At a recent media briefing on Physical AI organized by Arm, Drew Henry, Executive Vice President of Arm's Physical AI Division, explained that the technical challenges faced by Physical AI computing platforms are fundamentally different from those of cloud computing platforms. He stated that such platforms require specialized architecture designs, with "latency" being the most critical metric for understanding Physical AI.
The concept of "Physical AI" was first explicitly proposed in July 2025 during a dialogue between Jensen Huang, founder and CEO of NVIDIA, and Wang Jian, Director of Zhejiang Lab and founder of Alibaba Cloud, who suggested that the next wave of artificial intelligence would be "Physical AI." This sparked widespread interest and discussion within the industry.
According to Drew Henry, Physical AI involves deeply embedding AI into various intelligent devices to achieve tangible, real-world implementation. In other words, integrating AI into actuators, robotic platforms, autonomous vehicle systems, and other self-moving equipment constitutes Physical AI. Henry identifies "latency" as the most crucial metric for Physical AI. Latency refers to the time interval within an electronic system from sensing a signal to executing a physical action. In a car, it could mean the time from detecting an obstacle ahead to applying the brakes; in a robot, it could be the time from observing a target object to moving a robotic arm or the robot itself. In specific scenarios, the computation from "signal perception to control execution" must be completed within microseconds or milliseconds. "This represents a computational challenge entirely different from and far more complex than data center AI," Henry stated.
From Arm's perspective, realizing Physical AI requires a deep understanding of four computational layers. The first layer is perception-driven, focusing on autonomous operation. It concerns perception systems that enable robots or vehicles to "see" their surroundings and make rapid, real-time decisions accordingly. The core requirement for this layer is performing real-time computations within extremely short timeframes, with a key performance indicator being the latency from sensor signal detection to actuator initiation.
The second computational layer is interaction-driven. When a passenger travels in an autonomous vehicle, they still need to interact with it: they might want to check the route and navigation, confirm the trip, or watch a video to pass the time. To provide a smooth experience for passengers inside the vehicle or users interacting with humanoid robots, the interaction system must offer corresponding computational support, necessitating a specially designed interaction computing layer. This layer does not require the strong real-time performance of the perception layer and employs a different computational architecture.
The third computational layer is the actuation layer. It is responsible for precisely controlling the various micro-actuators in a robot's hand, as well as managing the control and execution of braking and steering systems in autonomous vehicles. This system comprises numerous micro-devices that require unified coordination and scheduling from upper-level systems, making the overall design extremely complex.
The fourth computational layer is the cloud layer, primarily aimed at enabling interaction between humanoid robots, autonomous systems, and robotic platforms with cloud environments. On one hand, users can train new models in the cloud and then download them to these terminal devices; on the other hand, all devices can be integrated via the cloud into a cluster for cooperative operation.
Furthermore, Drew Henry mentioned that these systems must ensure functional safety and information security. The technical challenges for Physical AI computing platforms are fundamentally different from those of cloud computing platforms.
Regarding whether the embodied intelligence industry or IC companies will be the stronger driving force for technological change, Henry indicated that over the next decade, embodied intelligence and its required models will inevitably continue to evolve. The application demands for humanoid robot platforms and autonomous driving platforms will also constantly change, with both models and requirements undergoing successive generations of upgrades. Each product generation will demand higher performance, energy efficiency, and cost-effectiveness. As the industry's exploration of pathways to achieve embodied intelligence deepens, the related workloads and models will continue to be optimized and adjusted. This is a computational challenge that requires dedicated effort for ten years or more, which is precisely why this field has the potential to become one of the largest markets ever.
In the field of Physical AI, because perception-driven intelligent systems—the process from sensors collecting input data to converting it into actual device actions—must be completed within microseconds or even milliseconds, it means the systems designed are not primarily focused on ultimate performance and ultra-high memory bandwidth. Instead, the core design principle is to achieve the fastest and most efficient command execution within an extremely short time, creating an immediate closed loop from data input to action output. This presents a class of computational challenges entirely distinct from those faced by cloud-oriented computing platforms, belonging to a different category of computing platform altogether, thus requiring specialized architectural designs for Physical AI. This distinct set of challenges is expected to drive systemic architectural transformations over the next decade.
Comments