Tesla's FSD has landed, and Chinese players are still far behind

Automotive Business Review Magazine

2024-07-01 18:01Auto Business Review official account

The full text is 3125 words, and it takes about 9 minutes to read, so help me highlight the key points

Highlight the point

01Tesla FSD V12 adopts an end-to-end solution, and the proportion of trips without takeover has increased from 47% to 72%, and the average takeover mileage has increased to 333 miles.

02 The "End-to-end Autonomous Driving Industry Research Report" points out that the core definition of end-to-end autonomous driving technology should be the lossless transmission of perception information and the global optimization of the autonomous driving system.

03The report divides the autonomous driving technology architecture into four phases: perception "end-to-end", decision planning modeling, modular end-to-end, and One Model end-to-end.

04 At present, the players in the market for end-to-end autonomous driving include Tesla, Xpeng Motors, Huawei, NIO, etc.

05The report looks forward to the mass production time of the modular end-to-end solution of domestic autonomous driving companies may be in 2025, and the end-to-end landing time of One Model will be 1~2 years later than that of modular end-to-end.

Technical support is provided by Tencent Hybrid Model

Written by Tu Yanping

Editor / Huang Dalu

Design / Zhao Yongshun

In May 2023, Musk revealed that Tesla's FSD V12 will adopt an end-to-end approach. In March 2024, Tesla began to push V12 on a large scale.

According to statistics from FSD Tracker, a third-party website, after the Tesla FSD V12 update, compared with the previous version, the proportion of trips without takeover has increased from 47% to 72%, and the average takeover mileage (MPI) has increased from 116 miles to 333 miles.

On Tesla's FSD V12, end-to-end self-driving technology shows great strength.

As the development of AI and large-scale model technology enters a new stage, end-to-end technology has become the focus of attention in the autonomous driving industry. So, what is end-to-end? How far has it evolved?

In June 2024, Chentao Capital, Nanjing University Shanghai Alumni Association Autonomous Driving Branch, and Jozzon Intelligent Driving jointly released the 2024 End-to-end Autonomous Driving Industry Research Report (hereinafter referred to as the "Report"), which provides a comprehensive analysis of the basic concepts, participants, development drivers, landing challenges and future prospects of end-to-end autonomous driving.

What end-to-end is

The report makes a systematic conceptual review of end-to-end autonomous driving technology, and proposes a set of terminology systems for reference, and defines the basic concept of end-to-end.

The report points out that in the early days, the core definition of end-to-end was "a single neural network model from sensor input to control output", and in recent years, the concept of end-to-end has been extended to a larger extent. According to the report, the core definition standard of end-to-end should be: lossless transmission of perception information and global optimization of autonomous driving systems.

Based on this definition, combined with the degree of application of AI in autonomous driving systems, the report divides the autonomous driving technology architecture into four stages:

Phase 1: Perception "end-to-end". The perception module realizes "end-to-end" at the module level through BEV (Bird Eye View) technology based on multi-transmitter fusion.

Stage 2: Decision planning modeling. Functional modules from forecasting to decision-making to planning have been integrated into the same neural network.

Phase 3: Modular end-to-end. The perception module no longer outputs results based on human understanding definitions, but more eigenvectors. The comprehensive model of the prediction decision planning module outputs the results of motion planning based on feature vectors. The two modules cannot be trained independently, and must be performed simultaneously by gradient conduction.

Phase 4: One Model end-to-end. There is no longer a clear division of perception, decision-making and planning functions, and the same deep learning model is directly adopted from the input of the original signal to the output of the final planning trajectory.

The latter two phases meet the end-to-end definition criteria described above.

In addition, the report also analyzes the differences and connections between end-to-end and confusing concepts such as large models, world models, and pure vision sensor solutions.

Players from all walks of life

2016年4月，英伟达团队发表了一篇名为End to End Learning for Self-Driving Cars的论文，展示了基于卷积神经网络（Convolutional Neural Network, CNN）的端到端自动驾驶系统DAVE-2。

The system processes the camera image in front of the vehicle via CNN and outputs the steering angle directly. During training, the model learns from simulated driving data. The report calls it a pioneering effort in the field of end-to-end autonomous driving in recent years.

In the current market, there are many players in end-to-end autonomous driving, and the report divides them into several categories:

First, the main engine factory. Such as Tesla, Xpeng Motors, Hongmeng Zhixing, NIO, Zero One, etc.

On May 20, 2024, Xpeng Motors held an AI Day press conference to announce the launch of the end-to-end large model.

On April 24, 2024, Huawei launched a new version of its intelligent driving system, Qiankun ADS 3.0, which implements the modelling of decision-making and planning, and lays the foundation for the continuous evolution of the end-to-end architecture.

In addition, NIO plans to launch end-to-end active safety functions in the first half of 2024, and Zero One Auto, a new energy heavy-duty truck technology company, plans to deploy end-to-end autonomous driving in vehicles by the end of 2024.

Second, autonomous driving algorithm and system companies. Such as Yuanrong Qixing, Jianzhi Robot, SenseTime, Pony.ai, etc.

In 2022, Horizon proposed Sparse4D, an end-to-end algorithm for autonomous driving perception; In 2023, UniAD, the industry's first publicly published end-to-end autonomous driving model published by Horizon Robotics, won the best paper in CVPR 2023.

In August 2023, Pony.ai connected the three traditional modules of perception, prediction, and regulation into an end-to-end autonomous driving model, which has been simultaneously installed in L4 robotaxis and L2 assisted driving passenger cars.

On the eve of the 2024 Beijing Auto Show, NVIDIA revealed the development plan of its autonomous driving business, mentioning that the second step of the plan is to "achieve new breakthroughs in L2++ systems, and put LLM (large language model) and VLM (visual language model) large models on the car to achieve end-to-end autonomous driving."

At the Beijing Auto Show in April 2024, Yuanrong Qixing showcased the high-end intelligent driving platform DeepRoute IO and the end-to-end solution based on DeepRoute IO, which will be mass-produced. SenseTime launched UniAD, an end-to-end autonomous driving solution for mass production, and DriveAGI, a next-generation autonomous driving technology, with the former being a modular end-to-end" type and the latter being a "One Model end-to-end".

During the 2024 Beijing Auto Show, Du Dalong, co-founder and CTO of Jianzhi Robot, revealed that GraphAD, the original end-to-end model of autonomous driving of Jianzhi Robot, has been mass-produced and deployed, and is being jointly developed with leading car companies.

Third, generative AI companies for autonomous driving, such as Lightwheel Intelligence, Lightwheel Intelligence, etc.

Founded in early 2023, Lightwheel Intelligence has developed a self-developed end-to-end data and simulation full-link solution. In September 2023, Lightwheel Intelligence launched DriveDreamer, an autonomous driving world model, which can realize full-link closed-loop simulation of end-to-end autonomous driving, and can be expanded to directly output end-to-end action instructions.

Fourth, academic research institutions, such as Shanghai Artificial Intelligence Laboratory, Tsinghua University MARS Lab, etc.

Many challenges

Since 2023, driven by Tesla's benchmarking role, the AGI technology paradigm represented by large models, and the anthropomorphic and safety needs of autonomous driving, the autonomous driving industry's attention to end-to-end has been heating up, and end-to-end has gradually become the consensus of the autonomous driving industry.

According to a survey of more than 30 front-line experts in the autonomous driving industry, 90% of the respondents said that the company they work for has invested in the development of end-to-end technology.

The report analyzes the challenges faced by the implementation of end-to-end solutions, including technical routes, data and computing power requirements, testing and verification, and organizational resource investment.

First, the divergence of technical routes.

For example, "modular end-to-end" uses a supervised learning training paradigm, and "One Model end-to-end" may focus more on autoregressive and generative training paradigms, both of which companies are betting on. The report judges that in the next 1~2 years, as more companies and research institutions increase investment in the end-to-end field and launch products, the technical route will gradually converge.

Second, the demand for training data has increased at an all-time high.

In the end-to-end technical architecture, the importance of training data has never been greater, and issues related to data volume, data labeling, data quality, and data distribution may become challenges that limit end-to-end applications. The report proposes that synthesizing data and establishing a data sharing platform may be the solution.

Third, the demand for training computing power is getting higher and higher.

Most companies say that 100 GPUs with high computing power can support end-to-end model training. The end-to-end has entered the mass production stage, and the demand for training computing power has increased sharply.

On the 2024Q1 earnings call, Tesla said that it already has 35,000 H100GPU and plans to increase it to more than 85,000 H100 in 2024. Previously, it also deployed a larger A100 GPU training cluster.

In August 2023, Xpeng announced the completion of the "Fuyao" autonomous driving intelligent computing center, with a computing power of up to 600PFLOPS (based on the FP32 computing power of NVIDIA A100 GPUs, which is equivalent to about 30,000 A100 GPUs).

In addition, SenseTime has deployed a nationwide integrated intelligent computing network, with 45,000 GPUs and a total computing power scale of 12,000PFLOPS, which will reach 18,000PFLOPS by the end of 2024.

According to the report, "most companies that develop end-to-end autonomous driving currently have training computing power at the kilocalorie level, and as end-to-end gradually moves towards large models, training computing power will be stretched." ”

Fourth, the test and verification methods are not yet mature.

Existing test and validation methods are not suitable for end-to-end autonomous driving, and the industry urgently needs new test and validation methodologies and tool chains.

Fifth, the challenge of organizational resource input.

End-to-end requires organizational reshaping, and resources need to be tilted to the data side, challenging existing models.

In addition, there is a view that the lack of computing power and interpretability of the vehicle end are the limiting factors for end-to-end implementation, and the report puts forward the opposite conclusion.

Outlook: Mass production in 2025

The report looks forward to the mass production time of the modular end-to-end solution of domestic autonomous driving companies may be in 2025, and the end-to-end landing time of the One Model will be 1~2 years later than the modular end-to-end time, and the mass production will begin from 2026 to 2027.

This will lead to the progress of upstream technology and the evolution of the market and industrial structure.

In terms of technology, the end-to-end implementation will accelerate the progress of the upstream tool chain and chips on which it depends; On the market side, the improvement of the end-to-end autonomous driving experience will bring about an increase in the penetration rate of high-level assisted driving, and may also drive the application of autonomous driving across geographical regions, countries and scenarios. In terms of the industrial landscape, end-to-end has further increased the importance of data and AI talents, which may give rise to new industrial divisions and business models.

In the early stage of development, autonomous driving drew on the accumulation of many levels of perception algorithms, planning algorithms, middleware and sensors in the robotics industry. In recent years, autonomous driving technology and industry maturity have improved, and the AGI technology paradigm provided by end-to-end autonomous driving has a feedback effect on the general humanoid robot industry.

The report believes that with the continuous progress of AI technology, the integration and mutual learning of autonomous driving and robotics technology will be deepened, and AGI (general artificial intelligence) will jointly promote AGI (general artificial intelligence) to the physical world and create greater social value.

View original image 863K

• 

• 

• 

• 

• 

• 

• 

• 

•