Artificial intelligence can already write, create images, analyze documents and hold conversations. The next stage is much more complicated: letting an AI model receive information from the physical world and decide how a real machine should act within it.
Mercedes-Benz and Britain's Wayve announced a production agreement under which Wayve AI Driver software will be integrated into future Mercedes models. According to the companies, the first cars are expected to begin receiving the technology over the next two years, with advanced assistance in urban and highway driving from the starting point toward the destination.
But despite the name "AI Driver", Mercedes has not announced a car that simply drives for you. The system's automation level has not been published, its intended models are unknown, and at this stage the companies themselves define the future application as a "driving assistance system".
The genuinely interesting part lies beneath the marketing name: Wayve is trying to change how a car learns to drive.
What exactly are Mercedes and Wayve building?
Wayve develops a driving system based on what is called Embodied AI, or artificial intelligence operating through a physical system.
Instead of an AI model receiving a request such as "write me an email" and returning text, it receives information from the car's sensors and must infer how the vehicle should act.
According to Wayve, its AV2.0 system can receive camera and radar information and turn it into driving actions, such as changing direction or slowing down. The company describes a deep-learning model connecting more of the environment-understanding and decision-making process into an end-to-end learned network.
This is an important difference.
For years, advanced driving systems were built as long chains of components: one identifies lanes, another identifies cars and pedestrians, another predicts where they will move, and a planning component decides what the car should do.
Wayve is trying to transfer more of this work to a model learning from driving data.
In simple terms: instead of separately teaching the car thousands of rules for different situations, the idea is to let the model learn driving patterns from large quantities of examples.
This does not mean rules, safety systems or traditional control components disappear. A car is not a chatbot on wheels. It still needs safety mechanisms, vehicle control, redundancy and operating restrictions surrounding the AI model.
Why is it called Embodied AI?
In recent years, most people encountered artificial intelligence through a screen. ChatGPT, Gemini, Claude and image generators work in a digital world: they receive information and generate information.
Embodied AI transfers the same principle to an environment where a model's decision has a physical consequence.
A robot deciding how to lift an object is one example. A car interpreting a road, pedestrian, motorcycle and junction traffic and deciding how to respond is a much more complex example.
This also explains why driving is such a significant test for artificial intelligence. A chatbot answer error can be annoying. An error by a model controlling a car can be a safety issue.
The Wayve-Mercedes partnership is therefore part of a bigger trend: moving from content-generating models to models trying to understand the world, predict how it will change and take actions in it.
Without detailed maps for every street
One of Wayve's central claims is that its approach can reduce dependence on detailed HD maps and systems restricted in advance to a particular geographic area.
Many autonomous systems rely on highly detailed information about their intended operating environment. Launching them in a new city sometimes requires substantial mapping, testing and adaptation work.
Wayve claims its AI system can generalize what it learned and adapt to new cities, vehicles and driving environments without full dedicated training for each city. The company presents this capability as one of AV2.0's main advantages.
This is a significant claim, but needs boundaries.
"Does not require HD maps" does not mean the car uses no navigation or maps. Nor does it mean a system trained in London can be deployed in Israel tomorrow and expected to drive perfectly.
Regulation, signs, infrastructure, weather, driving culture and road-user behavior differ between countries. A system learning to generalize still needs testing and approvals before deployment in a particular market.
What will the car actually be able to do?
Particular care is needed here.
Wayve and Mercedes describe "advanced point-to-point assistance" in urban and highway driving. This means the vision extends beyond adaptive cruise control or keeping a car centered in its lane.
The intention is a system combining navigation information with the driving task across different parts of a route.
But the companies have not yet published a precise list of actions the system will perform.
For example, it has not been officially clarified whether the customer system will independently cross junctions, turn, change lanes, exit highways or handle parking without intervention.
Driver requirements are also unknown. Mercedes and Wayve have not published whether hands may be removed from the wheel, for how long, or how attention monitoring will work.
It is therefore more accurate at present to speak of an AI-based assistance system than a "self-driving car".
Mercedes already has advanced systems. What is new?
This matters particularly because Mercedes already offers other advanced driving technologies.
MB.DRIVE ASSIST PRO, for example, is defined by the manufacturer as an SAE Level 2 system. It combines navigation and driving assistance and can even provide urban assistance from parking toward the destination, but the driver remains responsible. It uses approximately 30 sensors, including ten cameras, five radar sensors and 12 ultrasonic sensors.
By contrast, Mercedes DRIVE PILOT has already received Level 3 approval in Germany under certain conditions. In its approved version, it can operate on highways at up to 95 km/h, and within its approved operating domain it can take on the driving task under defined conditions.
Wayve AI Driver is something else.
It is a future AI platform the companies intend to integrate into production models, but no SAE level, operating conditions or model list has been published for it yet.
The current innovation is not "Mercedes reached a higher autonomy level", but the attempt to bring a driving approach based more heavily on a learned model into the production architecture of one of the world's largest automakers.
AI's big physical-world challenge: edge cases
A driving model can see millions of hours of driving and still encounter a situation it has never seen in exactly the same form.
A vehicle stopping strangely in the middle of the road, temporary infrastructure work, someone gesturing by hand, faded road markings or another driver breaking traffic rules are exactly the situations making real-world driving a difficult problem.
Wayve believes a model learned from large quantities of information can handle new situations better than a system relying mainly on predefined scenarios and rules.
But this remains one of the field's big questions.
A model can be excellent in 99.9% of cases, and in driving, the remaining fraction is precisely what interests safety engineers and regulators.
There is also explainability: when a system of programmed components makes a decision, tracing the chain leading to it is sometimes relatively easy. In a large, complex neural model, explaining exactly why it selected a particular action is harder.
This is one reason moving from onscreen AI models to physical systems requires much more than training a large model.
Mercedes has already integrated the system into its production architecture
The announcement is not merely a general statement of intent.
According to Wayve, AI Driver has already been integrated into Mercedes-Benz's production architecture using Mercedes hardware, MB.OS and its map interfaces. The companies say this work lays the foundation for development toward mass production.
The hardware required in customer cars remains unclear. Camera and radar counts, possible lidar use and the computer running the system have not been published.
The first models have not been revealed either.
When will we see it on the road?
The official target is to begin bringing the technology to customers over the next two years.
This does not mean every Mercedes will get an "AI driver" by 2028 or that the system will launch worldwide simultaneously.
The companies have not yet published the first markets, and ambitions for global deployment are not equivalent to regulatory approval in every country.
At the time of writing, there is also no confirmation that the system will be offered in Israel.
The bigger story is not just cars
It is easy to see this announcement as another chapter in the autonomous car race. But for artificial intelligence, it is interesting for another reason.
For several years, the big race was to build a model better able to understand and generate text, images, voice and video. Now another race is intensifying: to build models able to understand and act in the physical world.
Cars, robots and industrial machines demand something entirely different from AI than a chatbot does. They need to perceive a changing environment in real time, predict what will happen, decide and do so with exceptional speed and reliability.
The Wayve-Mercedes agreement remains far from proof that this problem has been solved.
But it does show the industry's direction: AI is no longer confined to the chat window. More and more models now being developed are ultimately intended to make decisions in the real world.
And for Wayve and Mercedes, that real world is the road.