ASK EEWORLD'S AI ANYTHING: POWERED BY ENGINEERS FOR ENGINEERS

Why Physical AI needs more physics

//

Share

Bookmark

From humanoids catching shifting loads to predicting GPU board failures, Paul Golding of Analog Devices explains how to partition intelligence across a physical system without letting hallucinations touch the hardware.

“Physical AI” is everywhere lately, though nobody seems to agree on exactly where its boundaries are. The phrase grew up in the robotics community, where the physical parameters that matter most are kinematic: navigating a Cartesian world without running into things. Paul Golding, head of physical intelligence at Analog Devices, thinks that framing is too narrow, and after a half hour on the phone with him, I came away with a better mental model for where AI actually belongs in a physical system.

Golding is less interested in defining the term than in where intelligence sits inside a physical system. Today’s foundation models, such as transformers, attention networks, and their relatives, learn primarily from text and images. Sensor data has different rules. “You could point the camera out the window at apples falling from a tree, and it would eventually be able to predict their pathways,” he told me. “But you could just tell it F=ma and be done with it.”

The more consequential failures show up in the data itself. A standard trick for training vision systems is data augmentation: rotate the dog photo slightly, and now you have two dog photos. Try that with inertial measurement data from a human body, and you eventually produce a human standing upside down, cheerfully defying gravity.

That is Golding’s version of physical intelligence: letting the model learn from data, but refusing to let it invent a body that can stand upside down against gravity. The approach has roots in physics-informed machine learning, but it’s now colliding with large reasoning models and their well-documented habit of hallucinating.

Where probabilistic inference stops being acceptable

I asked Golding where a probabilistic guess stops being good enough in a real system. His answer: it’s a spectrum, and two variables set your position on it. The shorter the decision time and the larger the consequence, the more deterministic the technology has to be. A robot in a healthcare setting encountering an unexpected hospital bed needs local reasoning. A robot hand trying to catch a slipping vial of blood needs reflexes, not a chain of thought.

This is territory ADI already occupies at the component level. The company supplies many of the sensors and actuators these systems depend on: IMUs, time-of-flight depth sensors, magnetic turn sensors, and actuators. Golding argues that making the parts confers an advantage in making them intelligent. “Our knowledge of physics is far superior to public knowledge, because we make the parts,” he said. “We know how they work in practice.” That knowledge gets embedded directly into the algorithms sitting closest to the sensor.

The architecture that keeps falling out of this, in humanoids and elsewhere, is a hierarchy Golding describes as systems 0, 1, and 2. System 0 sits closest to the hardware: real-time, physically constrained, deterministic. System 1 handles sensor fusion, the machine equivalent of balancing without thinking about it. System 2 does the slow, contextual planning, where probabilistic reasoning is tolerable because mistakes are recoverable. His example: a pet food company asked ADI to help curate a dataset for humanoids lifting 20 kg bags whose center of mass shifts mid-lift. Coordinating two hands to save the bag is system 1. Deciding how to cross the factory safely is system 2. Nobody’s brain needs to say “don’t drop the bag.”

The same pattern shows up far from robotics. In data center health monitoring, classical statistical processing runs locally on the board, watching for anomalies. Aggregation happens one level up. Only anomalies that cannot be resolved locally get escalated to a reasoning model that can pull historical diagnostics and board manuals into context before deciding whether a GPU board is about to fail. In Golding’s architecture, probabilistic inference diffuses upward through the stack, while determinism concentrates at the bottom. ADI is prototyping these architectures in an internal reasoning lab. And because its parts are already fielded by the millions in everything from data centers to 5G phased array antennas, the same approach extends to telemetry, for system-level diagnostics and predictive maintenance.

Beyond the datasheet

The part of the conversation that should make design engineers sit up came at the end. Golding argues that within roughly three years, AI will be making many system-level partitioning decisions, which means components need to be “AI legible.” Customers are increasingly building world models of their systems, predictive digital twins that know the current state and can forecast the next one. A part that ships as silicon plus a datasheet doesn’t plug into that. A part that ships with its own model does.

Whether datasheets evolve or are supplemented by richer models remains to be seen. Golding’s point is that components will need to arrive with more than silicon and a PDF. As AI-driven design tools build predictive models of entire systems, they will need machine-readable representations of the parts inside them. While he wouldn’t speculate on future plans, he was clear that the shift toward AI-readable models is already underway: “This is real stuff that we’re building. It’s not slideware.”

Golding’s point is that components will need to arrive with more than silicon and a PDF. As AI-driven design tools build predictive models of entire systems, they will need machine-readable representations of the parts inside them

Leave a Reply