Gasgoo Munich-Baidu Apollo unveiled Apollo World, an agent-driven, multimodal world model designed for physical AI, at a dedicated Apollo session during Baidu AI Day in Shenzhen on November 3, according to Baidu. The company also introduced a set of "2+3" criteria for evaluating world models.
Image source: Baidu Apollo
A "2+3" Framework for Evaluating World Models
The "2" represents two fundamental requirements for a system to qualify as a world model: causality and interactivity. Causality means following the one-way progression of time, using past information and the current state to predict what happens next, rather than reconstructing the past by piecing together visual information. Interactivity refers to the ability to respond dynamically to external interventions, allowing actions to influence subsequent world states.
The "3" refers to three advanced attributes for evaluating model performance: physical accuracy, persistence and real-time responsiveness. Physical accuracy measures whether simulated states reliably follow the laws of the physical world and whether capabilities learned in simulation can be transferred to real-world applications. Persistence concerns the model's ability to operate continuously and remain stable over time without objects or details drifting. Real-time responsiveness requires low latency, with inference time remaining stable as the complexity of the generated world increases, enabling more efficient simulation and testing.
Virtual Testing for Autonomous Driving
Apollo World is already being used in testing and validation for Baidu's Apollo Go robotaxi service as the company expands its autonomous driving operations globally.
According to Baidu, Apollo World can simulate up to 100,000 kilometers of driving per day. This daily virtual testing distance is equivalent to approximately three years of continuous road testing by a real-world vehicle equipped with a safety driver.
Apollo World converts complex interactions and high-risk scenarios encountered in real-world operations into repeatable, controllable and scalable virtual testing environments. It can also generate scenarios that have not yet been encountered on actual roads. This allows autonomous vehicles to undergo driving simulations in virtual environments before deployment, addressing testing costs and deployment risks in simulation. The approach effectively allows AI to train AI for autonomous driving.
Real-World Driving Data Supports Model Training and Validation
The effectiveness of virtual validation depends on its ability to reflect real-world conditions, supported by the driving data accumulated through Apollo Go's operations. The service currently operates in 28 cities worldwide and has completed more than 23 million rides. Its autonomous driving fleet has accumulated over 350 million kilometers, including more than 240 million kilometers of fully driverless operation. Apollo Go has also conducted driverless operations in both left-hand-drive and right-hand-drive markets.
Using real-world driving data, Apollo World identifies high-value scenarios for virtual testing. The results are then fed back into vehicle-side development and iteration, creating a cycle in which real-world operations and virtual validation continuously improve one another.
Potential Applications Beyond Autonomous Driving
Beyond Apollo Go, Baidu's Apollo World technology stack could also be extended to broader physical AI applications, including robotics.
