If the Xpeng model is unfamiliar to you, it might quickly be about to alter. China’s main electrical automobile maker goals to turn into a frontrunner in every kind of future know-how, from robots to electrical airplanes and extra.
As for automobiles, it sells the G6 and G9 SUVs, the P7 sedan and the X9 minivan, along with the MONA M03, in China. In the UK and Australia, Xpeng is well-known for the current launch of the G6, and shortly the G9. In the US, you may’t purchase the corporate’s electrical automobiles but, so it is maybe finest recognized for flying automobiles and CES demos.
Xpeng AI Day 2025: seeking to the longer term
AI Day is Xpeng’s annual showcase, and lately passed off in Guangzhou, with the theme 2025 across the idea of Emergency, with a giant give attention to new capabilities for firms’ ‘bodily AI’ techniques – these designed to work together with the bodily world, in automobiles, robots and airplanes. The keynote was VLA 2.0, Xpeng’s vision-to-action mannequin that introduces new capabilities to effectively convert digital camera imaginative and prescient to manage output.
During the occasion, the model introduced the brand new VLA 2.0 mannequin, teased a mass-production robotaxi, demonstrated a next-generation humanoid robotic, and confirmed off the upcoming hybrid tiltrotor flying automotive.
We additionally obtained many updates on beforehand introduced merchandise, plus new data on how Xpeng plans to construct an AI ecosystem with companions and even make its newest AI mannequin open supply.
Xpeng goes on air
Xpeng introduced a brand new tiltrotor plane referred to as the A868 and gave us a number of new details about the Land Aircraft Carrier hybrid launchable helicopter automobile, together with the information that it’s near mass manufacturing.
Watch the video beneath to see how this progressive system works, and to learn extra, take a look at my Xpeng protection in our sister publication, Tom’s Guide.
VLA 2.0: AI mannequin from imaginative and prescient to motion
Xpeng’s VLA 2.0 mannequin takes a vision-to-action method to processing visible information, so as a substitute of translating what its cameras see into textual content after which changing it into management output, the mannequin goes immediately from pixels to driving or manipulation.
This contrasts with the “imaginative and prescient, language, motion” path used, for instance, in Google’s current work on robotics and in autonomous driving analysis. Visual Language Action has an AI mannequin that first describes a scene in pure language after which decides what to do from that description.
This method is versatile and may embrace further data, equivalent to written directions or net context, however provides further processing, energy consumption, and latency that present {hardware} and batteries battle to supply in cell makes use of, equivalent to automobiles and robots.
From imaginative and prescient to motion is extra like quick automated reflexes, and Xpeng makes use of variants of the VLA 2.0 mannequin in automobiles, airplanes and robots, every with its personal particular coaching. For automobiles, Xpeng says it educated the AI ​​on round 100 million movies and runs a distilled mannequin with parameter counts within the billions. By comparability, Xpeng has mentioned its rivals are utilizing fashions with parameter counts within the tens of thousands and thousands.
For this method to work reliably on the street a number of computing energy remains to be wanted and the efficient computing on the gadget is quoted at 2250 TOPS (Tera Operations Per Second, a standard technique to measure the efficiency of AI {hardware}) in client automobiles and as much as 3000 TOPS in Robotaxi and Iron robotic.
By comparability, Tesla’s present Hardware 3 platform has 144 TOPS, and unofficial estimates for the not-yet-released Hardware 4 vary from about 430 to 1,150 TOPS.
Given the restrictions of present power and thermal budgets inside automobiles and cell robots, imaginative and prescient for motion is a practical alternative. It sacrifices some flexibility in comparison with language-in-the-loop techniques, however can supply nice advantages in effectivity, responsiveness, and talent to deploy at scale. Additionally, Xpeng’s method of dedicating a number of computing energy to the issue reveals that it’s critical about making it work now, quite than attempting to optimize it for extra environment friendly fashions or {hardware} sooner or later.
While we await the discharge of VLA 2.0 to customers and real-world suggestions,
VLA 2.0 in automobiles
VLA 2.0 will run on Ultra automobiles – the high-spec Xpeng 2025 fashions with the complete suite of sensors and computing platform already built-in. Using VLA 2.0, Ultra automobiles get what Xpeng calls XNGP, quick for Navigation Guided Pilot. For now, in passenger automobiles, that is equal to SAE (Society of Automotive Engineers) Level 2 driver help, the place the driving force should stay attentive and supervise the AI ​​management.
The system is designed to deal with most driving conditions, together with slim streets, tight spots for parked automobiles, advanced merges, roundabouts and development zones. Lane alternative and hole choice are a part of the coaching, so Xpeng says you may adapt to visitors, negotiate with different drivers, and get better if one thing surprising occurs. It additionally helps parking and low-speed maneuvering that makes use of curb detection and path planning to deal with tight areas and keep away from bumps or scrapes.
The VLA 2.0 rollout will attain new fashions and current Ultra automobiles by way of over-the-air updates, beginning with a small early entry take a look at in China after which a broader rollout as soon as security objectives are met. Xpeng emphasizes that its method relies on low latency, predictable habits and street abilities quite than piling on extra options. For instance, VLA 2.0 is designed to work with primary maps (or none in any respect), to raised deal with street adjustments between map updates.
In the driving force help panorama, Xpeng and Tesla are the principle firms pushing using digital camera imaginative and prescient alone (quite than including sensors like LIDAR) for client automobiles. Older Xpeng automobiles included further sensors, so this can be a daring method with many advantages, but in addition dangers, and we have seen Tesla battle to carry camera-only imaginative and prescient to larger ranges of autonomy. However, primarily based on demo movies, Xpeng’s high-computing technique appears very promising, and having the identical system underpinning a number of platforms ought to assist it scale sooner.
Xpeng additionally introduced a partnership with Volkswagen, which confirmed that it plans to make use of Xpeng’s know-how in its automobiles for the Chinese market. There can also be a partnership between Xpeng and VW for 2 collectively developed automobiles that shall be manufactured and bought in China.
At AI Day 2025, Xpeng additionally introduced three upcoming Robotaxi fashions that might use the identical VLA 2.0 mannequin, however right here they operated in totally driverless autonomous mode in outlined zones. This is equal to SAE degree 4, as firms like Waymo have achieved with their sensor-packed automobiles.
Xpeng says it’s going to use what it learns about passenger automobiles to speed up the validation course of for Robotaxis, with a pilot program deliberate after which mass manufacturing focused for 2026. Passenger automobiles are additionally anticipated to get a totally autonomous mode utilizing the identical Robotaxi {hardware} in choose fashions.
Iron humanoid robotic.
Iron, Xpeng’s next-generation humanoid robotic, leans closely towards anthropomorphism and has options equivalent to a bionic backbone, versatile pores and skin, and 82 levels of freedom (82 independently controllable axes of movement).
There’s additionally a giant emphasis on palms: Xpeng emphasizes a better diploma of freedom than many rivals, aiming for superb manipulation to make them extra helpful in on a regular basis duties and higher suited to duties at present carried out by people with out the necessity for re-equipment.
Iron makes use of three AI fashions collectively: VLA (Vision-Language-Action), VLT (Vision-Language-Task) and VLM (Vision-Language Model) to talk, stroll and work together. All calculations are executed on the robotic itself because of highly effective {hardware} that gives 3000 TOPS of processing. This is similar core {hardware} configuration as Xpeng’s automobiles and once more reveals that the corporate is keen to make use of a number of computing sources to make it work.
Compared to competing robots like Tesla Optimus and the Figure 01, Iron’s wager is on dexterity and processing on the gadget. This is necessary as a result of object manipulation is a key bottleneck for robots doing actual work in the present day, and having all the pieces run on the robotic ought to cut back latency and enhance privateness.
Detailed specs for Iron have but to be launched, however Xpeng is focusing on full-scale mass manufacturing by the top of 2026. By utilizing the identical underlying techniques in robotics and automobiles, Xpeng has a giant benefit in scaling mass manufacturing of core {hardware} and software program. The firm means that preliminary use instances may embrace jobs equivalent to conducting sightseeing excursions, performing retail duties, and back-office logistics.
Xpeng plans for the following 12 months
- December 2025: Pioneer pilot of supervised co-piloting in Ultra automobiles.
- First quarter of 2026: Complete VLA 2.0 implementation on Ultra fashions.
- 2026: Robotaxi metropolis pilot program utilizing solely digital camera imaginative and prescient.
- 2026: Iron models on sale, mass manufacturing goal on the finish of 2026.
- 2026: Mass manufacturing of land plane carriers.
Do you wish to know extra? You can watch Xpeng’s full AI Day 2025 presentation beneath:
XPENG 2025 AI Day: Emergency – YouTube
