With the deepening influence of intelligence and informatization in the global automotive industry, consumer demand for intelligent driving features has significantly increased. The level of intelligence has become a crucial factor in purchasing decisions, especially among young consumers who prioritize "intelligence" as the core of their car-buying decisions. The importance of intelligence in car-buying decisions is increasingly prominent, prompting many automakers to accelerate their R&D and engineering investments in intelligent technologies. Automakers that fail to keep pace with this trend may gradually lose market competitiveness. The development of autonomous driving technology has entered a critical stage, and the introduction of the end-to-end concept has made the path to 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 automotive market has entered a stage of parallel development of intelligence and electrification. Market research data from Jiaozi Guangnian shows that consumer demand for intelligent functions such as autonomous driving, smart cabins, and OTA upgrades has surged, particularly among young consumers who place greater emphasis on a car's intelligent driving performance. This trend reflects that automotive intelligence is no longer merely a technological showcase but has become a significant competitive advantage for companies vying for market share.

Autonomous Driving, Smart Cabins, and OTA Capabilities Gain Widespread Attention, Source: Jiaozi Guangnian

1.2 The Convergence of Electrification and Intelligence Drives Technological Upgrades

In the process of technological evolution, electrification has laid a solid foundation for intelligence. Electric vehicles (EVs) have superior electrical infrastructure compared to traditional internal combustion engine vehicles, enabling intelligent control and advanced autonomous driving technologies to be quickly implemented in EVs. Notably, 2022 was hailed as the "first year of NOA technology mass production," with NOA (Navigation Assisted Driving) technology penetration exceeding 10% on highways and breaking the 3% threshold on urban roads. These figures demonstrate that the widespread adoption of intelligent driving technologies has paved the way for future end-to-end autonomous driving.

Continuous Growth in Highway NOA Penetration (%), Source: Jiaozi Guangnian

Continuous Growth in Urban NOA Penetration (%), Source: Jiaozi Guangnian

1.3 Hierarchical Development of Autonomous Driving Technology

On the path to intelligence, autonomous driving technology exhibits a multi-layered development from L2 (Partial Automation) to L5 (Full Automation). L2 and L3 autonomous driving have achieved mass production and are applied in real-world driving scenarios, while L4 and L5 represent the ultimate goal of fully autonomous driving. End-to-end autonomous driving technology plays a crucial role in this technological evolution. It not only transcends the limitations of modular architectures but also achieves highly automated global optimization control.

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

2.1 Limitations and Challenges of Modular Architectures

Traditional modular autonomous driving architectures rely on multiple independent functional modules, including perception, decision-making, control, and planning. These modules process data sequentially and make corresponding decisions. However, the limitations of this architecture are gradually emerging: firstly, information loss occurs during transmission between modules, leading to inefficient computation; secondly, error accumulation between modules may compromise system safety. Additionally, modular architectures require complex engineering designs, resulting in high development and maintenance costs.

Modular Deployment of Traditional Autonomous Driving, Source: Jiaozi Guangnian

In modular architectures, the perception module plays a crucial role. It collects environmental information through cameras, radars, LiDARs, and other sensors, transmitting data to the prediction module. However, the sheer volume of sensor data makes it difficult for modular processing to balance real-time performance and accuracy. As autonomous driving technology matures, this model is gradually being replaced by end-to-end autonomous driving technology.

2.2 Advantages of End-to-End Autonomous Driving Architecture

The end-to-end architecture converts sensor data directly into driving decisions through a unified neural network model, avoiding information loss and computational delays inherent in modular designs. This architecture offers higher computational efficiency and stronger generalization capabilities. The integration of Bird's Eye View (BEV) and Transformer architectures enables end-to-end solutions to handle complex driving scenarios more accurately.

Taking Tesla's FSD V12 as an example, Tesla has achieved adaptive driving in complex road environments by constructing an end-to-end perception-decision-control integrated network. This system, trained on vast amounts of data, significantly enhances decision-making flexibility and accuracy, avoiding error accumulation in traditional modular systems.

2.3 Data-Driven Global Optimization Capabilities

The core advantage of end-to-end solutions lies in their global task optimization capabilities. Traditional modular systems tend to optimize subtasks locally, whereas end-to-end architectures optimize the entire autonomous driving process through a unified network model. This approach not only improves system response speed but also effectively reduces information redundancy and transmission loss between different tasks.

Characteristics and Advantages of End-to-End Autonomous Driving, Source: Jiaozi Guangnian

End-to-end solutions further reduce the need for engineers to manually define rules through automated data annotation and model training. Data-driven closed-loop systems provide robust data support for continuous iteration in autonomous driving. As data volumes increase, end-to-end autonomous driving systems can adapt more quickly to complex driving environments and gradually achieve full autonomy at L4 or even L5 levels.

Technical Realization and Corporate Practices of End-to-End Autonomous Driving

3.1 Tesla's FSD: Pioneering End-to-End Architecture

Tesla leads the industry in the application of end-to-end autonomous driving technology, with its Full Self-Driving (FSD) system achieving mass production of end-to-end driving in 2024. Through the accumulation of vast amounts of real-world road data, Tesla has established an efficient end-to-end deep learning model, significantly enhancing the vehicle's adaptive capabilities in complex scenarios.

Tesla's computing power will reach 100 ExaFLOPS by October 2024, equivalent to the combined power of 300,000 Nvidia A100 GPUs, providing robust support for continuous training and optimization of its end-to-end autonomous driving model. Tesla's success demonstrates that computing power and data are crucial factors for the efficient operation of end-to-end architectures.

3.2 Wayve's End-to-End Innovation

Wayve, a London-based autonomous driving technology company, specializes in developing highly adaptable end-to-end systems. Through its LINGO large model and GAIA visual generation model, Wayve has further enhanced the perception capabilities of end-to-end autonomous driving systems. Compared to traditional modular systems, Wayve's end-to-end system can handle more complex road scenarios, especially in urban traffic environments.

Wayve's innovative practices show that visual generation models have broad application prospects in autonomous driving. Through precise 4D scene reconstruction and video generation, end-to-end systems can accurately model dynamic environments and make more flexible decisions in high-risk driving scenarios.

3.3 Huawei's ADS 3.0: Transition from Modular to End-to-End

Huawei's ADS 3.0 is another significant breakthrough in end-to-end autonomous driving technology. Unlike previous modular systems, ADS 3.0 achieves end-to-end intelligent driving functions through the Prediction, Decision, and Planning (PDP) network and GOD large network. This system not only effectively navigates complex urban traffic environments but also offers high traffic efficiency and self-learning capabilities.

Huawei's ADS 3.0 system can update its model once every five days, with a daily learning mileage of 30 million kilometers, providing data and computing power support for rapid iteration. This data-driven end-to-end system lays the technical foundation for more complex L4 and L5 autonomous driving in the future.

Challenges and Solutions for End-to-End Autonomous Driving

4.1 Bottlenecks in Computing Power and Data

Despite the great technical potential of end-to-end autonomous driving systems, challenges in practical applications remain severe. Efficient operation of end-to-end systems relies on powerful computing support. Domestic manufacturers lag significantly behind Tesla in computing power, with some having less than one-tenth of Tesla's capacity. This gap limits progress in end-to-end model training for domestic manufacturers.

End-to-end systems heavily rely on data. Essentially, large autonomous driving models extract and compress driving knowledge from extensive, high-quality driving video clips. High-quality data must be large-scale, diverse, and generalizable to ensure stable system performance across different driving scenarios.

4.2 The Issue of Unexplainability

Another significant challenge for end-to-end systems lies in their "black box" nature. As end-to-end architectures achieve global optimization through deep learning, the decision-making process within the model is often difficult to explain. This poses a major obstacle in the highly safety-critical field of autonomous driving. When the system malfunctions, it is challenging to trace error sources through traditional logical analysis, increasing risk management complexity.

4.3 Challenges in Commercialization Loops

Despite significant technological progress in end-to-end autonomous driving, its commercialization faces hurdles. While consumer interest in intelligent driving features has increased, willingness to pay for them has declined. This suggests that autonomous driving features may increasingly become a cost borne by automakers to enhance product competitiveness rather than a direct revenue source.

Future Prospects for End-to-End Autonomous Driving

5.1 Technological Trend: From Data-Driven to AGI

With the continuous advancement of AI technology, the technological path for end-to-end autonomous driving systems is becoming clearer. In the future, with the widespread application of large models and multimodal models, end-to-end systems will further enhance their performance in complex driving scenarios. Especially in areas such as multimodal data fusion and self-supervised learning, end-to-end systems are expected to break through existing technological bottlenecks and achieve higher levels of automation.

5.2 Transformation and Opportunities in Business Models

Although the commercialization loop for end-to-end autonomous driving is not yet fully formed, automakers may find new profit points in areas such as OTA upgrades and subscription services in the future. End-to-end systems not only enhance driving experiences but also continuously optimize through data accumulation and self-learning capabilities, providing consumers with a "better driving experience over time."

5.3 Driving Forces from Policies and Regulations

In the future, the further popularization of autonomous driving technology will be driven by policies and regulations. End-to-end systems need to meet higher standards in safety and explainability to satisfy the requirements of regulatory agencies worldwide. As technology matures and regulations improve, end-to-end autonomous driving systems are expected to achieve commercialization in more regions.

Conclusion

End-to-end autonomous driving technology represents the primary development direction for future autonomous driving. Driven by large-scale data, powerful computing support, and continuous optimization of deep learning models, end-to-end systems lay the foundation for achieving L4 and L5 autonomous driving. However, current challenges include computing power bottlenecks, data demands, and unexplainability issues. With continuous technological iteration and market maturity, end-to-end autonomous driving will achieve wider application in the coming years, bringing new development opportunities to the field of intelligent driving.