Physical AI brings AI into machines that must sense their surroundings, make decisions, and act in the physical world. The market opportunity is substantial: McKinsey estimates that AI-powered agents and robots could unlock about $2.9 trillion in annual economic value in the U.S. by 2030. However, realizing this opportunity depends on moving from working prototypes to products that can operate reliably, safely, and economically at scale. That exposes a different engineering problem. In a prototype, teams can often add more compute, sensors, cooling, or power to prove that a capability works. In deployment, every added capability must fit within the size, weight, power, and cost budget, often abbreviated to SWaP-C, of the system. Those constraints shape how a robot performs: How long it can operate between charges, how much productive payload it can carry, and how reliably it operates over time. The challenge is not simply adding more intelligence but integrating it effectively into deployable products. A robot must continuously interpret its surroundings, decide how to respond, and translate those decisions into physical action. Doing that requires capabilities from perception and AI inference to planning and motion control to work together, while meeting different compute and timing requirements. System constraints shape intelligence Every physical AI system has a finite design envelope. Processors, batteries, cooling, sensors, actuators, and mechanical components all compete for space, power, thermal headroom, and cost. In a deployable robot, every gram matters. A larger battery may extend operating time but adds weight and takes up space. More compute can enable more advanced AI models, but places greater demands on cooling and energy. Consider a quadruped robot navigating challenging terrain. Much of its onboard energy and weight budget must support the basic demands of locomotion, while the same system needs sensing, perception, navigation, and control to operate autonomously. More sophisticated reasoning cannot be added in isolation. It has to fit alongside motors, batteries, sensors, and other components within the same finite system envelope. This makes physical AI a capability integration exercise. As Arm’s Federico Pecora explains in a recent paper titled The Missing Science of Robotic Systems: Capability Integration in the Model-Centric Era , robot intelligence depends on how perception, planning, control, safety, and orchestration work together across different timing, compute, and memory requirements. The “robot brain” is not only the largest model that runs on an accelerator, but the system architecture that enables these capabilities to work together at the right rates and with the right level of assurance. The tradeoffs also vary significantly by machine. A quadruped navigating rough terrain, a humanoid manipulating objects, and a stationary industrial robot may all be described as robots, but they require different combinations of capabilities—and different levels of compute performance, energy efficiency, responsiveness, and safety. As robotic systems become more capable, the industry needs more consistent ways to describe those capabilities and connect them to the underlying system requirements. That is why we are working with the ecosystem to establish a Robotics Capability Framework , which seeks to establish a common language to define and compare intelligent robotic systems. AI changes the design equation AI is changing how robotic and autonomous systems are designed. These machines increasingly need to interpret changing environments, respond to live data, and operate predictably as conditions change. Doing that efficiently requires compute, software, memory, sensors, and safety functions to be planned together. Multiple workloads operate simultaneously with different timing requirements and priorities. Heterogeneous compute gives engineering teams flexibility to match each workload to the right process