China's Baidu has launched "Apollo World," an AI system that generates virtual driving environments to train its Apollo Go robotaxis. The company announced the model at its AI Day in Shenzhen, Reuters reported. The rollout comes as Chinese and global self-driving firms continue to search for cheaper ways to develop autonomous technology.

A day of virtual driving, per Baidu

The headline claim is capacity. Baidu says Apollo World can simulate up to 100,000 km (62,000 miles) of driving per day. According to the company, that volume equals roughly three years of road testing by a single self-driving car operating with a safety driver.

That comparison is Baidu's own, however, and not an independent benchmark. The three-years-a-day figure refers only to simulated driving. It does not demonstrate how safe the cars are on real roads or whether they are ready for daily use.

The open question is whether the simulations accurately match the varied roads and driving conditions in the cities where Apollo Go plans to operate. The company also has yet to show how much additional real-world testing the vehicles will require.

Baidu added that it is already using Apollo World to train Apollo Go robotaxis for deployment in other countries. Apollo Go is China's largest robotaxi operator, and its rivals include WeRide, Pony.ai and Momenta Global.

Sensor streams, not just road scenes

Apollo World goes beyond generating static images. It can simulate how roads and their surroundings change over time and predict what might happen next. Live data from cameras and lidar sensors helps the self-driving software understand its environment and navigate through it.

The model is trained on data from real road tests. Baidu has reportedly recorded more than 350 million kilometers of driving data, which Apollo World also uses to generate new road situations the vehicles have not previously encountered. Those previously unseen scenarios matter because rare, hazardous events — the kind most likely to expose a self-driving system's limits — seldom turn up often enough in everyday driving to be collected through road testing alone.

Baidu has also reportedly proposed a "2+3" standard for testing world models — AI systems that model how an environment evolves and generate simulated situations for other software, such as a self-driving stack, to learn from. If other developers adopt it, Baidu could help shape how such systems are judged over the long term.

The company described Apollo World as an AI system capable of understanding and acting on information from multiple sources. It said the technology could also be applied to developing robots, not just self-driving cars.

The cost problem driving the push

One of the biggest challenges in the robotaxi business is the cost of developing and operating the technology. Many Chinese companies do not own all the vehicles they deploy. Instead, they supply the self-driving technology and partner with carmakers, ride-hailing firms and other overseas operators. That arrangement keeps running costs down, but building and training the underlying systems remains expensive. Simulated mileage speaks to that expense: generating driving distance in software avoids the years a physical car with a safety driver would need to cover the same ground.

Frank McCleary, a self-driving expert at consultancy Arthur D. Little, told Reuters that companies continue to spend heavily on training their AI systems and preparing vehicles. He said the pace of a launch depends on how well a system can learn to drive in a new place, and stressed that AI has helped companies make even faster progress.

China's robotaxi market is set to grow quickly over a short period. Morgan Stanley expects the country to have 1 million robotaxis by 2035, Reuters reported. XPeng, Volkswagen's partner in China, also plans to launch its robotaxi service in 2027. Tesla is taking a different approach with its Cybercab fleet, using AI models trained on videos of people driving. How simulations hold up on real roads — and whether other developers adopt Baidu's proposed testing standard — are among the markers to watch as that market takes shape.