Recently, Tesla announced that it will launch the FSD (Fully Autonomous Driving) feature in the Chinese market in the first quarter of next year. This news has once again sparked the automotive industry's attention to "end-to-end" technology. The current Tesla FSD V12 solution is based on iterative promotion of end-to-end technology. In response to challenges, since the beginning of this year, many companies, including car companies such as "Weixiaoli" and service providers such as Huawei and Horizon Robotics, have increased their efforts in end-to-end autonomous driving technology. Some industry insiders even believe that "end-to-end" is currently the only feasible solution to the end of autonomous driving.
However, the author believes that it is still too early to draw this conclusion.
The so-called 'end-to-end' actually originates from the concept of deep learning, which refers to an AI model that only needs to input raw data to output the final result. Applied to the field of autonomous driving, this means that with just one model, the perception information collected by sensors such as cameras, millimeter wave radar, and LiDAR can be converted into specific operational instructions such as the rotation angle of the vehicle's steering wheel, the depth of the accelerator pedal, and the braking force, enabling the car to achieve autonomous driving.
Compared with traditional "perception decision control" intelligent driving systems, "end-to-end" technology simplifies the original architecture of combining multiple models such as perception, prediction, and planning into a single model architecture of "perception decision integration" due to the lack of rule intervention in the middle, which has more advantages in information transmission, inference calculation, and model iteration. At the same time, the "end-to-end" architecture is built with data-driven modules, reducing the proportion of manually maintained modules. Therefore, the "end-to-end" system can not only significantly improve computing efficiency, but also reduce maintenance costs.
However, these advantages do not mean that "end-to-end" technology has no shortcomings. In fact, "end-to-end" autonomous driving still faces many challenges in achieving mass production and popularization.
Firstly, the training cost of "end-to-end" models is very high, requiring manufacturers to continuously increase their GPU procurement scale, which raises the entry threshold for "players". Taking Tesla as an example, currently its FSD has accumulated over 20 million human driving video clips for learning, and the collection cost alone requires 5 billion to 8 billion yuan.
Secondly, the training of "end-to-end" models requires a large amount of data, and the collection, cleaning, and screening of massive amounts of data is a highly challenging task for enterprises. Powerful computing power can process massive amounts of data in real-time, reduce data transmission latency, and facilitate the implementation of high-level autonomous driving. However, many car companies, including "Weixiaoli", are currently facing significant challenges in terms of computing power growth.
In addition, the safety of end-to-end intelligent driving has not been fully guaranteed, especially for data in uncommon or extreme scenarios, which puts extremely high demands on the generalization performance of perception models. The 'end-to-end' technology needs to be able to handle these long tail scenarios to ensure proper response in any situation.
At present, the exploration of "end-to-end" technology in the industry is just beginning, and there is still a lack of in-depth discussion on many fundamental issues. Therefore, it is currently uncertain whether "end-to-end" is the ultimate solution for autonomous driving, thus abandoning other paths. Only by competing and developing different technological routes together can we promote the safer and faster development of autonomous driving to higher stages.