Opinion
The biggest blind spot in autonomous driving is the car itself
“The industry is teaching vehicles to understand the world around them,” writes Amir Hever, CEO and Co-founder of UVeye, “but even the right driving decision can fail when it is based on the wrong assumptions about the machine carrying it out.”
An autonomous vehicle can identify a pedestrian, predict where they will step, calculate a safe braking path and make the right decision in milliseconds - and still get the outcome wrong because the car did not respond as the AI expected.
Perhaps its front camera was not properly recalibrated after a windshield replacement. Perhaps a tire still reports normal pressure but is wearing unevenly after a curb strike altered the alignment. Perhaps damage beneath the vehicle has changed a physical component that the software cannot see. Data can keep flowing, dashboards can stay quiet and the autonomy stack can appear healthy even as the machine underneath it has changed.
That is the uncomfortable blind spot in autonomous driving. We have spent years teaching cars to understand pedestrians, lane markings, traffic lights and the behavior of other drivers, while largely treating the car itself as a constant. Every mile changes a vehicle.
Glass is replaced. Panels are repaired. Sensors are moved. Wheels absorb potholes. Tires lose tread. Suspension geometry drifts. The underbody takes hits that few drivers ever see. Some changes trigger a fault code. Many do not. They accumulate until the real car no longer matches the car the software believes it is controlling.
An autonomous system therefore has to answer two questions: What is happening around the vehicle? And can this particular vehicle, in its current physical condition, execute the response the software has chosen? Today, the first question receives extraordinary attention. The second is too often assumed.
Modern vehicles already know a great deal about themselves. They monitor pressure, temperature, battery health, engine performance and thousands of electronic signals. That telemetry is essential, but it remains a collection of self-reported signals from systems designed to monitor specific conditions. It is not the same as an independently verified picture of the vehicle's physical state.
That distinction matters because autonomy is only as reliable as the link between a software decision and the physical response that follows. A system may correctly decide to brake, turn or accelerate, but the outcome still depends on grip, alignment, sensor calibration and the mechanical condition of the vehicle carrying out the command.
Verified vehicle condition should become a first-class data layer, not maintenance paperwork that sits outside the intelligence stack. That layer begins with consistent, repeatable inspection of what electronic telemetry cannot fully describe: tires, the underbody, exterior components and the physical environment around sensors.
Onboard diagnostics and external condition data provide complementary views. One reports what the vehicle's systems sense internally. The other verifies physical condition from the outside. Together, they can narrow the gap between assumed state and observed reality.
A one-time inspection is useful, but autonomy also needs memory. A time-stamped record across a vehicle's life can show when a change first appeared, whether it is getting worse, whether a repair actually resolved it and whether the same pattern is emerging across similar vehicles.
That is the leap from inspection to condition intelligence. Instead of waiting for a problem to become obvious, operators can distinguish an anomaly from a trend, connect physical changes to specific events and intervene before a small deviation becomes an operational failure. The condition record does not merely document the past. It improves the decisions made next.
The principle extends beyond cars. Robots, trucks, drones and industrial machines may run the same software tomorrow, but they will not be the same machines. They accumulate wear, impacts, environmental exposure and maintenance interventions. Software can be copied exactly from one day to the next. Hardware changes with use.
As AI takes on more responsibility in the physical world, the changing state of the machine becomes part of the intelligence problem. We cannot build trustworthy physical AI on static assumptions about changing hardware.
The next breakthrough in autonomy will not come only from better models or more sensors pointed outward. It will also require a continuously updated, independently verified view of the machine being controlled. Autonomous driving has spent years asking whether a car can understand the road. The overdue question is whether it understands its own condition.
The smartest driving software in the world still runs on glass, rubber, metal, cameras and mechanical systems that change every mile. If the car changes, the data must change with it.
Amir Hever is the co-founder and CEO of UVeye.
