Article · AI translation
IT Home reported on November 3 that at today's Baidu AI Day Apollo event in Shenzhen, Baidu Apollo unveiled Apollo World for the first time, an agent-driven, fully multimodal world model for physical AI, and proposed a “2+3” set of criteria for evaluating world models.
According to IT Home, Apollo World, which is designed for autonomous driving, must not only generate realistic images but also accurately predict how the real world will change. Baidu Apollo proposed a “2+3” set of criteria for evaluating world models:
“2” refers to the basic requirements:
Causality: A world model should follow the one-way passage of time, with the past determining the future, and predict the next moment based on the current state, rather than stitching together images or inferring the past from the future.
Interactivity: It can respond dynamically to external interventions, with injected actions affecting the future state of the world.
“3” refers to the advanced criteria:
Physical accuracy: How reliable are the simulated states? Do they conform to the laws of the world, and can capabilities acquired through learning still be used when brought back into the physical world?
Persistence: Can the duration and stability of continuous operation keep up, and do objects and details drift during prolonged operation?
Real-time performance: Low latency, with increases in the complexity of the generated world having no effect on model inference time, enabling efficient simulation testing.
In the virtual city built by Apollo World, traffic participants such as vehicles and pedestrians are not “props” following fixed scripts but AI agents capable of making decisions and interacting.
For example, in the same traffic scenario, different actions by an autonomous vehicle will cause the virtual world to change accordingly:
Apollo World can also generate multimodal observation data, such as camera images and lidar point clouds, in real time, allowing an autonomous driving system to perceive its surroundings as it would on real roads.
The driving system uses this information to decide whether to steer, accelerate or decelerate, then sends its actions back to Apollo World. Apollo World then updates the traffic environment and the behavior of other agents and generates observations for the next moment.
In this way, perception, decision-making, action and feedback repeat continuously, forming an end-to-end closed loop. This loop is not confined to a single intersection or stretch of road.
Apollo World supports autonomous vehicles driving from any starting point to any destination in the virtual city, passing continuously through different roads and traffic scenarios to complete an entire trip, from accepting an order and picking up a passenger to dropping them off.
Apollo World extracts high-value scenarios from real-world data and turns complex road conditions and risks into repeatable virtual tests. Autonomous vehicles can repeatedly practice and validate driving strategies in the virtual world. The test results then feed back into system development, and the updated driving system returns to real roads for evaluation, forming a closed loop of continuous improvement.
This creates a continuously cycling data flywheel:
Baidu says Apollo Go is the only ride-hailing platform in China operating L4 autonomous vehicles at scale using end-to-end technology. Data show that Apollo Go has expanded to 28 cities worldwide and logged more than 350 million kilometers of autonomous driving, including more than 240 million kilometers of fully driverless driving.
Apollo World is currently being used for Apollo Go's global testing and validation, helping bring autonomous driving into large-scale deployment.