Baidu World Model Generates Three Years of Robotaxi Driving a Day
Baidu has launched Apollo World, a model that simulates road environments for training robotaxis, and says it can produce up to 100,000 km of simulated driving every day, roughly three years of real world testing by a car with a safety driver on board.
The system generates camera and lidar observations in real time, so a self driving stack can sense and navigate a synthetic environment the way it would a street. Baidu says Apollo Go, its robotaxi arm and the largest operator in China, is already training on it in preparation for fleets it intends to operate outside the country.
But the 100,000 km figure is the claim to hold loosely rather than the number to repeat. Simulated distance is not the same as road experience, because a model produces the situations it was built to imagine, and the part that matters is whether the scenes it invents match the corners real traffic throws up.
But the economics explain why every operator is now chasing this approach. Simulated kilometres cost compute, real ones cost vehicles, safety drivers, permits and time, and Chinese robotaxi firms are all running asset light expansion abroad while carrying the full cost of developing the driving stack at home. A kilometre generated overnight does not need a permit, a driver or a city willing to host a test fleet, which is why the training budget has shifted toward the simulator.
Frank McCleary, an autonomous mobility specialist at Arthur D. Little, put the constraint plainly to Reuters, saying companies still carry the capital cost of training models and vehicles. He added that how fast a company can launch comes down to how quickly its technology can adapt and learn a new environment.
And the competitive field is crowded. Apollo Go is China's largest robotaxi operator and faces WeRide, Pony.ai and Momenta, while XPeng has said it plans to run robotaxis in 2027, and Morgan Stanley has estimated China's robotaxi fleet could reach a million vehicles by 2035.
And the training method is where the approaches diverge. Tesla trains an end to end neural network on video of human driving, which learns behaviour from observation, while a world model generates new situations to train against, which is a bet that synthetic variety beats recorded variety. The weakness in the second approach is that the model can only produce what someone thought to simulate, so the rare events that make driving hard are exactly the ones most likely to be missing.
So the useful question is not whether a simulator can drive 100,000 km a day but whether it can be wrong in the ways the real world is wrong. A model that only produces clean lanes and cooperative traffic teaches a car nothing about the edge cases regulators ask about.
So the test will be what the simulator teaches the fleet. Baidu says Apollo World is already preparing the Apollo Go cars for markets outside China, and a model that reproduces Chinese city traffic well would shorten the work of entering each new city, which is the same case every operator is now making for training in software rather than on tarmac.
Baidu has launched Apollo World, a model that simulates road environments for training robotaxis
It says the system can generate up to 100,000 km of simulated driving a day, about three years of real road testing
Apollo World is already being used to train the Apollo Go fleet
WeRide, Pony.ai, Momenta and XPeng are all pushing into robotaxi services
The training runs in software, which cuts the need for cars, drivers and permits
Source: Reuters
