Recently, XPeng’s second-generation VLA-powered NGP and Tesla’s FSD went head-to-head on the same roads in Amsterdam in what was described as their first global real-world comparison on the same route.
André, a Tesla owner from Germany who participated in the test, shared his assessment of the two systems.
Under the same route, at the same time, and under the same traffic conditions, XPeng NGP and Tesla FSD faced each other in a real-world driving test. The overall result was relatively close, but XPeng showed an advantage in some narrow roads and more complex urban driving scenarios.
Tesla FSD has long held a prominent position in the global advanced driving-assistance market. It has gone through years of development and has accumulated a massive amount of real-world driving data.
For XPeng, performing at a comparable level to such a well-established system is already a significant milestone. What makes the test particularly noteworthy is that it took place on unfamiliar European roads.
How can an intelligent driving system trained primarily on Chinese roads adapt to European streets with relatively little additional local training data? And how can it perform alongside an established local benchmark?
The Amsterdam test highlights a broader question: Can Chinese intelligent driving technology move beyond vehicle exports and become a globally reusable technology platform?
01. XPeng Goes Head-to-Head with a Global Benchmark
Traditional autonomous driving systems are generally built as a series of separate modules.
Perception identifies objects such as vehicles, pedestrians, traffic lights, road signs, and lane markings. Prediction estimates how those objects may move. Planning determines what the vehicle should do next, while the control system converts the decision into steering, acceleration, or braking commands.
This modular architecture has an inherent challenge: the different modules need to communicate with each other through intermediate representations.
Each module effectively has its own "language." Information must be translated from one stage to another, and every translation can potentially introduce information loss or accumulated errors.
This challenge becomes even more apparent when an intelligent driving system moves between countries.
A perception system trained primarily on Chinese roads may be highly familiar with Chinese road signs, traffic lights, lane markings, and driving patterns. When the same system encounters European road layouts, different signs, unfamiliar lane markings, bicycles, trams, and different pedestrian behavior, its perception and decision-making may need additional adaptation.
This is one reason why deploying conventional intelligent driving systems across different regions can require substantial localization work.
The Amsterdam test provides an interesting contrast.
Amsterdam is a particularly demanding environment for urban driving. Its historic streets can be narrow and winding, road layouts are not always standardized, and cars, bicycles, pedestrians, and trams frequently share limited road space. Parked vehicles can further reduce the available driving corridor.
For an intelligent driving system, this creates a difficult combination of perception, prediction, planning, and vehicle-control challenges.
According to the test described by XPeng, its second-generation VLA-powered NGP was able to navigate these unfamiliar conditions with little additional overseas training data.
Tesla FSD, meanwhile, has already been deployed in Europe and has accumulated experience with local road structures, traffic signs, driving behaviors, and road rules.
That makes the comparison particularly interesting.
It was not simply a competition on a neutral test track. XPeng was operating in an unfamiliar environment against a system with significant experience in the local market.
In several situations described in the test, XPeng demonstrated strong spatial planning when the road became particularly narrow or complicated.
For example, when passing temporarily parked vehicles, XPeng was reported to have anticipated the available space and completed the maneuver in a single movement, while the Tesla vehicle reportedly needed to reverse twice before getting through.
The overall performance was close, but the test suggested that XPeng had an advantage in certain challenging low-speed urban scenarios.
02. VLA + World Models: Giving AI a Better Understanding of the Physical World
The most interesting part of the comparison is not simply the driving result. It is the technology architecture behind it.
One of the key ideas behind XPeng's second-generation VLA system is reducing the dependence on traditional intermediate representations.
Conventional VLA systems generally involve a visual input being interpreted through an intermediate language or reasoning layer before generating an action.
XPeng's second-generation VLA takes a different approach by reducing the intermediate language translation step and moving toward a more direct mapping from visual information to driving actions.
In simple terms, instead of following a sequence like:
See → Describe → Understand → Decide → Act
the system aims to move closer to:
See → Understand → Act
This is intended to make the model behave more like an experienced human driver, who does not necessarily describe every object in words before deciding how to steer, brake, accelerate, or yield.
The approach also reflects a broader transition in AI-assisted driving—from simply executing predefined instructions toward building a more comprehensive understanding of the physical world.
Adding Time to the Model
XPeng's second-generation VLA has also introduced longer-term temporal understanding.
The company's Infini-VLA architecture allows the model to retain information from the previous several seconds of driving context, reportedly enabling it to remember events from up to around 30 seconds earlier.
This matters because driving is inherently a continuous process.
A vehicle cannot make every decision based solely on the current camera frame. It needs to understand what happened moments ago and how the surrounding environment is changing.
XPeng has also introduced X-Foresight, which incorporates predictive world-model technology into the VLA architecture.
The system is designed to simulate potential future developments over the next several seconds, allowing the vehicle to consider how pedestrians, vehicles, and other road users may move.
This creates two complementary capabilities:
Memory of the past + prediction of the future.
That is much closer to how human drivers naturally operate.
A driver does not only look at what is happening now. They remember what happened a few moments ago and anticipate what may happen next.
Where Does the Training Data Come From?
This raises another important question.
If overseas real-world driving data is limited, how can a model learn to handle such a wide variety of environments?
XPeng's answer involves world models.
A world model can be understood as an AI-based simulation system that attempts to reproduce the physical behavior of the real world.
Based on the current environment, the model can generate possible future scenarios involving vehicles, pedestrians, road users, and changes in traffic conditions.
Instead of relying exclusively on millions of additional real-world driving scenarios, the system can use simulated but physically consistent scenarios to expand the model's training experience.
According to Liu Xianming, head of XPeng's General Artificial Intelligence Center, the world model and second-generation VLA are not competing technologies. Instead, they provide different training signals that work together to improve the model's understanding of the physical world and its ability to act within it.
This approach could help explain how an intelligent driving system trained primarily on Chinese roads can adapt to roads in other countries.
The core idea is not simply to collect more local data for every market.
Instead, it is to improve the model's ability to understand the underlying physical principles shared across different environments.
Road signs may change.
Lane markings may change.
Road layouts may change.
Driving conventions may change.
But vehicles still have to obey the same basic physical constraints, pedestrians still move through physical space, and vehicles still need to maintain safe distances and trajectories.
If a model can learn these deeper relationships, its ability to generalize across markets could potentially improve.
03. From Selling Cars Overseas to Exporting Technology
For Chinese automakers, international expansion is no longer a new concept.
The first stage of global expansion was largely about exporting vehicles.
The next stage is increasingly about exporting technology.
According to Chinese industry statistics cited in the original report, China exported 8.324 million complete vehicles in 2025, representing year-on-year growth of 29.9%. New-energy vehicle exports reached 2.615 million units, approximately doubling from the previous year.
In the first four months of 2026, China's new-energy vehicle exports reached 1.384 million units, representing year-on-year growth of approximately 120%.
Europe has also become one of the most important overseas markets for Chinese new-energy vehicle manufacturers.
But selling a vehicle overseas is fundamentally different from deploying an intelligent driving system overseas.
A car can be shipped from one country to another.
A driving system has to understand a completely different road environment.
It must deal with different traffic signs, lane structures, road rules, driving behaviors, regulations, weather conditions, and interactions between different road users.
That means the next stage of competition may not simply be about which automaker can sell more vehicles.
It may also be about which technology platform can adapt to more markets with less localization.
The Amsterdam test offers one example of this transition.
Three Major Changes in Global Intelligent Driving
The test highlights at least three broader trends.
First, model generalization is becoming increasingly important.
XPeng's approach attempts to use physical-world understanding and world-model technology to reduce dependence on market-specific real-world driving data.
If this approach continues to prove effective, intelligent driving systems may become easier to adapt to new countries and unfamiliar road environments.
Second, the competitive advantage may be shifting from data volume to model capability.
According to XPeng, its second-generation VLA and world-model approach can reduce overseas real-world data collection requirements by around 70% and shorten market adaptation cycles from approximately 18 months to between six and eight months.
These figures are company claims and should be evaluated against independent testing and deployment data as the technology expands into more markets.
Nevertheless, the underlying direction is significant.
The question is gradually changing from:
"Who has collected the most driving data?"
to:
"Who can build a model that learns and generalizes more effectively?"
Third, Chinese automotive technology is increasingly moving beyond vehicle exports.
If a driving model developed primarily in China can operate effectively on European roads, it suggests that China's new-energy vehicle industry is expanding from exporting products toward exporting technology, software, standards, brands, and broader ecosystems.
This does not mean every market can be entered without localization. Regulations, safety requirements, road infrastructure, mapping systems, and local driving practices remain important factors.
However, the ability of a model to generalize across different physical environments could become a major competitive factor in the global intelligent-driving market.
The Globalization of Chinese Intelligent Driving
China's push toward intelligent connected new-energy vehicles is also increasingly connected with international regulatory and technical standards.
Chinese government departments have set long-term targets for strengthening the country's intelligent connected new-energy vehicle industry and improving its position within the global automotive sector.
International regulation will be equally important.
In June 2026, the United Nations' global technical regulation framework for automated driving systems, ADS GTR, was approved, representing another step toward establishing international regulatory foundations for automated driving technologies.
XPeng has also participated in work related to automated driving regulations through the United Nations World Forum for Harmonization of Vehicle Regulations, or WP.29.
Regulatory frameworks will ultimately determine how advanced driving functions can be deployed across different markets.
Technology alone is not enough.
A globally deployable intelligent driving system needs to combine:
- AI model capability
- Real-world safety performance
- Physical-world understanding
- Data and simulation infrastructure
- Local regulatory compliance
- Vehicle hardware
- Software updates
- International standards
This is why the Amsterdam comparison matters beyond a single road test.
It represents a larger experiment in whether an intelligent driving model can move across borders without being rebuilt from the ground up for every market.
Conclusion: The Next Battle Is About Global Adaptability
The XPeng NGP versus Tesla FSD test in Amsterdam is more than a comparison between two advanced driving systems.
It reflects a larger shift taking place across the global automotive industry.
Tesla has accumulated years of real-world driving experience and a huge volume of data through its global vehicle fleet. XPeng, meanwhile, is attempting to build a different technical path centered on VLA, long-term temporal understanding, and world-model technology.
The Amsterdam test suggests that intelligent driving competition is becoming increasingly focused on how well an AI system understands and generalizes across the physical world.
For Chinese automakers, the first stage of globalization was about putting Chinese-built vehicles on roads around the world.
The next stage may be about taking the underlying technology with them.
A vehicle can be exported relatively easily.
A globally capable intelligent driving system is much harder.
It must understand unfamiliar environments, adapt to different road systems, comply with local regulations, and make reliable decisions in situations it may never have directly encountered during training.
If Chinese-developed intelligent driving systems can continue demonstrating this kind of cross-market adaptability, the global automotive competition may gradually move beyond vehicle manufacturing and sales toward a deeper contest involving AI models, software platforms, technical standards, and physical-world intelligence.
The Amsterdam road test is only one example, and broader independent testing will be needed to determine how consistently these systems perform across different countries and scenarios.
But the direction is becoming clear:
The future of automotive globalization may not simply be about exporting cars. It may be about exporting the intelligence that drives them.