Baidu Apollo Unveils World Model Apollo World for the First Time, Proposes "2+3" Evaluation Criteria

“For world models, the key is not whether the rendered image quality looks good, but whether the simulator can faithfully reproduce the causal relationships of real-world physical laws and behavior.” On November 3, at Baidu AI Day·Apollo held in Shenzhen, Baidu Apollo disclosed for the first time Apollo World, an agent-driven omni-modal world model for physical AI, and proposed a “2+3” evaluation standard for world models. Currently, the model has been used in the global testing and validation of Apollo Go, helping to advance the large-scale deployment of autonomous driving.

Wang Liang, Chief Architect of Baidu's Intelligent Driving Group (IDG) and Chairman of the IDG Technology Committee, said that a world model cannot merely be “a painter that generates flashy videos”; it must also become an “examiner” that validates the safety of autonomous driving. Apollo Go is China's only ride-hailing platform operating L4 autonomous driving at scale with end-to-end technology, with over one million kilometers of fully driverless end-to-end operations, and Apollo World provides it with a virtual training environment to help the “student” AI Driver learn, validate and iterate.

“Apollo World models all traffic participants in a virtual scene as AI agents with decision-making and interaction capabilities, supporting driverless vehicles to travel between any origin and destination in a virtual city, and creating virtual-scenario orders that validate the ‘infinite future’ in advance,” further introduced Yang Ze, Distinguished Architect of Baidu's Intelligent Driving Group (IDG) and head of the world model. In addition, the technology stack of Apollo World can be extended to broader physical AI scenarios such as robotics.

Industry observers analyze that Apollo World can “travel 100,000 kilometers a day,” and the virtual road-test mileage it completes in one day is equivalent to about three years of continuous testing by a real road-test vehicle equipped with a safety driver. Relying on the capability leap brought by the world model, combined with the data flywheel effect, the industry will show a pattern where the strong get stronger, and this trend is also pushing the industry toward the long-awaited “AGI moment.”

A world model cannot merely “look the part”: Apollo sets out a “2+3” yardstick for evaluating world models

Morgan Stanley regards world models as AI's “imagination engine,” enabling AI to simulate in virtual space before acting, achieving “think first, act later.” But transferring this capability to autonomous driving is not easy. Fei-Fei Li once proposed that “a car is a robot with wheels,” but an autonomous vehicle is more like a two-dimensional box for obstacle avoidance, with an obvious gap from three-dimensional household robots.

Wang Liang said that robots that can dance and fold clothes are impressive, but learning to drive is another challenge. Autonomous driving has never been the “EASY mode” of embodied intelligence: robots only need to execute planned actions in confined spaces, whereas autonomous driving takes place on open roads, continuously playing games with a massive number of unpredictable traffic participants. For a vehicle weighing nearly two tons and traveling at tens of kilometers per hour, every judgment and action it makes is directly related to traffic safety, which raises the requirements for safety, real-time performance and conflict gaming by an order of magnitude—the margin for error in the open world is almost zero, and the safety decision cycle must be strictly limited to within 1/30 of a second. This means that a world model cannot be satisfied with merely “looking like” reality.

From the perspective of physical AI, Apollo proposes the “2+3” evaluation standard: the “2” are the basic thresholds that determine whether something counts as a world model, and the “3” are advanced indicators that measure how capable a model is. The two basic thresholds: first, causality, following the one-way nature of time, with the past determining the future, predicting the next moment based on the current state rather than piecing together images to infer the past; second, interactivity, being able to dynamically respond to external interventions, with injected actions affecting future world states. The three gain attributes: physical accuracy—how reliable the simulated states are, whether they conform to the laws of the world, and whether the capabilities learned can still be used when brought back to the physical world; persistence—whether the duration and stability of continuous operation can keep up, and whether objects and details drift during long runs; real-time performance—low latency, with growth in the complexity of the generated world not affecting model inference time, achieving high simulation testing efficiency.

Apollo World's three core capabilities: agent-driven, omni-modal perception, and trip closed loop

Yang Ze introduced: “Apollo World hopes to endow traffic participants with intelligence, so that the world can respond to every action of the driverless vehicle.”

In traditional simulation, other traffic participants are more like props running according to scripts. In the virtual city built by Apollo World, traffic participants such as vehicles and pedestrians are all modeled as AI agents with decision-making and interaction capabilities, able to adjust their actions according to changes in the surrounding traffic. A driverless vehicle slowing down, changing lanes or yielding may trigger different reactions from other participants, enabling the autonomous driving system to be trained and tested in a continuously changing traffic environment.

At the same time, Apollo World can generate multi-modal observation data such as camera and LiDAR data in real time, allowing the autonomous driving system to directly perceive the virtual environment. After the system makes a driving decision, the environment changes accordingly and provides new feedback to the system, forming an end-to-end closed loop.

Yang Ze said: “Many world models in the industry still focus on simulating isolated scenes, while Apollo World pays more attention to closed-loop simulation of complete tasks.” In the virtual city built by Apollo World, driverless vehicles can drive continuously between different origins and destinations, completing the full trip task from accepting an order and carrying passengers to arrival, enabling Apollo Go to be trained and tested in continuous driving across road segments and scenarios.

For autonomous driving to move toward scale, it must simultaneously balance safety, intelligence and scale, and there is an “impossible triangle” among these three. Long-tail scenarios are difficult to fully cover through real-vehicle mileage, deploying in new areas involves testing costs and trial-and-error risks, and the ability to drive continuously across road segments and scenarios requires large-scale validation. Apollo World turns the complex interactions and risk scenarios in real operations into a repeatable, controllable and scalable virtual testing environment, and generates new scenarios not yet covered in reality, allowing driverless vehicles to complete driving simulations in the virtual world before real-vehicle deployment, absorbing testing costs and deployment risks in the virtual world first, and letting AI teach AI to drive.

To make virtual validation truly close to reality, it is inseparable from the real road data continuously accumulated by Apollo Go. Apollo Go has deployed in 28 cities worldwide, completed more than 23 million trips in total, and its total autonomous driving mileage has exceeded 350 million kilometers, including more than 240 million kilometers of fully driverless mileage, and it has carried out driverless practices in both left-hand-drive and right-hand-drive markets. Relying on real driving data, Apollo World extracts high-value scenarios, and the virtual test results in turn feed back into vehicle-side iteration, forming a closed loop in which real operations and virtual validation reinforce each other.

The physical AI market is growing rapidly. Grand View Research estimates that the global physical AI market will grow from about $81.6 billion in 2025 to about $960.4 billion in 2033, with a compound annual growth rate of 36.1% from 2026 to 2033. Moving from “generating images” to “validating actions” and cracking the “impossible triangle,” Apollo World is providing a new technical yardstick for the large-scale deployment of autonomous driving, helping autonomous driving reach the AGI moment.

Leiphone Leiphone