A new approach to self-driving cars: When AI starts to reason

Views: 0

For years, the development of self-driving cars has followed a fairly clear logic. In simplified terms, vehicles have relied on enormous systems of rules and predefined scenarios. The car “sees” its surroundings with sensors, cameras, and radar, compares that information with a large internal library of known situations, and, based on that, decides what to do: brake, accelerate, turn, or stop.

This approach has worked reasonably well in most cases. Traffic is, to a certain extent, predictable. Traffic lights, signs, pedestrians, lanes… all of that can be modeled and trained with millions of kilometers of data.

The problem arises when something unforeseen happens. When the real world deviates from the script.

A very illustrative example occurred in San Francisco in December 2025, when a failure knocked out all the traffic lights in several areas of the city. For human drivers, it was an awkward but manageable situation. For Waymo's self-driving cars, it wasn't.

Many of them came to a complete stop. Not because they were broken down, but because their system couldn't find a valid response in its database of scenarios. Faced with the unexpected, the safe option was to stop.

This highlights a significant limitation: current systems react, but they don't reason like a person.

This is where NVIDIA's new proposal, called Alpamayo, comes in. Instead of relying exclusively on predefined rules, this system proposes something different: that the car reasons using artificial intelligence.

The idea is ambitious. Alpamayo uses sensors to observe the environment, but instead of simply consulting a library of cases, it employs an AI model capable of learning from the context, from how human drivers behave, and from how traffic evolves in real-world situations.

In theory, over time the system not only improves its driving, but can also explain why it makes a particular decision. This is key, both from a safety perspective and for future regulation.

A human driver doesn't need to have seen an exact situation before to react. They use experience, intuition, and reasoning. Alpamayo attempts to approximate that model.

If traffic lights stop working, a human observes the behavior of others, identifies implicit priorities, and acts cautiously. NVIDIA's proposal is for the autonomous car to do something similar, learning from what it sees and adapting in real time.

It's a radically different approach, and also much more complex. Reasoning, learning from mistakes, and generalizing situations are among the major current challenges of artificial intelligence.

NVIDIA has already confirmed that Mercedes-Benz will use Alpamayo technology in its future CLA models. This makes the proposal more than just a laboratory experiment.

From NVIDIA's perspective, the strategy is very clear. The company is a world leader in AI chips and can offer its chips in extremely powerful computers designed for integration into vehicles. This is how they will continue selling, not only for autonomous driving, but for all the car's systems: driver assistance, energy management, infotainment, and safety.

It is becoming increasingly clear that the car of the future, simply put, will be a battery, a very powerful computer, and a software system on wheels. The engine remains important, but the real differentiating factor will be software and AI.

That's why it's no surprise that companies like NVIDIA want to position themselves in this market. Tens of millions of cars are sold every year, and each one will essentially be a mobile computing node.

Lots of promises, but still few certainties. Reasoning like a human isn't easy. For now, NVIDIA has explained what it wants to do and has shown promising demonstrations. As they say, anything can be written on paper.

You can watch a video about Alpamayo here:  https://youtu.be/EiEPeVNfZDE

We'll have to see how Alpamayo performs in the real world, with chaotic traffic, unpredictable drivers, and extreme situations. But if it works, it could be a game-changer for autonomous vehicles. And other companies will surely try similar approaches; there's a lot at stake.

As always, time will tell.