Baidu unveils Apollo World, a world model that simulates whole robotaxi trips for Apollo Go

Baidu has unveiled Apollo World, which it calls an agent-driven, all-modality world model for physical AI, at the Apollo session of Baidu AI Day in Shenzhen on 3 November. The announcement pitches it as a way to run a robotaxi through a full trip in a virtual city, from accepting an order to dropping off the passenger, before the car drives the real streets. Baidu says it is already used for testing and validation of Apollo Go, its robotaxi service.

What Baidu says it does

In Apollo World’s virtual city, other cars and pedestrians are not scripted props, Baidu says, but AI agents that make decisions and react. If the robotaxi slows down or overtakes, the simulated world responds differently.

The model also generates the sensor feed a driving system needs. Baidu says it produces camera images and lidar point clouds in real time, and that weather changes show up in both. The driving system reads those observations, decides to steer, accelerate or brake, and sends the action back. Apollo World then updates the traffic and the other agents and renders the next moment. Baidu describes this as an end-to-end closed loop that runs from any start point to any destination, not one junction at a time.

Baidu also says Apollo World pulls high-value scenarios from real driving data and turns hard road conditions into repeatable virtual tests, with results fed back into the driving system before it goes back on the road. It frames the stakes in its own terms: for a two-tonne car at tens of kilometres per hour, the safety decision cycle has to stay within 1/30 of a second.

The numbers, as Baidu gave them

According to Reuters, Baidu said Apollo World can generate “up to 100,000km” of simulated driving a day, which it equated to about three years of real-world testing by an autonomous car with a safety driver. Reuters also reports Baidu saying the system is being used to train Apollo Go’s fleet.

In its own post, Baidu says Apollo Go covers 28 cities worldwide and has logged more than 350 million km of autonomous driving, over 240 million km of it fully driverless. None of these figures come with a method or outside audit.

A “2+3” yardstick for world models

The more portable part of the announcement is a checklist. Baidu proposes two baseline requirements:

• Causality: the model moves forward in time, predicting the next moment from the current state rather than stitching frames together or working backwards.
• Interactivity: it responds to outside actions, so an injected action changes the future state of the world.

And three advanced measures:

• Physical accuracy: whether simulated states follow real-world rules, and whether skills learned inside transfer back to the physical world.
• Persistence: how long it can run stably without objects and details drifting.
• Real-time performance: low latency, with inference time not growing as the scene gets more complex.

Those criteria map neatly onto the gaps we have flagged in media-focused world models. Genie 4 and Tencent’s open-weight Hy-Nomas are judged mostly on how worlds look and how they can be steered, while a driving simulator lives or dies on physical accuracy and on whether a lesson learned in simulation holds up on a real street.

Why this matters

The “2+3” list is a practical checklist for evaluating any world model you are considering for simulation or agent training, especially the physical accuracy and persistence points that demos rarely show. Apollo World itself is an internal Baidu tool: the announcement mentions no paper, weights, API or third-party test, so unlike Hy-Nomas, there is nothing you can run. The real test of Baidu’s claims will be Apollo Go’s record in the cities it validates with Apollo World.