News · SI
Baidu's Apollo World simulates 100,000 km of robotaxi driving a day
Baidu has unveiled Apollo World, an agent-driven world model that builds virtual cities where cars and pedestrians react to a self-driving car in real time. Baidu says it can generate up to 100,000 km of simulated driving a day and is already used to test and validate Apollo Go's robotaxis worldwide.
Baidu has unveiled Apollo World, an AI system that simulates road environments for training and testing robotaxis, and says it can generate up to 100,000 km of simulated driving a day[1]. Reuters reports that is about three years of real-world testing for a safety-driver-equipped autonomous vehicle[1]. Baidu Apollo showed it at Baidu AI Day's Apollo session in Shenzhen on November 3[2].
01 / 04
A city that reacts back
In Apollo World's virtual cities, cars and pedestrians aren't scripted props. Each is an AI agent that makes decisions and reacts to traffic, so if the robotaxi slows, changes lanes or yields, the agents around it respond differently[2]. The model generates camera images and lidar point clouds in real time, the driving system decides to steer, accelerate or brake, and the world updates to produce the next observation, closing the loop end to end[3]. Vehicles can run complete trips in the simulation, from pickup to drop-off[3].
Wang Liang, chief architect of Baidu's Intelligent Driving Group, put it this way: a world model shouldn't just be "a painter generating cool videos", it should be the examiner that verifies a driverless car is safe[2].
02 / 04
Already training Apollo Go
Baidu says Apollo World is already used for global testing and validation of Apollo Go, its robotaxi service[1][2]. The system pulls high-value scenarios from real operations, turns them into repeatable virtual tests, and feeds results back into the driving system before it returns to real roads[3]. Baidu says Apollo Go now runs in 28 cities, has completed more than 23 million rides, and has passed 350 million km of autonomous driving, over 240 million km of it fully driverless[2].
03 / 04
Baidu's '2+3' test for world models
Alongside the model, Baidu proposed a "2+3" standard for judging world models[3]:
- The 2 baseline requirements. Causality, meaning time only runs forward, and interactivity, meaning the world responds to outside interventions.
- The 3 advanced measures. Physical accuracy, persistence so long runs stay stable, and real-time, low-latency inference.
04 / 04
Why it matters
Reuters notes autonomous-driving companies are embracing world models to cut development costs, and Baidu's Apollo Go competes with WeRide, Pony.ai and Momenta Global[1]. If simulated miles can stand in for years of road testing, the race shifts toward whoever has the most real operating data to seed those simulations. Baidu also says the same stack can extend to robots and other physical AI[2].
Sources
Sources for this story.
FAQ
Frequently asked questions.
01What is Apollo World?
Apollo World is Baidu Apollo's agent-driven, full-modal world model for physical AI, unveiled at Baidu AI Day's Apollo session in Shenzhen on November 3. It builds virtual cities where vehicles and pedestrians are AI agents that react to a self-driving car, and streams camera and lidar data so the driving system can train and be tested in a closed loop.
02How much driving can it simulate?
Baidu says up to 100,000 km of simulated driving a day, which Reuters reports is roughly three years of real-world testing by a safety-driver-equipped autonomous vehicle.
03Is it in use yet?
Yes. Baidu says Apollo World is already used for global testing and validation of Apollo Go, its robotaxi service, which it says operates in 28 cities and has passed 350 million km of autonomous driving, more than 240 million km of it fully driverless.
04What is the '2+3' standard?
Baidu's proposed test for a useful world model. The '2' are baseline requirements of causality and interactivity, meaning time runs one way and the world responds to interventions. The '3' are physical accuracy, persistence over long runs, and real-time, low-latency inference.
That’s the story
Building with what’s new?
Spectrum Web Co ships products, places talent, builds social narratives and builds custom business tools as monthly services.
Related stories