Gasgoo Munich-At its Baidu AI Day·Apollo event in Shenzhen on Nov. 3, Baidu Apollo unveiled Apollo World — an agent-driven, full-modal world model built for physical AI. The company also introduced a “2+3” framework to evaluate what truly qualifies as a world model.
Image credit: Baidu Apollo
The “2” sets the baseline threshold, covering two core dimensions. First is causality: time flows in one direction, meaning the past dictates the future. The model must predict the next moment based on the current state — not stitch together frames to reverse-engineer the past. Second is interactivity: the system must dynamically respond to external interventions, allowing injected actions to alter future states.
The “3” acts as an advanced metric for model capability, adding three enhancing attributes. Physical accuracy measures how reliably simulated states align with real-world physics — and whether skills learned in simulation hold up in reality. Persistence gauges whether the model can run continuously without stability degrading or object details drifting over time. Real-time performance demands low latency: as the generated world grows more complex, inference time shouldn’t bloat, ensuring high simulation efficiency.
The model is already powering global testing and validation for Apollo Go, Baidu’s robotaxi service, accelerating the scaled deployment of driverless vehicles.
Apollo World can virtually “drive 100,000 km a day.” A single day of its virtual road testing matches roughly three years of continuous driving by a real test vehicle staffed with a safety driver.
Image credit: Baidu Apollo
Here’s how it works: Apollo World converts complex interactions and edge-case risks from real-world operations into repeatable, controllable, and scalable virtual environments. It also generates novel scenarios not yet encountered in reality. This allows robotaxis to game out driving maneuvers in the virtual world before physical deployment — absorbing testing costs and deployment risks digitally. In effect, it lets AI teach AI to drive.
Yet virtual validation only rings true when grounded in real-world data — a reservoir Apollo Go continues to build. The robotaxi service has expanded to 28 cities globally, completing over 23 million trips. Its autonomous driving mileage has surpassed 350 million km, including more than 240 million km in fully driverless mode. It has even begun driverless operations in both left- and right-hand-drive markets.
Image credit: Baidu Apollo
Anchored in this real-world driving data, Apollo World extracts high-value scenarios for virtual testing, then feeds the results back into vehicle-side updates. The result is a closed loop where real-world operations and virtual validation continuously sharpen one another.
Beyond Apollo Go, Apollo World’s tech stack is reportedly extensible to broader physical AI applications — including robotics.
