XPENG YOYO bets a robotaxi on cameras alone

A GX SUV pulls up in Guangzhou with nobody behind the wheel, the cabin already cooled to the temperature you picked and a fragrance you chose drifting from the diffuser. There's no lidar dome on the roof. That missing hardware is XPENG's whole argument.

On October 8 the Chinese automaker gave its robotaxi business a name, XPENG YOYO, and launched a ride-hailing mini program that lets members of the public book rides with invitation codes. TechNode described the move as a shift from internal validation to real-world user testing. It's a meaningful step, but a small one, and the gap between the two matters more than the branding.

What XPENG actually launched

The vehicle is a robotaxi version of the GX, XPENG's flagship SUV. According to the company's own account, as reported by China EV Home and TechNode, it carries four in-house Turing AI chips rated at 3,000 TOPS of effective onboard compute, and it runs the second-generation Vision-Language-Action model, VLA 2.0. Automotive World notes that XPENG calls this the highest in-vehicle compute in the world, which is the company's claim, not an independent benchmark.

The timeline has moved quickly. XPENG announced the robotaxi plan in November 2025, started regular road testing in January 2026, set up a dedicated robotaxi division in March, and rolled its first mass-production robotaxi off the line in May 2026. In July it ran the full service loop for the first time: online booking, autonomous pickup, drop-off. Employee rides followed, and Automotive World reports those were limited to staff paying a nominal fare.

In August, XPENG received a remote-testing qualification in Guangzhou covering road tests with nobody in the driver's seat on designated Class I, II and III roads. ArenaEV adds that a cloud-based remote takeover platform can step in if the onboard system gets into trouble. By the October launch, China EV Home counted more than 2,200 employee trial rides. TechNode's figure, as of August, was more than 2,000, so the two numbers fit together.

What XPENG did not say is just as telling. No outlet that covered the launch reported a fleet size, a coverage map, operating hours or rider pricing. "Public trial" here means invited members of the public on designated roads, with a remote operator on call.

The vision-only wager

Most L4 programs are built around lidar and high-definition maps. ArenaEV contrasts XPENG directly with Waymo and Baidu, both of which rely on lidar and pre-mapped cities. XPENG uses neither. Its system is end-to-end and vision-based, with the VLA model turning camera input straight into driving decisions.

The appeal is cost and scale. China EV Home says XPENG aims to lower hardware costs by building on a mass-production vehicle architecture, and TechNode reports the company wants one unified platform running from its Level 2 driver assistance up to Level 4 autonomy. If that works, every consumer GX on the road is collecting data for the robotaxi stack, and the robotaxi is just a production car with more chips and a different job.

The risk is the same one that has followed camera-only autonomy for years. Lidar gives a direct depth measurement that holds up in glare, darkness and heavy rain, and HD maps give the car a prior on what the road should look like. Without them, the model has to infer everything from pixels, every time. XPENG is making up for that with very large onboard compute and a large model. Whether that is enough to run with no human in the loop is the question this pilot is meant to answer. It hasn't answered it yet.

Asset-light, by design

XPENG doesn't plan to run its own taxi fleet. CEO He Xiaopeng has described the company as a technology provider. According to TechNode, XPENG will supply vehicle platforms, driving software and AI capability while partners handle daily operations, and it will earn from vehicle sales, technology services and a share of ride-hailing orders. Automotive World first reported in November 2025 that XPENG was working with Amap, Alibaba's mapping and mobility platform, to put its robotaxis on that network.

He has also set a financial target. China EV Home reports he expects the Guangzhou business to reach vehicle-level break-even as early as the second half of 2027, and he argues that vehicle density within the operating zone matters more than total fleet size. That is a per-car figure, not company-wide profitability, and it depends on a fleet nobody has sized publicly.

The global ambitions are bigger still. The English name was chosen with export in mind, XPENG says it is talking with potential partners in Europe, the Middle East and Southeast Asia, and He expects markets outside China to eventually account for most of the robotaxi business. A dedicated robotaxi model is planned for 2027.

Where it stands against the field

For scale, Pony.ai reported a fleet of more than 1,700 robotaxis in its first-quarter 2026 results and raised its year-end target to over 3,500, across more than 20 cities. Baidu's Apollo Go reported 20 million cumulative rides by February 2026. Those programs are already running commercial driverless services at volume. XPENG is running invite-only trials with a remote safety net and an undisclosed fleet.

XPENG's possible advantage is the cost structure, not the head start. An automaker that already builds the car, designs the chips and trains the model could get to cheap, scalable driverless vehicles faster than a software company buying hardware from someone else, if the vision-only system holds up.

What to watch

The milestone that matters is the one XPENG set for itself: passenger service without a safety officer, targeted for 2027. Before then, look for three things. First, a fleet number and a coverage map, because a robotaxi service without them is still a demo. Second, how often the remote takeover platform is actually used, which XPENG has not disclosed and which is the best available measure of whether cameras alone can carry the load. Third, a named operator outside China.

XPENG has shown it can build a robotaxi, put it on public roads and take bookings. It has not yet shown that a camera-only car can be trusted on its own, at scale, in the conditions where lidar earns its keep. Until it does, YOYO is a well-funded test of an idea the rest of the industry has bet against.