Physical AI needs a nervous system, not a bigger brain

Physical AI systems such as robots, drones and autonomous vehicles must respond to events in the real world, where waiting for a distant server can be too slow. A drone may need to correct for a gust of wind, while a robot may need to catch a falling part. These actions must happen where the event occurs, often within milliseconds.

In a TechRadar opinion article, Rafael Sotomayor, identified in the article as president and CEO of NXP Semiconductors, argues that intelligence for machines needs to be distributed across different layers. He compares this with the human nervous system: the cerebrum supports reasoning, the cerebellum helps coordinate movement and balance, and the spinal cord handles reflexes.

Sotomayor describes an architecture with a slower reasoning layer, an intermediate coordination layer and a fast reflex layer at the edge. In his warehouse example, a humanoid robot uses reasoning to set its destination and coordination to stay balanced; if an item hits its arm, the reflex layer must quickly rebalance and secure its package.

He also points to Vision-Language-Action (VLA) models as a way to connect perception and movement. Running them within the power and memory limits of edge devices remains an engineering challenge. Sotomayor says progress will require energy-efficient processors for the different layers of robotic intelligence, along with tools and frameworks for developing autonomous intelligence at the edge. He argues that trust is essential because physical actions cannot simply be undone.