In 2024, the core of the development of China's new energy vehicle industry is one word - volume. The volume price, volume configuration, volume to the innumerable, and even the intelligent driving that began to push the full amount has also started the "volume" technology.
From last year's hot BEV/Occupancy Network, to this year's even hotter end-to-end. The technology of intelligent driving has never been as fierce as in the past two years, just like after the technology is clear, all car companies and suppliers are sprinting through the brand.
Recently, two major events have happened in the intelligent driving circle that are worth talking about.
First, He Xiaopeng, chairman of Xiaopeng Motors, issued a document highly praising Tesla's "excellent performance" in autonomous driving.
He Xiaopeng said: "In the past two days, I have experienced the version of FSD V12.3.6 in California, USA, and I have also played a Waymo experience. Overall, Waymo performs better in downtown San Francisco, while FSD performs extremely well in Silicon Valley and on the highway, can achieve high scores, and many road conditions are handled silky. I am very impressed that FSD has made great progress in the past few months, and we will also learn from FSD their excellent function points and user experience parts, and I believe that 2025 will be the moment of fully autonomous ChatGPT! In the autonomous driving industry, everyone is learning from each other quickly, iterating on their own, and improving everyone's happiness together. ”
Second, Wei Xiaoli and the three have coincidentally completed the organizational adjustment of the intelligent driving team.
NIO has set up a separate large model department to be responsible for end-to-end model research and development.
In the layoffs in May, Ideal retained the algorithm R&D team: managed by Jia Peng, mainly responsible for the R&D and implementation of NOA in cities without maps, as well as the pre-research of end-to-end intelligent driving.
In the Xpeng Zhijia team, there are at least 3 departments directly related to data, including big data, spatio-temporal data, and American platform support data, etc., and the division of labor is very fine.
According to relevant reports, most of the all-in-end-to-end companies are reducing the size of their original teams and adjusting the focus of their teams to AI large models and data infrastructure.
Everything aims at end-to-end and everything serves end-to-end, which has become the consensus of the industry.
Take a brief look at the end-to-end landing time of the leading car companies.
Xpeng Motors is the first domestic vehicle company to release an end-to-end model for mass production. He Xiaopeng said in May this year that by the third quarter of 2024, Xpeng Motors' goal is to "be able to drive all over the country and every road", and realize the intelligent driving experience of intelligent driving in urban areas comparable to high-speed in 2025.
On April 24, at Huawei's intelligent vehicle solution conference, Huawei released the ADS 3.0 end-to-end architecture, which will be officially launched with the Xiangjie S9 in August.
On the Li Auto side, a smart driving conference will be held on July 5, and it is expected to announce its own end-to-end car plan. Ren Shaoqing, vice president of autonomous driving at NIO, said in an interview that NIO has been laying out end-to-end and is expected to achieve mass production this year.
In fact, although most of the teams have already adjusted and given an end-to-end mass production time, referring to Tesla's "performance ramp-up" from FSDv11 to v12, the early performance of the end-to-end model may not be stronger than the classic technical solution optimized by the extreme optimization, which also tests the management's determination to transform the technical roadmap.
On the one hand, the end-to-end transformation will reduce the personnel size of the existing intelligent driving team, and on the other hand, the end-to-end investment in data and data infrastructure will increase.
For some company management, it is easy to make the decision to expand the team, but the investment of "invisible and intangible" resources such as data and tool chains requires a switch of cognitive mode.
Much of the controversy surrounding end-to-end autonomous driving stems from the lack of clarity in the definition of its concept, and the industry's views on the technology are polarized:
"Tech fundamentalists" will argue that the "end-to-end" hype promoted by many companies in the market is not really end-to-end;
"Pragmatists" will argue that as long as the fundamentals are in line and the product is performing, the end-to-end accuracy doesn't matter.
At present, the evolution of autonomous driving architecture can be divided into four main stages:
Phase 1: Perception "end-to-end"; The second stage: modelling of decision planning; Phase 3: Modular end-to-end; Phase 4: One Model End-to-End.
One Model end-to-end autonomous driving can be seen as an end-to-end solution. This concept was proposed much earlier than modular end-to-end.
In 2016, when the autonomous driving industry was just beginning, NVIDIA proposed to use a single neural network (convolutional + fully connected simple architecture) to achieve end-to-end autonomous driving, and the input and output are the most primitive sensor signals, steering wheel angles and throttle openings.
With the improvement of the Transformer network architecture and vehicle-side computing power (which can gradually support 0.1B~1B parameter network operation), the end-to-end solution of One Model has returned to people's field of vision, and will even evolve into an end-to-end final solution.
Based on this judgment, there is another concept that needs to be clarified. The concepts of end-to-end and large models are often confused.
But in reality, the two are not necessarily related. Large models pay more attention to the number of parameters and emergence, while end-to-end emphasizes more on the gradient conduction and global optimization of the structure.
The current large model provides a good alternative for end-to-end implementation, but end-to-end is not necessarily based on the large model implementation.
In terms of goals, large models are like generalists, they have a large number of parameters, and can handle a variety of complex tasks, including autonomous driving, such as natural language processing, image recognition, etc.
The data types required for large model training have a wider structure, and due to their wide range of applications, the requirements for interpretability and reliability are not high.
In contrast, the end-to-end goal is to enable the vehicle to navigate autonomously and drive safely. It needs to learn and adapt to various driving scenarios in order to make accurate driving decisions, and more training is based on annotated data, which requires higher reliability of the system.
On the relationship between the world model and end-to-end autonomous driving, there are basically two different views:
First, the world model exists as a data generator in the development system of the entire autonomous driving algorithm. Second, as long as the world model is slightly adjusted or some output links and modules are added, the end-to-end autonomous driving of the One Model can be realized. Both views are subject to future developments.
Back to Tesla's fiery end-to-end solution.
Tesla announced that it will adopt an end-to-end autonomous driving solution starting with FSD 12.3. It is generally accepted in the industry that these huge improvements in performance are largely based on the modular basic solution for decision planning, but it is not yet possible to determine which stage of Tesla's implementation is in the "decision planning modeling", "modular end-to-end", or "one model end-to-end" as defined in this report.
For domestic players such as Huawei and Xpeng, they are generally still in the stage of perceiving "end-to-end" and modelling decision-making planning.
Referring to Tesla's announcement of the technical architecture of BEV and Occupancy Network at the AI Day at the end of 2021 and 2022 respectively, and the domestic car companies began to OTA BEV/Occupancy Network-based functions in 2023-2024, and the R&D progress gap between China and Tesla is about 1.5~2 years.
2025 will be the first year of the full mass production of end-to-end solutions for domestic car companies, and it may also be the time point when Tesla's FSD officially enters China.
The One Model end-to-end system is expected to be implemented 1~2 years later than the modular end-to-end system, and will be mass-produced from 2026 to 2027.
Despite the popularity of end-to-end technology, we are still not sure that it is the ultimate solution for autonomous driving.
Optimistically, from the analysis of the characteristics and development speed of artificial intelligence, the emergence of artificial intelligence may indicate that end-to-end is only one of the many technical solutions in the future.
However, it is undeniable that whether it is the rapid progress of academia and industry in the field of end-to-end autonomous driving, or the attention of the capital market to related companies, they all point to a trend: end-to-end has opened a new round of autonomous driving industrial revolution.