XPeng plans to take its self-driving software worldwide in 2027, its founder and chief executive said on Friday, days after the UN cleared the first global rules for driverless cars.
He Xiaopeng tied the timing to this week’s vote in Geneva, where the UN’s WP.29 forum approved a framework for fully autonomous systems.
Posting on LinkedIn, the XPeng founder said its second-generation system, VLA 2.0, would reach overseas drivers in 2027 with voice control in mixed Chinese and local languages.
The regulation removes a barrier that had long confined many self-driving systems to a single market.
Global Rules
WP.29 adopted the world’s first regulation for Automated Driving Systems, or ADS, on June 24, the body said.
The framework covers vehicles from Level 3 to Level 5 and rests on a so-called safety-case approach rather than fixed tests, according to the UN Economic Commission for Europe.
Backers include the US, China, the EU, Japan, the UK and Canada, the commission said.
Early forecasts of widespread automated driving had failed to materialize over the past decade, the commission said.
Manufacturers must show their systems pose no unreasonable risk and perform at least as well as a competent human driver, under the rules.
Vehicles will have to record and store safety-relevant data and undergo continuous monitoring once deployed, under the framework.
Richard Damm, who chairs the UN working group that drafted the regulation, said the framework was “not a compromise on safety.”
Entry into force is expected in January 2027, according to the group.
China has said it will base a national standard on the global text, the commission said.
Separately, the forum amended about 90 existing regulations so they apply to cars without steering wheels or pedals.
A related update to Level 2 driver-assistance rules, known as DCAS, is due to become mandatory in the European Union within about six months.
Under DCAS, the driver remains responsible even when the car steers itself in cities.
XPeng joined the process, presenting its software to UN delegates at a working-group session in Shanghai in February.
Vision to Action
VLA 2.0, unveiled in November 2025, maps camera input directly to driving commands, skipping the language-processing step of earlier systems, XPeng said.
The design cut response time to below 80 milliseconds and improved driving efficiency by 23%, according to the company.
A companion Vision-Language Model, or VLM, handles scene reasoning and the multilingual voice features planned for export markets, the company said.
Unlike modular systems, the model combines perception, reasoning and action in a single network, XPeng said.
Xianming Liu, who leads the company’s General Intelligence Center, has described language as “poison” for driving models.
Designed in-house, XPeng‘s Turing chips deliver up to 2,250 trillion operations a second.
Training used nearly 100 million video clips of difficult driving, mostly from Chinese roads, according to the company.
Management describes VLA 2.0 as its first mass-produced model trained to interpret the physical world.
Volkswagen AG agreed to become the first foreign carmaker to license the software and also selected the Turing chip.
Tesla Benchmark
He Xiaopeng has set a goal of matching Tesla Inc.’s Full Self-Driving system by Aug. 30.
After a five-hour drive in San Francisco, he praised the US system before setting the deadline for his team.
Liu has said XPeng already reached parity with FSD version 13 on narrow Chinese roads.
In a personal bet, Liu said he would run across the Golden Gate Bridge if XPeng misses the target, telling Electrek the goal was already met in testing.
XPeng spends about 300 million yuan ($42 million) a month on model training, or roughly $500 million a year, Liu told the outlet.
Tesla charges $99 a month for Full Self-Driving in the US, while XPeng includes VLA 2.0 in the vehicle price.
Cameras drive the main system, with radar and ultrasonic sensors kept only as a backup, Liu said.
Chinese roads offer a richer source of difficult cases than other markets, Liu said, calling it a training advantage.
Physical AI
VLA 2.0 also underpins XPeng‘s expansion into robotaxis, humanoid robots and flying cars, the company has said.
The company began producing its first robotaxi in May on its GX platform, fitted with four Turing chips and about 3,000 TOPS of compute.
Three robotaxi variants are planned on the platform, in five-, six- and seven-seat versions.
Data from XPeng‘s passenger fleet helps train the robotaxi, which shares the same core software, the company said.
Passenger pilots are scheduled for the second half of 2026, with driverless operation targeted for early 2027, according to XPeng.
Built around cameras rather than lidar or high-definition maps, the robotaxi is suited to cross-border use, the company said.
Shared software runs XPeng‘s IRON humanoid robot and a flying car developed by its ARIDGE unit.
Management has earmarked 7 billion yuan for physical-AI research this year, with a target of 1 million sales abroad by 2030.
Overseas Push
XPeng rolled out VLA 2.0 to its Ultra models from March, starting with the P7, G7 and X9.
Pioneer users received early access at the end of 2025, ahead of the wider rollout, the company said.
The company has promised at least one major over-the-air update each quarter.
International road testing will begin before the 2027 launch, XPeng said in March.
Volkswagen will be the first launch customer for VLA 2.0 in China, the companies have said.
First-quarter revenue rose to 13.03 billion yuan on 62,682 deliveries, according to company filings.
Private robotaxi fleets in China and the US doubled in 2025 to about 8,000 vehicles, the International Energy Agency said in May.
By 2035, the IEA expects 700,000 to 3 million robotaxis worldwide.
He Xiaopeng said he expects “full autonomy will arrive within the next one to three years.”
