During the 2026 Paris Motor Show in October, XPENG will bring its NGP equipped with VLA 2.0 to public roads in Europe for large-scale ride-and-drive experiences. Participants can ride in the XPENG L03, under the supervision of a safety driver, to experience the system's performance in scenarios such as Parisian roundabouts, narrow streets, motorcycles, and pedestrians. This marks the first large-scale public ride-and-drive experience of XPENG's NGP in Europe.
This is not XPENG's first time validating this technology in Europe. According to disclosures from XPENG, on July 14, He Xiaopeng participated in the localization acceptance test of the second-generation VLA model in Germany. By October, XPENG conducted public ride-and-drive experiences in Paris.
The purposes and formats of the two events differ. The German test was part of localization validation, while the Paris experience allowed the media and the public to observe the system's performance on real roads in Paris. From the German test to the public experience in Paris, XPENG is advancing the application of its intelligent driving technology in different road environments.
Behind this lies a question worth pondering: the model has already accumulated extensive driving experience on Chinese roads, so why does it still require continuous localization adaptation and validation upon entering the European market?
01. Why Can't Capabilities Trained on Domestic Roads Be Directly Transplanted?
The experience accumulated by intelligent driving systems domestically does not mean they can be applied to other markets without validation. This is not only because road signs and traffic rules vary across countries, but also because road structures, the behavior of traffic participants, and driving decision requirements in different scenarios differ significantly. Taking roundabouts as an example, the vehicle must not only recognize the road structure and lane markings of the roundabout but also judge the relative positions of vehicles inside and outside the roundabout, the timing to enter, and the path when exiting.
When encountering narrow streets, the system needs to comprehensively consider road width, parked vehicles on the roadside, oncoming traffic, and available passing space. In scenarios where pedestrians, bicycles, and motorcycles mix, the system must continuously judge the movement changes of different traffic participants and decide when to decelerate, wait, or adjust its trajectory. These tasks cannot be solved simply by changing a map.
Maps can provide road structures, routes, and navigation information, but they cannot replace the vehicle's understanding of real-time traffic conditions, nor can they directly dictate what driving actions should be taken in complex scenarios. This also explains why going global with intelligent driving cannot solely focus on whether the model can recognize local road signs.
For intelligent driving systems, environmental understanding must ultimately translate into driving decisions. The system must comply with local traffic rules while combining real-time road conditions to make reasonable choices within the limits permitted by the rules.
02. Where Exactly Does the Generalization Capability of VLA 2.0 Lie?
The NGP brought to Paris by XPENG this time is based on VLA 2.0. According to official disclosures from XPENG, the model supports 4D spatiotemporal understanding, can retain context information for up to about 30 seconds, and can predict how the scene might develop in the next approximately 6 seconds. These capabilities aim to help the system combine previous scene information to judge the potential movement changes of traffic participants.
It should be noted that the 30 seconds refers to the context information the model can retain, not that it completely saves 30 seconds of raw video; the approximately 6 seconds into the future is the prediction time range disclosed officially, and it cannot be directly understood as the system being able to accurately predict traffic conditions 6 seconds later in all scenarios.
These capabilities are highly beneficial for adapting to overseas roads, but they cannot be simply equated. Context information helps the system understand the previous movement states of traffic participants, while spatiotemporal prediction helps judge how vehicles, pedestrians, and other traffic participants might act next.
However, whether the prediction results can support safe decisions depends on whether the system can combine local traffic rules, road relationships, and specific scenarios to translate judgments into appropriate driving actions.
A media participant in the ride-and-drive noted that the vehicle could generally handle local traffic signs, roundabouts, narrow roads, and scenarios involving pedestrians and two-wheeled traffic participants along the test route, but there is still room for improvement in lane-changing decisions and merging into traffic. Public reports also record safety drivers intervening and taking over in edge cases such as temporary construction and bridges.
Furthermore, evaluating the generalization capability of a model cannot solely rely on whether it can complete a trip on an unfamiliar road. Whether the vehicle can recognize the scene is just the starting point; it is also necessary to observe whether it can continuously make reasonable and stable decisions under different traffic conditions, and whether it can adopt appropriate driving strategies when encountering uncertain situations.
Therefore, model generalization capability and localization validation are two related but distinct issues. The former focuses on whether existing capabilities can be transferred to unfamiliar environments, while the latter requires local road testing, scenario validation, and issue iteration to confirm whether the system's performance in a specific market meets practical requirements.
Issues discovered during validation may involve model capabilities, or they may involve driving strategies, system configurations, or local regulatory requirements. In other words, not all issues must be resolved by retraining the model.
03. What Validations Are Still Required from Public Experience to Official Deployment?
Public road experiences allow the outside world to observe the performance of intelligent driving systems in real traffic environments, but this is not equivalent to achieving the same capabilities in mass production.
Preset routes, limited experience durations, and safety driver supervision mean that such activities can only cover partial scenarios and cannot replace more systematic testing.
The European market also involves regulatory compliance and product access requirements. During the Paris Motor Show, XPENG stated that its goal is to become one of the first automakers to meet the requirements of the 02 series of UN ECE Regulation No. 171 and obtain the corresponding approval.
Here is some supplementary information: UN R171 positions DCAS as SAE Level 2 driving assistance, not autonomous driving. The regulation not only emphasizes that the driver is the responsible entity but also translates this responsibility into testable system obligations through hands-on/eyes-on monitoring, HOR→EOR→DCA graded warnings, and ultimately risk mitigation stopping. At the same time, the regulation clearly states that DCAS only assists and never replaces the driver, with no transfer of control responsibility; the 02 series further requires that the system can only operate when the driver is in control, and strengthens hands-off monitoring and safety incident reporting.
Therefore, obtaining UN R171 approval does not mean, in either legal or technical terms, that the vehicle has autonomous driving capabilities, nor does it mean that the driver is exempt from the obligation of continuous supervision and readiness to take over at any time. For driving assistance systems, technical performance, regulatory compliance, and market access are requirements at different levels, and none of them can be replaced by a single public road experience.
Going from China to Europe, the challenges for intelligent driving enterprises are not limited to algorithm-level adaptation. Local road testing, traffic rule adaptation and validation, vehicle configuration, functional boundaries, and regulatory compliance may all affect when and in what form the system enters the local market.
XPENG's previous localization acceptance in Germany, followed by large-scale public experiences in Paris, reflects validation work at different stages, but it still cannot be concluded that the system has covered all road conditions in the European market.
For Chinese intelligent driving enterprises, overseas roads are not only scenarios to test the generalization capability of models but also important environments to discover new problems and improve systems. However, a vehicle completing a trip on an unfamiliar road does not mean it has adapted to local roads; a good performance in a single experience cannot replace the validation of long-term safety and stability.
What going global with intelligent driving truly needs to prove is not just whether the model can cross geographical boundaries, but whether the system can continuously make safe and reasonable decisions under different traffic rules, road structures, and driving environments. This is not only an issue that XPENG needs to continue validating but also an engineering requirement that Chinese intelligent driving enterprises must face when entering the global market.