Recently, "end-to-end" has become popular in the automotive industry! Tesla's benchmark demonstration effect based on the "end-to-end" FSDV12 (fully autonomous driving) solution, coupled with rumors of its entry into China, has driven car companies such as "Weixiaoli" and service providers such as Huawei and Horizon to shift their focus towards end-to-end autonomous driving technology. The so-called "end-to-end" actually comes from the concept of deep learning, which is "End to End (E2E)" in English, referring to an AI model that can output the final result as long as the raw data is input. Applied to the field of autonomous driving, it means that only one model is needed to convert the perception information collected by sensors such as cameras, millimeter wave radar, and laser radar into practical commands such as the rotation angle of the vehicle steering wheel, the depth of the accelerator pedal, and the braking force, enabling the car to achieve autonomous driving. In the words of He Xiaopeng, the founder of Xiaopeng Motors, the performance is "very smooth" and more like "human driver driving". Previously, most of the auto drive system on the market were traditional modular, that is, a hybrid system of artificial and intelligent: perception depended on neural networks, and planning control used algorithms designed manually by humans. The advantage of this system is that it has clear division of labor, making it easy to detect and solve defects in modules. However, the problem is that this modular auto drive system performs well in relatively simple driving tasks, while its ceiling is obvious in front of complex driving tasks. Even the advanced intelligent driving features claimed to be far ahead in cities still have a mechanical feel and may crash when merging into expressways or passing through large intersections. Considering that the core challenge of autonomous driving is to solve endless edge scenarios, the cost and time of solving the infinite long tail problem with limited manpower are difficult to estimate, and dataization and modeling have become inevitable trends. However, end-to-end is also a highly challenging technical task that requires careful polishing by experienced professionals. On the one hand, end-to-end training requires massive amounts of high-quality data to be fed. Unlike the big language model, which can crawl massive text data on the Internet for training, end-to-end smart driving requires extremely high cost and difficulty in acquiring video data. Taking Tesla as an example, currently its FSD has accumulated over 20 million human driving video clips for learning, and the data collection cost alone for this scale requires 5 billion to 8 billion yuan. On the other hand, end-to-end support requires powerful computing power. Autonomous driving involves technologies and solutions such as LiDAR, image perception, and V2X vehicle road collaboration. Powerful computing power is not only beneficial for real-time processing of massive data and reducing data transmission latency, but also better supports full scenarios such as smart cities, smart transportation, and high-level autonomous driving. However, domestic enterprises such as Huawei Auto BU, Baidu Jiyue, NIO, Ideal, Geely, Great Wall, and Xiaopeng are currently facing significant bottlenecks in their computing power growth. The problem lies in the fact that the constraints of computing power and data can significantly affect the development of algorithms. Although the end-to-end autonomous driving model UniAD proposed by the domestic academic community won the Best Paper Award of CPVR in 2023, providing a reference direction for domestic enterprises, the development of UniAD under open-loop verification system and small sample data still requires some time for engineering transformation and large-scale data training to get on board. In addition, the upper and lower limits of the auto drive system will be amplified from end to end. Because the end-to-end construction is a neural network black box, it sacrifices some of the interpretability of traditional modular solutions in the process of obtaining higher upper limits. How to retain the interpretability in the auto drive system, and characterize the rules that should not be overridden, such as not running the red light, into the neural network, so as to ensure the end-to-end safe landing, application and evolution, will also be an important topic for the planning and control engineers. There are two routes to climb Mount Everest: one is the northern slope of Xizang, China, and the other is the southern slope of Nepal. Regardless of whether you choose to climb from the south or north slope, you will eventually reach the same peak. This is similar to the current development path of autonomous driving. Although it is still difficult to determine whether end-to-end is the optimal or final solution for autonomous driving, this does not hinder enterprises from exploring innovation. After all, end-to-end can better handle extreme cases than traditional modular methods and represents a more efficient approach to reducing manual coding dependencies. Based on this path, perhaps autonomous driving can lead to higher stages. (New Society)