With the determination of the timing of FSD's entry into China, the "end-to-end" autonomous driving technology route has once again sparked heated discussions within the industry.
During the earnings call on the early morning of July 24th, Tesla CEO Elon Musk stated that he will apply for regulatory approval in Europe and China to implement supervised FSD, and is expected to receive approval before the end of this year.
In fact, starting from March 2024, Tesla has been widely promoting FSD v12 in North America. This end-to-end intelligent driving system performs excellently, allowing practitioners and users to experience the improved driving experience brought by intelligent technology, and is also an important driving force for the formation of a large-scale consensus on the end-to-end autonomous driving technology roadmap in the short term.
Xiaopeng Motors Chairman He Xiaopeng recently stated on social media that Tesla's FSD system has made significant technological improvements compared to the past this year, and he appreciates this progress.
He also mentioned that he recently discussed the application of end-to-end technology with several leaders of L4 autonomous driving projects. Although these officials believe that end-to-end technology is more suitable for L2 or L3 level autonomous driving, He Xiaopeng firmly believes that combining end-to-end technology with large models will ultimately achieve L4 level autonomous driving.
How fragrant is competing to enter the game end-to-end?
Simply put, 'end-to-end' refers to an AI model that can output the final result as long as the raw data is input. When applied to intelligent driving in automobiles, it means allowing the vehicle to automatically take you from point A to point B.
When describing FSD Beta v12, Musk mentioned that it has the ability to input images, output vehicle control signals such as steering, braking, and acceleration, making it "end-to-end".
An algorithm solution development engineer told a reporter from Huaxia Times that integrating perception and decision-making into the same model enables end-to-end models to effectively avoid errors between levels without any manual rule intervention, making it a more advanced intelligent driving approach that is closer to human driving behavior.
He also mentioned that due to end-to-end technology eliminating the boundaries between modules and simplifying the system architecture, it has improved operational efficiency, enabling the integrated model to process data faster and enhance the system's response speed. At the same time, it also reduces reliance on LiDAR and high-precision maps, lowering costs.
Actually, 'end-to-end' is not a completely new concept.
In 2016, Nvidia proposed using a single neural network to achieve end-to-end autonomous driving. However, due to the overly simple structural design and small scale of the model, this solution can only support autonomous driving under high-speed or simple road conditions, and has only completed small-scale demo verification.
Until Tesla announced the FSD V12 version in August 2023, mentioning the introduction of end-to-end technology, it became the hottest concept in the autonomous driving industry.
In China, companies such as Xiaopeng, NIO, Ideal, Huawei, Great Wall, SenseTime, Yuanrong Qixing, and Horizon have actively followed suit and successively launched end-to-end autonomous driving solutions and models for mass production.
On January 30th, He Xiaopeng stated that Xiaopeng Intelligent Driving will achieve end-to-end model integration in the future. On May 20th, Xiaopeng Motors held an "AI DAY" in Beijing and announced that it will start pushing intelligent driving and cabin systems based on end-to-end big models to users from today.
On April 24th, at the Huawei Intelligent Automotive Solution Conference, Huawei unveiled its new intelligent automotive solution brand, Qiankun, with intelligent driving as its core, and launched ADS 3.0, which adopts an end-to-end architecture. It is reported that the Xiangjie S9, which was just launched in June, has already been equipped with the ADS 3.0 intelligent driving system for the first time. NIO also announced its end-to-end cloud computing power scale in April and revealed that the end-to-end solution will be released within this year.
On July 5th, Ideal Auto unveiled its end-to-end autonomous driving technology architecture for the first time at the 2024 Intelligent Driving Summer Conference.
Traditional car companies are also unwilling to be outdone. On April 15th, during the live debut of Wei Jianjun, Chairman of Great Wall Motors, the end-to-end intelligent driving solution equipped on the new Wei brand Blue Mountain model attracted attention from the outside world.
In terms of industrial chain, Yuanrong Qixing and Shangtang Jueying each showcased end-to-end products at the 2024 Beijing Auto Show. The former showcases the upcoming production of the advanced intelligent driving platform DeepRoute IO and end-to-end solutions based on DeepRoute IO, while the latter launches the production oriented end-to-end autonomous driving solution "UniAD".
Horizon released the SuperDrive full scenario intelligent driving solution in May, which uses a perception end-to-end architecture that integrates dynamic, static, and occupancy networks. At the same time, Horizon has also designed and developed data-driven interactive games, no longer rule-based decision networks.
The challenge is arduous and the commercialization path has not yet been opened up
Despite its great potential, according to the 2024 "End to End Autonomous Driving Industry Research Report" jointly released by Chentao Capital and multiple institutions, end-to-end technology is still in its early stages of development, and there are still many application challenges and pain points that need to be solved urgently, such as divergent technological routes, high demand for data and computing power, weak interpretability, and the lack of consumer payment awareness.
Firstly, there is currently no unified understanding of the concept of end-to-end in the field of autonomous driving. There are different viewpoints on the technical roadmap and system architecture, such as modular joint end-to-end, One Model end-to-end, and LLM/world model based end-to-end methodologies.
Secondly, in terms of data and computing power reserves, in recent years, both mainstream domestic automotive companies and emerging car manufacturing forces have accelerated the construction of computing power reserves to meet the training requirements of autonomous driving models. However, compared to Tesla, there is still a significant gap in the computing power level of domestic manufacturers.
At the 2024 Q1 earnings conference call, Tesla announced that it already has 35000 H100 GPUs and plans to increase it to over 85000 H100 GPUs by 2024, reaching the same level as Google and Amazon.
In China, most companies that develop end-to-end autonomous driving currently only have a training computing power scale of kilocalories. As the end-to-end approach gradually moves towards larger models, the training computing power becomes insufficient. Moreover, against the backdrop of GPU export restrictions in the United States, domestic enterprises still have a long way to go to reach international first-class computing power reserves, "said Yuan Rongqixing, a relevant person in charge, to the reporter of Huaxia Times.
Once again, when problems arise, end-to-end models are unable to analyze intermediate results like traditional autonomous driving tasks, resulting in weak interpretability and difficulty in providing effective evidence for accidents, after-sales accountability, and directly improving the problematic areas.
Finally, consumer awareness of software payment has not yet formed, and the commercial closed loop of autonomous driving has not yet been implemented.
In the increasingly fierce "involution" competition in the automotive industry, consumers face an inherent contradiction when buying a car. McKinsey pointed out in the "2024 McKinsey China Automotive Consumer Insights" report that consumers' interest in various autonomous driving functions is increasing, but their willingness to pay extra for them is decreasing, with a particularly significant decline in willingness to pay in first tier cities.
At the same time, the subscription service model has not yet achieved large-scale popularization, and the autonomous driving function has become a "hard cost" that car manufacturers have to bear in order to improve the driving experience and enhance product quality.
Although there are many difficulties, the temptation of end-to-end technology for intelligent driving is too great. Currently, domestic enterprises are continuously promoting the implementation of end-to-end technology, which will bring different problem-solving ideas.