicon around Arm’s compute platforms. Extending that approach to physical AI suggests that Arm wants to align chip design, operating systems, AI models, simulation tools and robot hardware earlier in the development process.
The initial framework structure was informed by organizations including ANYbotics, Fourier, GALBOT, Lenovo, McKinsey and Robotec.ai, according to Arm. That mix reflects the breadth of the problem: industrial inspection robots, humanoids and autonomous machines may share computing needs, but their operating environments and acceptable risks differ.
Arm estimates that physical AI could represent an annual computing opportunity of more than $200 billion in the 2030s. That is the company’s projection, not a measured market outcome. The more immediate significance is strategic: the company is positioning its architecture and partner ecosystem as a foundation for a robotics market that is still deciding how systems should be described and evaluated.
Not the first attempt—and not yet the final answer
Robotics researchers and standards bodies have already proposed taxonomies for autonomy, safety and performance. Independent analysis from Fierce Sensors notes that organizations including NIST, ISO, ACM and consulting groups have approached parts of the same classification problem.
That means Arm’s framework will need industry testing and alignment rather than simple adoption by announcement. Different robot classes may also require different evidence. A warehouse vehicle, surgical assistant, home robot and humanoid factory worker cannot be evaluated by a single generic demonstration.
Still, an imperfect shared framework can be useful if it encourages vendors to disclose assumptions, supervision needs and operating limits. The best outcome would be a language that complements existing safety standards and makes capability claims easier to compare rather than creating another competing label system.
Why this matters for embodied AI
Physical AI is moving rapidly from a conference theme into product road maps. RoviVox’s coverage of robotics at IFA 2026 showed how companies are trying to bring AI-driven machines into homes and daily environments. The challenge is no longer only whether models can perceive and plan, but whether complete systems can act safely and consistently.
The same issue appears in humanoid research platforms such as Unitree H2 Plus with NVIDIA GR00T. Shared models and hardware can accelerate experimentation, but progress is difficult to measure when every developer uses different terms, tasks and test conditions.
A capability framework cannot solve reliability or safety by itself. It can, however, clarify which questions need answers and expose the gap between a demonstration, a supervised deployment and a genuinely adaptive system.
RoviVox View
Arm’s proposal matters because robotics is reaching the stage where vocabulary can influence investment, procurement and regulation. Clear capability levels could help the market reward demonstrated performance instead of broad claims about intelligence or autonomy.
The risk is fragmentation. If every chip vendor, robot maker and standards group publishes a different scale, buyers may face more labels rather than more clarity. Arm will need open participation and compatibility with established safety work for the framework to become genuinely useful.
Watch for concrete evaluation methods, reference workloads and documented case studies. The framework will gain credibility only if companies can use it to compare real systems—and if it clearly separates reactive automation, supervised autonomy and self-improving behavior.