Things That Go Bump In The Daytime

There is no argument that autonomous technology is better at certain things than systems controlled by people. A computer-guided system has only one mission — to stay on the road, avoid object, and reach the end destination. It doesn’t get tired, text, or look out the window. And it can park within a millimeter of a wall or another vehicle without hitting it, and do that every time — as long as everything is working properly.

But technology ages, and as it does it behaves differently. So a sensor that is 95% to 100% accurate when is manufactured, may only be 65% or 35% accurate after years of use. Those numbers can vary greatly, depending upon where a vehicle is used. In areas with high winds and dust, for example, even sensors located inside a vehicle can be impacted by what is basically sandblasting of windows. Getting accurate images through pitted glass is much more difficult than through clear glass. For external sensors, the impact may be much worse.

There are ways to track these changes, basically comparing output to a reference model and recalibrating the sensors and adjusting the systems that interpret that data. But this also points to a fundamentally different way of building a car. Rather than focusing just on the reliability of the individual parts — which is important but not sufficient — the real focus needs to shift to the reliability of the data generated by sensors, external communication, and any other way of triangulating that data.

In this scenario, data becomes the key metric, not the longevity of the individual parts. That has to be programmed into the training algorithms used to make sense of that data. And it has to be implemented in a way that inferencing chips can respond to it, quickly and reliably.

This is extremely complicated stuff, and it’s not entirely clear who is best equipped to manage this transition. The skill sets include data science, system-level engineering, chip design, and various levels of testing and monitoring. In addition, all of that needs to be set in the context of software updates, regulatory changes and localized environmental conditions that may be different from what was originally programmed into these systems.

It’s no wonder that carmakers are pushing back the rollout dates for fully autonomous vehicles. While a new vehicle can be programmed to find its way from point A to point B, there are no guarantees it can reach the same goal several years later. All of that needs to be proven, and so far there is nowhere near enough data compiled across all of these systems to have sufficient confidence in designs.

Aging can be modeled, but it’s not just the aging of a single component that determines a system’s behavior. It’s the aging of that component in the context of other systems for a particular use case in a particular environment, and the number of variables that enter into that picture is infinite.

So even though it’s important for semiconductors to be built to stringent specs, that’s just scratching the surface for reliability. Fundamentally, it’s all about the quality and consistency of data, and that’s a topic most carmakers aren’t willing to discuss.