End-to-end solutions for autonomous driving from the perspective of car companies

As the wave of intelligence and information technology deepens in the global automotive industry, consumers' demand for intelligent driving functions has increased significantly. The level of intelligence has also become one of the important considerations when buying a car, especially many young consumers regard "intelligence" as the core of their decision when choosing a car.

The importance of intelligence in car purchase decisions is becoming increasingly prominent, which has also prompted many automakers to speed up the research and development of intelligent technology and engineering investment. Auto brands that fail to keep up with the pace of intelligence may gradually lose their market competitiveness. The development of autonomous driving technology has entered a critical stage, and the introduction of the end-to-end concept has also made the path of autonomous driving technology clearer.

Rapid development of automotive intelligence and autonomous driving technology

1.1 Automotive Intelligence: From Technological Innovation to Market Demand

Since 2020, the global automobile market has entered a stage of parallel development of intelligence and electrification. According to the market research data of Jiazi Guangnian, consumers' demand for intelligent functions such as autonomous driving technology, smart cockpits , and OTA upgrades has increased significantly, especially young consumers, who pay more attention to the performance of automobiles in intelligent driving. This trend reflects that automobile intelligence is no longer just a means of technology display, but has become an important competitive advantage for companies to compete for market share.

1.2 The integration of electrification and intelligence promotes technological upgrading

In the process of technological evolution, electrification has laid a solid foundation for intelligence. Electric vehicles have a higher electrification infrastructure than traditional fuel vehicles, which enables intelligent control and advanced autonomous driving technology to be quickly realized on electric vehicles . In particular, 2022 is known as the "first year of mass production of NOA technology". The penetration rate of NOA (navigation-assisted driving) technology on highways has reached more than 10%, and the penetration rate on urban roads has exceeded 3%. These data fully demonstrate that the widespread popularization of intelligent driving technology has paved the way for future end-to-end autonomous driving.

1.3 Hierarchical Development of Autonomous Driving Technology

On the road to intelligence, autonomous driving technology has shown a multi-level development from L2 (partial automation) to L5 (full automation). L2 and L3 levels of autonomous driving have been mass-produced and applied in actual driving scenarios, while L4 and L5 represent the ultimate goal of fully autonomous driving. End-to-end autonomous driving technology plays an extremely important role in this technological evolution route. It not only breaks through the limitations of modular architecture, but also realizes highly automated global optimization control.

Technical Path and Advantages of End-to-End Autonomous Driving

2.1 Limitations and Challenges of Modular Architecture

The traditional modular autonomous driving architecture relies on multiple independent functional modules, including perception, decision-making, control and planning. These modules are connected in series and make corresponding decisions after processing data step by step. However, the limitations of this architecture are gradually emerging: first, information will be lost when it is transmitted between multiple modules, resulting in low computing efficiency; second, the accumulation of errors between modules may affect the safety of the system. In addition, the modular architecture also requires complex engineering design, and the development and maintenance costs are high.

In the modular architecture, the perception module plays a vital role. The perception module collects environmental information through sensors such as cameras , radars , and lidars , and passes the data to the prediction module. However, the amount of data collected by sensors is huge, and the modular processing method is difficult to strike a balance between real-time and accuracy. As autonomous driving technology matures, this model is gradually replaced by end-to-end autonomous driving technology.

2.2 Advantages of end-to-end autonomous driving architecture

The end-to-end architecture directly converts sensor data into driving decisions by building a unified neural network model, thus avoiding the information loss and computational delay problems in modular design . This type of architecture has higher computational efficiency and stronger generalization ability. The application of BEV (bird's eye view) combined with the Transformer architecture enables the end-to-end solution to handle complex driving scenarios more accurately. Taking Tesla 's FSD V12 as an example, Tesla has achieved adaptive driving of vehicles in complex road environments by building an end-to-end perception-decision-control integrated network. The system has significantly improved the flexibility and accuracy of decision-making through training with a large amount of data, avoiding the error accumulation in traditional modular systems.

2.3 Data-driven global optimization capabilities

The core advantage of the end-to-end solution lies in its global task optimization capability. Traditional modular systems tend to optimize each subtask locally, while the end-to-end architecture can optimize the entire autonomous driving process through a unified network model. This approach not only improves the system's response speed, but also effectively reduces information redundancy and transmission loss between different tasks.

Features and advantages of end-to-end autonomous driving, source: Jiazi Light Years The end-to-end solution can further reduce the need for engineers to manually formulate rules through automated data labeling and model training. The data-driven closed-loop system provides strong data support for the continuous iteration of autonomous driving. As the amount of data increases, the end-to-end autonomous driving system can adapt to complex driving environments more quickly and gradually achieve L4 or even L5 level fully autonomous driving.

Technical implementation and enterprise practice of end-to-end autonomous driving

3.1 Tesla's FSD: Pioneer of end-to-end architecture

Tesla is at the forefront of the industry in the application of end-to-end autonomous driving technology. Its FSD (Full Self-Driving) system has achieved mass production of end-to-end driving in 2024. Tesla has established an efficient end-to-end deep learning model through the accumulation of a large amount of real road data, which greatly improves the vehicle's adaptive ability in complex scenarios. Tesla's computing power reserves will reach 100EFLOPS in October 2024, equivalent to the total computing power of 300,000 Nvidia A100s, which provides strong support for the continuous training and optimization of its end-to-end autonomous driving model. Tesla's successful experience shows that computing power and data are key factors for the efficient operation of end-to-end architecture.

3.2 Wayve’s End-to-End Innovation

Wayve is an autonomous driving technology company based in London, UK, focusing on developing highly adaptable end-to-end systems. Wayve has further improved the perception capabilities of end-to-end autonomous driving systems through its LINGO large model and GAI A visual generation model. Compared with traditional modular systems, Wayve's end-to-end system can handle more complex road scenes, especially in urban traffic environments. Wayve's innovative practice shows that visual generation models have broad application prospects in the field of autonomous driving. Through precise 4D scene reconstruction and video generation, the end-to-end system can accurately model dynamic environments and make more flexible decisions in high-risk driving scenarios.

3.3 Huawei 's ADS 3.0: Transformation from modularization to end-to-end

Huawei's ADS 3.0 is another company that has made important breakthroughs in end-to-end autonomous driving technology. Unlike previous modular systems, ADS 3.0 realizes end-to-end intelligent driving functions through the PDP (Prediction Decision and Planning) network and the GOD large network. This system can not only effectively cope with complex urban traffic environments, but also has high traffic efficiency and self-learning capabilities. Huawei's ADS 3.0 system can complete a model update within 5 days, and the learning mileage reaches 30 million kilometers per day, which provides data and computing power guarantee for the rapid iteration of the system. This data-driven end-to-end system provides a technical foundation for more complex L4 and L5 autonomous driving in the future.

Challenges and solutions for end-to-end autonomous driving

4.1 Computing power and data bottlenecks

Although end-to-end autonomous driving systems have shown great technical potential, the challenges in their practical application remain severe. The efficient operation of end-to-end systems requires strong computing power support. The computing power reserves of domestic manufacturers are still far behind Tesla, and the computing power level of some manufacturers is less than one-tenth of Tesla. This computing power gap limits the progress of domestic manufacturers in end-to-end model training. End-to-end systems are extremely dependent on data. The large autonomous driving model is essentially a process of extracting and compressing driving knowledge from a large number of high-quality driving video clips. High-quality data not only needs to be large in scale, but also needs to have diversity and generalization capabilities to ensure that the system can perform stably in different driving scenarios.