Baidu Discloses ApolloWorld, a World Model for Physical AI

The global Robotaxi competition is ushering in a new watershed moment.

In February 2026, autonomous driving giant Waymo completed a $16 billion (approximately RMB 107.2 billion) financing round, reaching a post-money valuation of $126 billion (approximately RMB 844.2 billion).

Chinese companies are also accelerating their expansion. Baidu's Apollo Go has entered 28 cities worldwide, completed more than 23 million ride services in total, and surpassed 350 million kilometers in total autonomous driving mileage.

After more than a decade of development, autonomous driving has moved from technical validation to large-scale commercial operation, but safety, cost, and efficiency remain the "three mountains" facing the industry.

In particular, as global expansion accelerates, the traditional model of continuously accumulating test mileage through real vehicles and safety drivers is facing an increasingly obvious efficiency bottleneck.

And world models are becoming the key variable that is changing this situation.

In February this year, Waymo released a world model built on Google DeepMind Genie 3.

On November 3, Baidu also officially unveiled ApolloWorld, an agent-driven omni-modal world model for physical AI.

According to data released by Baidu, the virtual road-test mileage ApolloWorld completes in one day is equivalent to about three years of continuous testing by a real vehicle equipped with a safety driver.

More importantly, in the virtual city built by ApolloWorld, not only are the autonomous vehicles driven by AI, but surrounding traffic participants also have autonomous decision-making capabilities and can continuously interact with the autonomous vehicles, allowing the autonomous driving system to experience various complex traffic situations in advance.

1. The Global Robotaxi War Escalates, World Models Become the New Watershed

Looking back over the past decade or so, the global autonomous driving industry has roughly gone through three stages of development.

From around 2013 to 2023, the industry mainly competed around autonomous driving technology R&D and road testing, with the focus on enabling cars to drive autonomously.

Starting in 2022, as Baidu and other companies successively obtained fully driverless commercial operation permits, and then beginning in 2023, Robotaxi began to move from technical validation to paid operation, with the focus of competition gradually shifting to fleet size, operational capability, and business models.

▲Schematic of Robotaxi industry development (technical development across stages overlaps)

By 2026, the global Robotaxi industry has entered a new stage.

Waymo is accelerating expansion into operating cities, while Baidu's Apollo Go has already launched fully driverless commercial operations in Dubai and Abu Dhabi, is conducting public road testing in London and Switzerland, and has obtained Hong Kong's first fully driverless testing license. As global expansion accelerates, the industry has begun to compete on large-scale operational capability across cities and markets.

The capital market is also paying attention to this change. Waymo completed a $16 billion financing round in February this year (approximately RMB 107.2 billion), officially announced its world model four days later, and in October obtained $5 billion (approximately RMB 33.5 billion) in debt financing.

Capital is often very sharp-sighted, because leading investment institutions value not only a company's existing operational scale, but also whether it can build sustained leading technological and commercialization capabilities in the future.

And Robotaxi scaling is precisely facing three mountains: safety, cost, and efficiency.

On the one hand, real roads are full of low-frequency, complex, and even dangerous traffic scenarios, and it is difficult to fully cover all risks simply by increasing real-vehicle test mileage.

On the other hand, every time an autonomous driving company enters a new city, it needs to invest a large number of vehicles, manpower, and time to validate the local road environment and driving habits.

If every algorithm upgrade and every entry into a new market requires another lengthy road-testing process, the pace of corporate expansion will be significantly constrained.

In the past, the autonomous driving industry mainly accumulated experience by increasing actual test mileage.

▲World models can generate different content

This is not only a change in testing methods, but may also become an important watershed in the next stage of Robotaxi competition.

2. World Models Debut in Clusters, Baidu Sets a New Technical Benchmark

What exactly is a world model?

In June 2026, AI scientist Fei-Fei Li once divided world models into three types: renderers, simulators, and planners.

Simply put, a renderer is responsible for drawing the world, a simulator is responsible for predicting how the world changes, and a planner is responsible for deciding what action to take next.

Among them, the simulator is an important link connecting the virtual world with real actions.

Fei-Fei Li once compared autonomous vehicles to "boxes" moving on a two-dimensional plane, and argued that general-purpose robots need to complete more complex three-dimensional spatial operations.

▲Dr. Wang Liang, Chief Architect of Baidu's Intelligent Driving Group (IDG) and Chairman of the IDG Technology Committee

Dr. Wang Liang, Chief Architect of Baidu's Intelligent Driving Group (IDG) and Chairman of the IDG Technology Committee, offered another judgment.

Although the control actions of a car are relatively simple, autonomous driving faces dynamic games among a large number of traffic participants on open roads.

For example, when an autonomous vehicle is preparing to pass through an intersection, a nearby vehicle may actively yield or may suddenly accelerate. Different choices by the autonomous vehicle will further change the behavior of other vehicles.

Therefore, the difficulty of autonomous driving lies not only in understanding the road, but also in predicting the choices of other traffic participants and the chain reactions triggered by those choices.

▲Real-time interaction caused by different decisions

With the rapid development of generative AI, world models have become a hot concept in the AI industry. From video generation and 3D space construction to physics simulation, different technologies have begun to be labeled as world models, but whether they can truly be used for physical AI still varies markedly.

This time, Baidu did not rush to package its technology with a new concept. Instead, it chose to start from the actual needs of L4 autonomous driving, first clarifying what problems a world model should solve, and then designing the system around that goal.

In Baidu's view, a world model cannot merely be a tool for generating realistic images; it should be a testing environment capable of simulating real physical laws, understanding the behavior of traffic participants, and continuously validating autonomous driving capabilities.

For this reason, Baidu proposed a "2+3" world model evaluation standard for physical AI, attempting to establish a clearer evaluation method for this rapidly developing technology.

Among these, the two basic thresholds are causality and interactivity: the model must not only predict the future based on the current state, but also respond to external actions so that different decisions produce different outcomes.

For example, in a virtual scene demonstrated by ApolloWorld, an autonomous vehicle drives to an elevated road entrance.

When the autonomous vehicle chooses to drive onto the elevated road, vehicles on the adjacent branch road merge onto the main road; when the autonomous vehicle chooses to continue along the side road, the branch-road vehicles stop and wait.

At the same intersection, because the autonomous vehicle makes different choices, the entire traffic environment changes accordingly.

The other three advanced indicators are physical accuracy, persistence, and real-time performance, requiring the model to conform to real physical laws, perform stable inference over long periods, and respond promptly to vehicle actions.

▲Different driving decisions elicit different model responses

According to this standard, a world model for autonomous driving cannot merely generate a realistic road video; it also needs to serve as a testing environment for real autonomous driving systems.

Having obtained simulated data, enabling virtual LiDAR to accurately reproduce perception changes in the real environment is also an important part of the model's capabilities.

In rain and snow, raindrops and snowflakes affect the propagation and reflection of laser light, causing weakened echoes, missing point cloud data, or noise.

▲Synchronized simulation of LiDAR point cloud data

ApolloWorld must not only simulate rain, snow, reflections, and visibility changes in camera images, but also synchronously generate LiDAR point clouds that conform to physical laws, and ensure that both sensors see the same world.

This is more complex than simply generating realistic road videos, and it is also the key to whether virtual testing can truly reflect an autonomous vehicle's perception capabilities.

Furthermore, ApolloWorld can support an autonomous vehicle in completing a full trip from start to destination in a virtual city. This presentation demonstrated virtual driving tasks from the Shenzhen Binhai Boulevard side road to the A8 Music Building, and from the Hong Kong SkyCity Marriott Hotel to multiple destinations, and can also adjust traffic density along the same route.

▲Complete virtual road testing

A single intersection tests a single driving decision, while an entire trip tests the autonomous driving system's ability to continuously complete tasks.

Globally, world models have become a technological high ground contested by leading autonomous driving companies, but the directions different companies are exploring are not entirely the same.

▲Comparison of Baidu's and Waymo's autonomous driving world models

In February this year, Waymo announced a world model built on Google DeepMind Genie 3, capable of generating complex scenarios such as tornadoes, extreme weather, and rare obstacles, and also supporting driving action control, traffic participant behavior settings, and multi-sensor simulation.

Compared with Waymo's world model technology, Baidu ApolloWorld places greater emphasis on the practical application value of world models and is closer to the needs of large-scale Robotaxi operations.

As a result, Baidu ApolloWorld can not only simulate dynamic traffic games among multiple agents, but also extend simulation from a single scenario to a complete travel task, and connect to an L4 autonomous driving system already operating at scale, enabling world models to truly serve the training, validation, and commercialization of autonomous driving.

The new-generation end-to-end autonomous driving system of Apollo Go disclosed this time has already achieved fully driverless operations of more than one million kilometers, and ApolloWorld is precisely the training and validation environment provided for this system.

ApolloWorld can not only build an interactive virtual world, but also enable an L4 autonomous driving system already put into actual operation to continuously undergo testing within it.

World models thus move beyond frontier technology exploration and further become the technological foundation supporting large-scale Robotaxi operations.

III. From one day matching three years to global city expansion: Robotaxi enters an era of "the strong getting stronger"

No matter how advanced a world model is, it must ultimately answer one question: what can it bring to the autonomous driving industry?

A set of data disclosed by ApolloWorld is quite representative.

According to Baidu, ApolloWorld can complete about 100,000 kilometers of virtual road testing per day, equivalent to about three years of continuous testing by a real vehicle equipped with a safety driver.

Some of the testing work that previously required real vehicles, roads, and manpower can now be moved to the cloud and continuously scaled through computing power.

For example, on real roads it is difficult to repeatedly create dangerous situations such as pedestrians suddenly crossing or the vehicle ahead braking sharply for testing purposes, but in the virtual world these scenarios can be repeatedly generated and traffic conditions continuously adjusted.

The change brought by world models is not just an increase in testing mileage, but also enables autonomous driving systems to undergo testing across a large number of complex scenarios at lower cost and higher efficiency.

This is especially important for Robotaxi companies that are accelerating their global expansion.

Road structures, traffic rules, and driving habits differ across countries and cities. Take Hong Kong as an example: Baidu has already conducted fully driverless testing there on right-hand-drive, left-hand-traffic roads.

▲Illustration of Baidu Apollo testing in Hong Kong

Combining real operating experience with virtual testing helps reduce the validation cost of cross-market deployment and improve the efficiency of launching new cities.

And Baidu has another noteworthy advantage—the large-scale real-world operating data that Apollo Go has already accumulated.

Currently, Apollo Go's total autonomous driving mileage exceeds 350 million kilometers, of which fully driverless mileage exceeds 240 million kilometers.

This real driving experience can continuously provide high-value scenarios for ApolloWorld. The world model, in turn, can generate more test situations around real problems and then feed the validation results back to the vehicle-side autonomous driving system.

Real operations provide experience, the virtual world expands that experience, and the autonomous driving system continuously iterates through validation.

As a result, Baidu is forming a noteworthy technological closed loop. Real operations accumulate driving experience, the world model continuously expands virtual test scenarios, and after training and validation, the autonomous driving system brings new capabilities back to real roads.

This may also change the competitive logic of the Robotaxi industry.

When real driving experience can be repeatedly utilized in the virtual world, operating scale is no longer just an advantage in fleet size and order volume; it may also translate into an advantage in algorithm iteration efficiency.

For a leading company like Baidu, combining hundreds of millions of kilometers of real driving experience with ApolloWorld gives the operating data it has already accumulated the opportunity to generate greater technological value. And faster iteration speed will in turn support the company's expansion into new operating regions.

From competing over the scale of actual testing to competing over the combined capabilities of real operations and virtual training, world models are expected to further amplify the advantages of leading companies and push the Robotaxi industry into a new phase of "the strong getting stronger."

▲The advantage of continuous data iteration

From a longer-term perspective, this is also an important meaning of what Baidu calls the "AGI moment" of the autonomous driving industry.

And companies with real operating scale and world model technological capabilities are also expected to further expand their advantages in the next stage of global Robotaxi competition.

Conclusion: Whoever masters world models will hold the initiative in the second half of the Robotaxi race

For the Robotaxi industry, which has already entered global scale competition, the future contest is not only about who can drive better, but also about who can make driverless vehicles continuously evolve at lower cost and higher efficiency.

The emergence of world models such as Baidu ApolloWorld has also turned world models from a cutting-edge AI technology into an important variable affecting the competitive landscape of the autonomous driving industry.

In the future, whoever masters world models is very likely to seize the initiative in the second half of the Robotaxi race ahead of others.