Baidu launches AI system to simulate 100,000km of daily robotaxi driving

Simulating the open road

Baidu recently introduced a new AI system called Apollo World. This technology is designed to simulate various road environments specifically to train robotaxis. By creating these digital worlds, the company aims to predict how autonomous vehicles will behave in different scenarios.

The primary goal of this initiative is to help automakers reduce the high costs associated with developing self-driving technology. Developing these systems is a major financial hurdle for the industry. Apollo World provides a way to bypass some of the physical limitations of traditional testing methods.

As reported by The Straits Times, the system is already being utilized to train the Apollo Go fleet. Apollo Go is currently the largest robotaxi service operator in China. The company intends to use this training to prepare its fleet for global operations.

Synthetic data at scale

One of the most significant features of Apollo World is its massive output capacity. Baidu says the system can generate up to 100,000km of simulated driving every single day. To put that into perspective, this amount of data is equivalent to about three years of real-world testing for a single vehicle equipped with a safety driver.

The system functions by producing real-time multimodal data streams. These streams include camera observations and lidar data, which is a type of laser scanning used to map surroundings. By providing these inputs, the AI allows self-driving systems to sense and navigate the simulated environments as if they were real.

This high-volume data generation is crucial for teaching AI how to handle rare or dangerous events. Instead of waiting for a specific accident to happen on a real street, engineers can program the simulation to create those scenarios. This allows the software to learn from a much wider variety of experiences in a shorter timeframe.

The cost of autonomy

Despite the progress in AI, the development of robotaxis remains a very expensive endeavor. Experts note that companies face heavy costs for both training models and the physical vehicles themselves. These capital expenditures are a challenge that every autonomous vehicle company in the world is currently facing.

To manage these costs, many Chinese firms have adopted asset-light expansion strategies. This means they focus on supplying the underlying autonomous-driving technology to others. They often partner with automakers, ride-hailing platforms, and local fleet operators rather than owning the entire fleet themselves.

Frank McCleary, an expert in autonomous mobility and urban transportation at Arthur D. Little, noted that the speed of a launch depends on a technology's ability to learn. He explained that AI has helped accelerate these advances by improving how quickly a system can adapt to its environment. This highlights why simulation tools like Apollo World are becoming central to the industry.

The competitive landscape

Baidu is not alone in the race to dominate the robotaxi market. It competes with several other major players, including WeRide, Pony.ai, and Momenta Global. Each of these companies is vying for a significant share of the growing autonomous transport market in China and beyond.

Other manufacturers are also making moves to enter the space. For example, Volkswagen’s Chinese partner XPeng has announced plans to launch its own robotaxi operations in 2027. This indicates that traditional automakers are increasingly looking to integrate autonomous technology into their business models.

The scale of the opportunity is significant for the region. Morgan Stanley has estimated that the robotaxi fleet in China could grow to 1 million vehicles by 2035. This growth suggests that the infrastructure for autonomous transport is expected to expand rapidly over the next decade.

Future of training models

The industry is seeing a shift toward different types of AI architectures to achieve these goals. For instance, Tesla is pursuing an end-to-end neural network to develop its self-driving technology. This approach relies on training the system on vast amounts of video footage of human driving.

Baidu’s approach with Apollo World complements these trends by focusing on the generation of synthetic environments. By combining high-quality simulation with real-world data, companies can create more robust systems. This multi-pronged approach helps solve the problem of how to train a vehicle to handle the unpredictability of human behavior.

As these technologies mature, the focus will likely shift toward how quickly these models can be deployed in diverse cities. The ability to simulate different climates, traffic laws, and urban layouts will be essential. Apollo World represents a step toward making that global deployment more feasible and less costly for developers.