Physical AI’s Hard Problem Isn’t Vision. It’s Knowing Where the Robot Is.

Dr. Behrooz Rezvani has built and sold two semiconductor companies, and last week he published a warning aimed at everyone rushing to put intelligence into robots. The consensus in physical AI is that scale solves the hard problems: bigger world models, more data, more convincing demos. Rezvani argues the part that actually breaks deployments is the one nobody funds, the robot knowing, moment to moment, where it really is. I spent years building a robot that had to be right about that inside a human body, and I think he is naming the most underpriced problem in physical AI today.

The layer that doesn’t demo

Writing in The Robot Report on June 23, Rezvani, the founder and CEO of the physical-AI startup Atomathic, splits the robotics stack into four layers he calls physical AI 2.0: world models, physical state recovery, reasoning, and action. Vision and language live in the layers that give good demos. State recovery is the unglamorous middle. A camera gets blinded by glare, an object hides in shadow, two sensors disagree, and the machine’s estimate of its own situation quietly goes wrong. A machine that starts from a wrong idea of where it is cannot reason its way back to reality. His fix is not more data. It is a dedicated recovery layer built on physics-based constraints and richer sensing, so the system corrects its own state before the clever reasoning ever runs. Atomathic is backed by RTX Ventures and GM Ventures, which tells you defense and automotive already treat this as a real gap, not an academic one.

Why the boring layer is where the companies get built

This is the pattern I keep seeing in robotics, and it is why I bet narrow. The parts that photograph well, a humanoid folding laundry or a world model generating a synthetic kitchen, get the funding rounds and the keynote slots. The parts that decide whether the machine survives contact with a messy real environment do not. NVIDIA’s Cosmos platform can generate endless simulated worlds, and that is genuinely useful, but no volume of simulation fixes a sensor that cannot tell you the truth about the last inch of a grasp. The robots that already win bounded physical tasks earned it the hard way. Surgical robots run computer vision and machine learning on edge processors, on the device, because a robot in an operating room cannot wait on a data center or a clean camera frame. Da Vinci did not beat human hands by being more human-shaped. It won by being engineered around the specific task of controlling more instruments than a surgeon has hands.

What this means if you are building one

If you are a robotics or physical-AI founder, the useful question is not how general your platform sounds in the deck. It is what your machine can still do when the network is gone, the lighting is bad, and the object is not where the model expected it. Defense buyers already write that into their RFIs, asking for untethered robots that run without a connection, because a robot that needs the internet to function is a liability the moment someone cuts the link. That constraint also picks your form factor: a quadruped or a drone is far easier to make autonomous on edge compute than a general-purpose humanoid. Warehouse and industrial-automation founders face a softer version of the same test every shift. Borrow Rezvani’s four layers as a checklist. If your roadmap spends everything on perception and reasoning and nothing on knowing where you are and recovering when you are wrong, you have built a great demo and a fragile product.

Dave Saunders is the founder of Base Reality Group and a Fractional CPO for hard-tech founders. He was a founder and operator at Galen Robotics, where the surgical-robotics platform earned FDA De Novo authorization in 2023, and he managed a 35-patent portfolio licensed from Johns Hopkins. He wrote Founders Who Finish and publishes The Build. More about Dave →