How long does an autonomous driving test vehicle need to operate to accumulate sufficient real-world road data? Today, the simulated data generated by a single world model in one day may equal years of road testing for a vehicle.
On November 3, Baidu released Apollo World, designed to simulate road environments and generate multi-sensor simulation data for training and testing Robotaxis. Baidu stated that the system can generate up to 100,000 kilometers of simulated driving data per day and is currently being used in Apollo Go's training. In fact, companies such as Waymo and NVIDIA have also been continuously advancing world models for autonomous driving in recent years.
As everyone moves to deploy world models, will real-world road data, previously considered a core asset by the industry, become increasingly worthless?
01. Real Data Becomes More Like a Seed
Previously, autonomous driving companies would expand their test fleets, with one key objective being to acquire more real-world data. The more vehicles drove, the richer the road scenarios they encountered. Traffic rules, driving habits, weather changes in different cities, and various unexpected events difficult to design in advance all became important material for subsequent training. However, real-world road testing has a natural limitation: the most needed data is often the hardest to collect.
A typical urban road drive generates large amounts of data on car-following, lane changing, and intersection navigation. But what truly impacts autonomous driving safety are low-probability, high-risk scenarios. Situations such as a vehicle ahead suddenly driving against traffic, abnormal obstacles appearing in construction zones, or unpredictable interactions among multiple traffic participants cannot be stably obtained simply by increasing test mileage.
This is precisely what world models attempt to change. They no longer just replay data the vehicle has already experienced but continue to generate different environmental changes and traffic participant behaviors based on existing road scenarios. R&D personnel can proactively alter weather, road layouts, or vehicle movements, allowing the autonomous driving system to repeatedly face situations it rarely encounters.
The world model released by Waymo this year demonstrated using generative models to simulate rare scenarios and generate corresponding camera and LiDAR data. NVIDIA is also promoting the use of foundational world models to generate synthetic driving data, supplementing scenarios such as extreme weather and night conditions that are difficult to fully cover through real-world roads. This means that in the future, the value of a segment of real-world road data lies not just in how many kilometers it contains, but in whether it can serve as a basis for generating more effective scenarios.
02. Why Are Problems Still Unsolvable Despite Increasing Test Mileage?
The autonomous driving industry has long emphasized a specific metric: test mileage. The reason for focusing heavily on test mileage is simple: theoretically, the more mileage driven on real roads, the richer the covered scenarios. However, for autonomous driving, mileage does not equate to scenario coverage.
A vehicle driving 1,000 kilometers continuously on a highway may be repeatedly processing similar car-following and lane-keeping scenarios; conversely, a single abnormal interaction at a complex intersection might expose issues the system had never encountered before. World models offer a different approach: rather than waiting for test vehicles to accidentally encounter dangerous scenarios, proactively generate these scenarios and constantly vary the parameters within them. Testing thus shifts from collecting what has already happened to actively verifying what could happen. This is particularly important for autonomous driving R&D.
Traditional data replay has a problem: regardless of the decisions made by the system, the subsequent road footage continues to play according to the original recording. But real driving is not like this. If the autonomous driving system decelerates early, the relative positions of surrounding vehicles change; if it chooses to change lanes, the subsequent traffic relationships that emerge will be completely different. For autonomous driving, the value of world models lies not only in generating more realistic visuals but, more importantly, in constructing a virtual road that evolves dynamically with the vehicle's decisions. Only then can the system face different subsequent environmental changes under different decisions.
03. Data Can Be Generated, But the Real World Cannot Be Replicated
The development of world models easily leads to a misconception: since data can be generated, can real-world road testing be significantly reduced? At least for now, the answer is no. World models themselves still require real-world data for training. What kind of road scenarios they can generate largely depends on what they have previously seen and understood.
More importantly, generated data cannot inherently prove its conformity with the real world.
Recently, the University of Toronto and CUHK MMLab published the DrivingGen benchmark at ICLR 2026, systematically evaluating 14 driving world models. The results showed that while general-purpose models appear better, they violate physics; dedicated driving models exhibit more realistic motion but lag in visual quality.
Some models generate more realistic visuals but may still deviate in physical plausibility and motion consistency. This is the risk that autonomous driving world models must most carefully address. If a model generates a seemingly realistic dangerous scenario where vehicle trajectories, braking processes, or traffic participant behaviors deviate from the real world, systems trained on this data may learn incorrect patterns.
Therefore, world models do not render real-world road data worthless. On the contrary, real data remains a critical foundation for training world models, identifying real-world problems, and validating simulation results. What may truly change is how autonomous driving companies utilize data.
In the past, competitive advantage came primarily from possessing large volumes of data. In the future, competition may shift further toward who can identify the most valuable data and use it to construct scenarios that have not previously existed.
04. Autonomous Driving Data Competition May Enter a New Phase
According to official disclosures, Baidu's Apollo World can generate up to 100,000 kilometers of simulated driving data daily. If simulated data grows rapidly, will autonomous driving companies still need to invest in massive test fleets? At least for now, the answer remains yes. The real world is sufficiently complex; while world models can generate more scenarios, they cannot exist independently of real-world road data.
More importantly, the most dangerous and scarce scenarios often still require continuous discovery and validation through real-world roads. World models do not stop autonomous driving from collecting data; instead, they transform testing from passively waiting for scenarios to occur to actively constructing and verifying them.
In the future, data collected by test vehicles will remain a vital foundation of the entire system. However, the gap between autonomous driving companies may no longer depend solely on who has driven more miles, but on who can identify problems in the real world, generate variations around those problems, and ultimately validate them back on real roads.
Real data will not become worthless; rather, the logic that "more data is always better" may become increasingly insufficient. As autonomous driving begins to actively construct scenarios never encountered before, the question that truly needs answering in the next phase may be: what situations has the system never actually tested?