The Chinese tech giant also proposed a new '2+3' standard for judging world models at its AI Day in Shenzhen
Baidu wants its robotaxis to do more of their learning inside a simulation. At its AI Day in Shenzhen on November 3, 2026, the company introduced Apollo World, which it describes as an agent-driven multimodal world model built for physical AI.
Baidu also proposed a new way to grade this kind of technology: a ‘2+3’ evaluation standard for world models.
What Baidu actually announced
Apollo World debuted during the Apollo session of Baidu AI Day. Apollo is the company’s autonomous driving brand, and the new model sits squarely inside that business.
The model is already integrated into Apollo Go, Baidu’s robotaxi service, which has been tested globally. Inside Apollo Go, the system handles verification and simulation across what Baidu frames as infinite virtual scenarios.
The engine behind Apollo World draws on data gathered from more than 350 million kilometers of real-world testing by Apollo Go. That driving history gives the model a large library of actual road behavior to learn from before it starts inventing new situations.
Why ‘agent-driven’ is the key phrase
A world model is roughly what it sounds like. It is an AI system that builds an internal picture of how a physical environment works, then predicts what happens next.
Apollo World layers agent-based modeling on top of that idea. Vehicles and pedestrians inside the simulation are treated as agents, meaning each one makes its own decisions and reacts to the others.
According to the research summary, this setup is designed to support intelligent decision-making and interaction among traffic participants.
The ‘multimodal’ label signals that the model works across more than one type of input.
The ‘2+3’ standard and the grading problem
Baidu’s second announcement may prove just as interesting as the first. The company proposed a ‘2+3’ judgment for evaluating world models.
The research findings describe the standard as a more rigorous way to assess AI models, one that could influence future development across the industry.
Background: the push toward simulated worlds
The launch fits a broader industry trend toward world models that can simulate and predict physical environments. The appeal is straightforward: they can reduce how much a company depends on collecting fresh real-world data.
Baidu’s approach combines both worlds. It uses its 350 million-plus kilometers of real driving as the foundation, then uses Apollo World to stretch that foundation into scenarios the fleet has never encountered.
Per the research findings, the technology stack can be extended to robotics and other physical AI applications.
What this means
For Baidu, Apollo World is a bid to turn its accumulated driving data into a durable advantage. The goal, per the research findings, is to help Baidu scale its autonomous driving deployments.
The ‘2+3’ standard is the piece to monitor most closely. If other developers adopt it, Baidu gains influence over how the entire category defines quality.
Baidu is positioning Apollo World as a general engine for physical AI, not just a robotaxi tool, with the company noting potential applications in robotics and broader physical AI contexts.
