Top-line view from the three sources and TSO verification conclusion:
Source 1: Focuses on XPENG Robotaxi production, L4 road testing, the VLA 2.0 end-to-end model, and information such as no lidar and no high-definition maps, while also mentioning Guangzhou road-testing approval and the creation of a Robotaxi business unit.
Source 2: Focuses on the technical mechanism of VLA 2.0, saying it removes the “language translation” stage and enables end-to-end generation from visual signals to action commands, while also citing figures such as a 32x ultra-dense computing chain and a 33% reduction in prediction error.
Source 3: Focuses on the strong sales performance of the XPENG GX and rising orders, while also adding that the model is equipped with L4 autonomous-driving hardware, a steer-by-wire chassis, and computing power configurations, linking it to the same technical event chain as XPENG Robotaxi/GX.
TSO verification conclusion: The three sources are aligned on the main theme of “XPENG’s concentrated May release of end-to-end smart-driving and Robotaxi-related information.” The “end-to-end visual-to-action generation” aspect of VLA 2.0 can be cross-supported. However, details about the specific vehicle platform, hardware configuration, order volume, and commercialization progress are mainly provided by single sources and cannot all be cross-verified.
Commonly confirmed facts:
Multiple reports linked XPENG in May 2026 to progress in end-to-end autonomous driving and Robotaxi development.
VLA 2.0 is described as an end-to-end autonomous-driving solution, with the core direction being direct generation of action commands from visual signals.
Robotaxi, L4 autonomous driving, and production/road-testing progress form the main event chain in this coverage.
At least one source explicitly describes a linkage between XPENG’s related business/products, the Robotaxi platform, and smart-driving hardware.
Main differences or divergences:
Technical wording differences:
Source 2 explicitly mentions the removal of the “language translation” stage and emphasizes direct visual-to-action generation.
Source 1 only broadly describes VLA 2.0 as an end-to-end model and does not explain the mechanism in detail.
Source 3 focuses on GX model configuration and order performance, without detailing the VLA 2.0 technical path.
Differences in metrics and parameters:
Source 2 gives specific figures such as a “32x ultra-dense computing chain” and “33% lower prediction error.”
These parameters do not appear in Sources 1 or 3 and cannot be confirmed from the provided materials.
Differences in product/platform framing:
Source 1 discusses Robotaxi production, Guangzhou road-testing approval, and the establishment of a Robotaxi business unit.
Source 3 discusses GX sales strength, L4 hardware, steer-by-wire chassis, and computing configuration, but does not confirm whether this is the same production subject as the Robotaxi item in Source 1.
Unverifiable information:
The scope of “no lidar/no high-definition maps” is mentioned in Source 1 but not corroborated by the other sources.
The details of the “Guangzhou road-testing approval” cannot be confirmed from the given sources.
The “roughly 50,000 orders” claim appears in the title context of Source 3, but whether it is accurate or directly tied to the smart-driving event cannot be confirmed from the given sources.
Background and analysis:
Taken together, the three sources suggest that XPENG’s May narrative was not centered on a single vehicle, but rather on a continuous chain of “end-to-end smart-driving capability — Robotaxi production — hardware platformization.” Source 2 places the technical emphasis on VLA 2.0’s generation path, which is an algorithm/model-layer explanation; Source 1 extends that discussion to Robotaxi testing and production; and Source 3 connects the technical event to GX order performance and hardware configuration, showing a three-layer narrative of “technology launch — product implementation — market performance.”
Strictly speaking, however, the only safe conclusion from cross-verification is that XPENG did intensively promote VLA 2.0, Robotaxi, and L4 autonomous driving in May, and that the “visual-to-action” end-to-end direction is explicitly stated in Source 2. As for the exact implementation path, hardware combination, scope of road-testing approval, and commercialization pace, the available sources differ in granularity and cannot be treated as a fully confirmed single fact chain.
Three-source summary:
Source 1 summary: Reports XPENG Robotaxi production, L4 highway testing, the VLA 2.0 end-to-end model, as well as Guangzhou road-testing approval and the establishment of a Robotaxi business unit; some key points come from a single source.
Source 2 summary: Explains VLA 2.0 in depth, stressing the removal of the language-translation step, direct visual output of action commands, and improved compute/error metrics.
Source 3 summary: Centers on GX popularity and order growth, while also mentioning L4 autonomous-driving hardware, a steer-by-wire chassis, and computing configuration, linking them to the Robotaxi/GX platform.
Conclusion:
Overall, XPENG’s external messaging in May 2026 can be summarized as simultaneous progress in end-to-end smart driving and Robotaxi development. Among these, “VLA 2.0’s visual-to-action end-to-end approach” and “L4 / Robotaxi / production road testing” are the most solidly cross-supported themes. Other content involving hardware combinations, regional permits, order scale, and commercialization pace should currently be treated as source mentions rather than fully confirmed facts.