Arm Pushes Common Robotics Baseline Into Physical AI Buildout

Arm has introduced a physical AI design programme and robotics framework meant to give robot and autonomous-system builders a common technical baseline, AI News reported.

The launch pairs Arm Total Design for Physical AI with a Robotics Capability Framework.

The package is aimed at industries such as mining, agriculture, manufacturing and transport, where machines increasingly need to combine models, sensors, actuators, runtime software and specialised compute inside operational environments.

The commercial prize is large enough to explain Arm’s ecosystem push.

Physical industries represent trillions of dollars in activity, and Arm places the annual compute opportunity at about $200 billion by the 2030s.

Its first partner group spans more than 80 software, hardware and AI organisations.

Named participants include cloud and industrial names such as AWS and Siemens, model and developer players including Hugging Face, Liquid AI and Qwen, chip or platform specialists such as NXP and QNX, and robotics or mobility companies including ECARX, PlusAI, PSYONIC and Unitree Robotics.

The framework addresses a basic engineering gap before those systems move widely from demonstrations to field use.

Robotic platforms do not yet share a common way to describe and compare capability, Arm chief architect Richard Grisenthwaite wrote in an architectural manifesto.

That makes it harder for hardware makers and software developers to integrate components, size workloads and reduce deployment risk.

Arm’s answer borrows the idea of structured levels from driving automation, but applies it to physical AI.

The Robotics Capability Framework maps systems across rising tiers, from reactive machines to context-aware, cognitive and self-improving systems.

Each tier ties real-world uses to machine behaviour, expected outputs and hardware limits.

Those criteria turn the framework into more than a naming exercise.

Safety standards and deterministic operation sit beside power budgets, memory needs, where compute is placed and how much latency the application can tolerate.

That gives engineering teams a way to discuss whether a robot can meet a specific operating requirement rather than only whether it runs a particular model.

Arm built the first baseline with input from the robotics sector.

The contributing group includes consulting and device makers such as McKinsey and Lenovo, field robotics companies such as ANYbotics, Fourier, GALBOT and Gravis Robotics, and additional participants including Anaxi Labs, FMC³ Robotics and Robotec.ai.

Total Design for Physical AI extends a collaboration model Arm has already used in cloud AI infrastructure.

The new programme brings models, virtual platforms, digital twins, sensors, silicon and software stacks into earlier development cycles, so teams can test more of a system before final hardware is available.

Automotive work supplied one proof point for that approach.

Arm demonstrated the method with AWS, Google, HERE, RemotiveLabs and Siemens on an integrated digital cockpit reference solution, allowing engineers to develop and validate complex vehicle code on the Arm Zena CSS platform before physical silicon was ready.

The next step is open-ended rather than a finished standard.

Arm is asking the wider engineering community for technical contributions as physical AI deployments mature, leaving the framework’s value dependent on whether enough robotics builders use it to reduce fragmentation across real machines.