AI has learned to drive “like a human” – and is already doing it better than humans

Chinese startup ZYT showed an unusual approach to autopilot: their AI is not programmed according to rules and maps, but is “taught to see the world” through a huge amount of video, notes xrust. As a result, the system already confidently drives through difficult city streets and, according to the developers, controls the car better than a person — even in heavy traffic and near pedestrians.

What exactly did they do

The main idea of ZYT is to abandon the classic autopilot architecture. Typically, such systems consist of many separate blocks: one recognizes pedestrians, another — signs, a third — markings, and a fourth — builds a route. All this requires fine tuning for a specific city, country, and even driving style.

Here is a different approach. Instead of a set of modules, a single model is used that learns everything at once. It receives video data and itself develops an “understanding” of the traffic situation — in much the same way as a person behind the wheel does.

Key features of the technology

• Training without strict rules
  The system is not “taught” to recognize objects separately. She herself forms an idea of ​​​​what is happening on the road.
• The widest possible set of data
  Not only records from cars, but also video from drones, robots, motorcycles, household devices and even cameras that are simply carried in the hands. This gives the AI ​​more «real life» experience.
• Universality instead of locality
  Unlike the classic autopilots, the model is not tied to specific roads. It does not need to be trained separately for each city.
• Quick adaptation to different types of transport
  The technology, trained on passenger cars, was able to be transferred to trucks in a matter of weeks — without completely reworking the system.
• Economic effect already now
  In cargo transportation, AI can reduce fuel consumption by a few percent — and this is direct money.
• Working in difficult conditions
  The system already demonstrates confident behavior on narrow roads, in oncoming traffic and next to pedestrians — some of the most difficult scenarios.
• Black box effect
  Developers admit: they don’t always understand exactly how AI makes decisions. This is a sign of a new generation of models — powerful, but not completely transparent.

Why is this different from everything that came before class=»notranslate»>__GTAG20__

Traditional autopilots are essentially engineered systems with AI elements. They require a huge amount of manual configuration: maps, scenarios, rules of behavior. Therefore, such solutions work well only in pre-prepared conditions.

The ZYT approach is closer to modern neural networks: less manual logic, more self-learning. This makes the system potentially scalable — it can be quickly transferred between countries, machines and tasks.

To simplify: before the autopilot “knew the rules”, now it begins to “understand the world”.

Where will this be used first

Although everyone is waiting for driverless passenger cars, the main focus now is on freight transport. The reason is simple: the economic effect is visible faster there.

Even small fuel savings on long routes give significant money. Plus — less influence of the human factor, stable behavior on the highway and the ability to optimize logistics.

Therefore, trucks may become the first mass market for such systems.

Limitations and problems

class=»notranslate»>__GTAG9__ Despite the progress, the technology is not yet ready for mass implementation:

• Dear “hardware”
  Now the system requires powerful computing platforms, which are installed in robotaxis and prototypes, but not in ordinary cars.
• Optimization is still underway
  The developers are working to port the model to cheaper chips.
• Opacity of decisions
  If the AI acts like a “black box”, it creates safety and regulatory issues.
• Regulatory barriers
  Entering global markets will require compliance with various requirements and laws

When to wait in regular cars

According to the developers' plans, the first cars with this system may appear closer to 2027. The technology is currently being tested and refined, including on European roads.

Context: why this is important

We are seeing a paradigm shift in autopilots. The industry is moving away from “programmed” systems towards general purpose AI models that learn from data and adapt themselves.

This can lead to three key consequences:

1. Sharp acceleration in the implementation of drones
   If the model does not need to be trained for each city, scaling becomes easier.
2. Reducing the cost of technology
   After optimization for mass-produced chips such systems can find their way into regular cars.
3. Increasing global competition
   Even an advantage of several months is already gives a serious lead — the market is forming right now.

Based on materials from Reuters.

Xrust AI has learned to drive “like a human” — and is already doing it better than people