Arm has launched Total Design for Physical AI and proposed a Robotics Capability Framework intended to give hardware and software developers a common way to describe increasingly capable machines. The company is convening more than 80 partner organizations across AI, silicon, sensors, software and industrial systems.
The initiative addresses a practical source of friction in robotics: a deployed machine must combine models, runtime software, compute, sensors and actuators while meeting requirements for real-time control, power and safety. Arm argues that shared baselines can reduce integration risk and help projects move from demonstrations into repeatable deployments.
Capability tiers create a shared technical vocabulary
Arm’s framework responds to what chief architect Richard Grisenthwaite describes as the absence of a common method to compare robotic capabilities. It is patterned after the SAE levels used for driving automation, but applies to a broader range of machines.
The proposed tiers progress from reactive systems through context-aware and cognitive machines to self-improving systems. Each level connects practical use cases with expected behavior, outputs and hardware constraints. The framework covers latency, where computing occurs, memory allocation, power limits, deterministic operation and safety standards.
This is a starting point rather than a finished industry standard. Arm developed the initial baseline with sector feedback and is seeking wider technical contributions. Participants named in the source include Anaxi Labs, ANYbotics, FMC³ Robotics, Fourier, GALBOT, Gravis Robotics, Lenovo, McKinsey and Robotec.ai. Their involvement demonstrates breadth, but the source does not say that an independent standards body has adopted the tiers.
Total Design brings components together earlier
Arm Total Design for Physical AI extends a collaborative model already used for cloud AI infrastructure. It brings AI models, virtual platforms, digital twins, sensors, compute silicon and software stacks into a shared development structure. Initial ecosystem participants include AWS, ECARX, Hugging Face, Liquid AI, NXP, PlusAI, PSYONIC, QNX, Qwen, Siemens and Unitree Robotics.
The approach is meant to let component makers work against known assumptions before every physical part exists. That can be especially useful in autonomous transport and robotics, where perception, AI processing, real-time control, safety and energy efficiency must operate as one system.
Arm points to an automotive demonstration developed with AWS, Google, HERE, RemotiveLabs and Siemens. The companies created an integrated digital-cockpit reference environment that allowed engineering teams to develop, test and validate automotive software on the Arm Zena CSS platform before physical silicon was available.
The example shows the intended workflow, not proof that the same arrangement will remove integration problems from every robotic category. Industrial arms, mining equipment and mobile machines face different operating environments, physical loads and safety consequences.
A broad ecosystem still needs measurable adoption
Arm places the opportunity across mining, agriculture, manufacturing and global transport. These physical sectors represent trillions of dollars in economic activity, while the company cites an estimated annual compute opportunity of $200 billion by the 2030s. That estimate is a projection, not realized demand.
The partner list gives Total Design access to model developers, chip suppliers, software specialists and robotics manufacturers. Its value will depend on whether those organizations use the same capability language when specifying products and whether developers can build against the resulting hardware and software profiles without creating another layer of translation.
The framework also has to remain precise as machines advance. Terms such as “cognitive” or “self-improving” can describe very different behavior unless each tier is linked to testable requirements. Arm’s inclusion of latency, memory, power, determinism and safety offers a technical foundation for that work.
For now, the launch establishes a coordinated proposal and an engineering community around it. The decisive evidence will come when robot makers reference the tiers in real systems, suppliers align components to them and deployments show that earlier virtual development reduces cost, delay or integration failures.
