Baidu launches AI model for robotaxis, generating up to 100,000 km of driving scenarios per day

Reuters reported on Nov. 3, 2026, that Baidu had launched Apollo World, an AI model for robotaxis. Baidu says the model can simulate up to 100,000 km of driving scenarios per day. It generates camera and LiDAR observations in real time, allowing autonomous driving systems to perceive and navigate a virtual environment, the company says.

Baidu says Apollo World generates not only road environments but also camera and LiDAR observations, and is being used to train robotaxis. This demonstrates a system for training autonomous driving systems in virtual environments in addition to on-road driving, but the scale of the generated distance alone does not establish safety or performance on real roads.

The maximum output of 100,000 km per day indicates Apollo World’s scale, but does not mean it can simply replace on-road testing. Baidu’s comparison of the output to “about three years” refers to driving distance; Reuters’ report does not say how well the model reproduces events on real roads or how its safety has been verified. Generated distance should not be treated as proof of performance or safety; the reproducibility of the simulation and validation with real vehicles should be considered separately. A key part of the technical approach is that, in addition to generating road environments, the model outputs camera and LiDAR observations in real time, enabling autonomous driving systems to perceive and navigate virtual spaces directly. The configuration connects environmental input, sensor-data generation and driving-system training in one model. Baidu says it is using the model to train Apollo Go, but the report does not specify which types of roads and traffic conditions it covers or how post-training performance will be assessed with real vehicles. The business model also raises questions. Reuters’ report describes a trend among Chinese robotaxi companies to limit vehicle ownership and operations while expanding through technology supply and partnerships with local operators. It also cites an expert who said model training and vehicle development require capital expenditure. Apollo World may handle part of the training process, but it does not eliminate costs for vehicles and operations. How the company balances expansion based on technology supply with the expense of developing models and vehicles will be a key issue. This is a different layer of the challenge from the system addressed by NHTSA’s proposed standard for optional automated-driving identification lights, which concerns identifying vehicles while they are operating autonomously. Simulation is a development and training tool; it does not replace displays indicating a vehicle’s operating status or safety checks during operation. Questions to watch include how the generated environments are validated, how they are compared with real-world driving data, and how training is reflected in Apollo Go’s real-world operations.