Baidu unveils Apollo World to train Apollo Go robotaxis
TL;DR
- Apollo World can generate up to 100,000 km of simulated driving per day, which Baidu equates to about three years of real-world testing by a safety-driver-equipped vehicle.
- The engine draws on more than 350 million kilometers of Apollo Go road data, including over 240 million kilometers of fully driverless operation across 28 cities.
- Baidu proposed a '2+3' evaluation standard for world models and is positioning Apollo World as a general physical-AI engine that could extend to robotics.
Baidu unveiled Apollo World at its AI Day in Shenzhen on November 3, pitching the system as a world model that can generate up to 100,000 kilometers of simulated driving per day, which the company equates to about three years of real-road testing by a safety-driver-equipped vehicle. TechNode reported the engine is already in use training the Apollo Go robotaxi fleet, with real-time camera and lidar outputs and virtual pedestrians and vehicles that react to the ego car's moves.
The pitch leans on data scale. Baidu says the model is grounded in more than 350 million kilometers of Apollo Go road data, including over 240 million kilometers of fully driverless operation across 28 cities.
Wang Liang, chief architect of Baidu's Intelligent Driving Group, framed the design philosophy bluntly. A world model, he said, should not be "a painter that makes cool videos" but an "examiner" that checks driverless car safety. In a demo flagged by MartinCid, fellow architect Yang Ze showed the system "producing cyclists cutting across intersections and dogs running into a robotaxi's path," the kind of long-tail cases that are expensive to collect on real streets.
Baidu also proposed a "2+3" evaluation standard for world models, requiring that simulators be causal and interactive, then judging them on physical accuracy, long-run consistency, and speed, with safety decisions due within one-thirtieth of a second. The company is positioning Apollo World as a general engine for physical AI that can extend beyond driving into robotics.
The context is a tightening race. Frank McCleary of Arthur D. Little told reporters "the speed of a launch depends on how well the system can learn to drive in a new place," a line that cuts to why simulation now matters strategically for any operator pushing into new cities. Morgan Stanley projects China will field roughly a million robotaxis by 2035; Apollo Go's rivals WeRide, Pony.ai and Momenta Global are building their own simulation stacks, and this is the sixth Baidu story we've tracked in the last 90 days.
Originally reported by technode.com
Read the original article →Original headline: Baidu Launches Apollo World, a Driving Simulator Generating 100,000 km of Synthetic Scenes Per Day