physical AI systems cannot rely on large-scale reasoning alone, argues the president and CEO of NXP Semiconductors. Robots, drones and autonomous vehicles also need fast, dependable responses at the point where events occur, without waiting for a distant server.
The article contrasts tasks such as solving equations, which machines can handle well, with everyday physical skills like balancing, grasping objects and avoiding obstacles. In the real world, delays and errors can have consequences that cannot simply be corrected or reversed, making low latency, energy efficiency and reliability essential.
The proposed model takes inspiration from the human nervous system, which distributes different functions across the brain and spinal cord. For machines, the author describes three coordinated layers: a reasoning system for higher-level decisions, a coordination layer for movement, and a fast reflex layer operating at the edge.
In the example of a humanoid carrying a package through a warehouse, the reasoning layer chooses where to go, while the coordination layer helps it walk steadily. If an object hits its arm, an edge-based reflex system would need to react within milliseconds to restore balance and protect the package.
The article also says physical AI must learn to anticipate how objects behave, including their weight, momentum and friction, rather than merely react to what it sees. Vision-Language-Action models are presented as one way to connect perception, instructions and movement, though running them within the power and memory limits of edge devices remains an engineering challenge.
The author argues that progress will depend on processors suited to each layer of robotic intelligence, alongside tools that make it easier to develop autonomous systems for edge devices. Trust, the article concludes, must be built into the system from the silicon up so machines can behave predictably and respond safely to unexpected situations.
Sources: TechRadar.