On September 8, 2026, Arm expanded Total Design into Physical AI and put a Robotics Capability Framework on the table as one of the first shared projects in that circle. If you build robots, vehicles, or the software that sits between sensors and actuators, this is the announcement worth reading carefully.
A little context helps, then I will stay on this story. At CES in January 2026 Arm stood up a dedicated Physical AI business so automotive and a wider robotics mandate had a clear home. Arm licenses IP; partners ship the silicon and the machines. September 8 is the ecosystem follow-through: how those partners are supposed to build together on Arm without reinventing the same integration work alone.
What Total Design for Physical AI actually is
Total Design started as a collaborative model for cloud AI. The Physical AI edition pulls the same idea into autonomous systems. Arm is convening partners across software stacks, AI models, sensors, compute hardware, virtual platforms, and digital twins so OEMs and startups can reduce integration risk and move from proof of concept toward something they can deploy.
That matters because physical systems fail in the seams. Perception, planning, real-time control, safety, and power budgets all have to land on one board story. A shared platform for partners does not erase differentiation. It gives everyone a place to meet before the expensive surprises show up late in a program.
Who showed up
Arm says more than eighty companies are joining Total Design for Physical AI. The public list spans cloud, models, silicon, industrial software, and robot makers: AWS, Hugging Face, NXP, Siemens, Unitree Robotics, QNX, Liquid AI, PlusAI, PSYONIC, and many others. I am not going to turn this into a roll call. The useful signal is breadth. Model labs, RTOS vendors, chip partners, and robot OEMs are in the same room, which is exactly where integration pain usually hides.
A six-level vocabulary for robot sophistication
Alongside the ecosystem move, Arm introduced a Robotics Capability Framework. The goal is a common way to describe how capable a robotic system is, tied to real constraints builders already argue about: latency, where compute lives, memory and power, determinism, and safety.
Arm’s chief architect framed the need the way many of us feel it in product reviews. Robotics is moving fast, and teams still lack a shared way to compare machines. SAE Levels gave driving automation a vocabulary people could argue with. This framework aims for a similar spirit for robots: six levels of sophistication, from more reactive systems toward context-aware, cognitive, and self-improving ones, always connected back to use cases and system requirements.
Read that carefully. The framework is an evolving starting point shaped with partners across the robotics community. It is a shared vocabulary in progress. It is still short of a finished, certified industry standard in the SAE sense. Use it to clarify what a demo claims, what a plant floor needs, and where compute and safety actually sit. Do not treat a level label as a compliance badge.
Arm has invited the broader industry to help shape the next versions. That invitation is the product signal. Frameworks that stay useful usually grow in public, with OEMs and vendors correcting the language against real deployments.
Why builders and product people should care
If you are evaluating a robot, an AMR, a humanoid pilot, or a vehicle stack that leans on Arm partners, shared language saves months of translation. Procurement conversations get clearer when “capable” maps to latency, determinism, and safety requirements instead of a demo reel. Architecture reviews get clearer when compute placement and power are part of the same sentence as the model story. Supplier conversations get clearer when eighty-plus companies are already practicing inside one Total Design circle.
I keep a simple test for announcements like this. Can I take one robot I care about and place it on the framework using the words Arm actually published? Can I name which partners in the Total Design set would touch perception, control, or safety for that machine? If those two answers get easier after reading the primary pages, the announcement earned its keep.
Dig into the primary pages
Start with Arm’s Total Design for Physical AI announcement, then sit with the Robotics Capability Framework page and the partner perspectives there. Map one real system onto the six levels using latency, compute placement, memory and power, determinism, and safety. Bring that map to your next architecture or procurement meeting. The useful work begins when the vocabulary leaves the press release and shows up on your whiteboard.