Huawei leads the way, and China's L3 autonomous driving breakthrough is in sight

03. Algorithm and software iteration

In 2016, Tesla began to collect a large amount of data on the vehicle side. By 2018, it had initially built a data closed-loop system and gradually improved cloud computing resources, automated labeling, simulation and other links. By 2023, Tesla's perfect data closed-loop system has created its ultimate model iteration speed, and the public beta version of FSD BetaV11 is once every 20 days. At present, smart driving manufacturers represented by Xiaopeng and Huawei have gradually improved their infrastructure and data closed-loop systems and have become the leaders of China's ADAS industry. Huawei's engineering experience accumulated in multiple fields such as chips, communications , and mobile terminals can deeply empower the smart driving data closed-loop system. Huawei can develop its own cloud training chips and vehicle-side smart driving chips. There are very few manufacturers who can do this. Therefore, the company can achieve true deep collaboration between software and hardware to improve efficiency.

Yu Chengdong disclosed at the launch of the QM7 facelift in September 2023 that its cloud computing power is 1.8E FLOPS, which can learn 10 million kilometers of data every day. By November 2023, the cloud computing power has increased to 2.8E FLOPS (2-3 times the computing power resources of other domestic manufacturers), which can learn 12 million kilometers of data every day and iterate the model every 5 days. Xiaopeng is the first manufacturer among the new forces of automobile companies to establish a data closed-loop system and deploy a cloud supercomputing center. Since 2023, the company's data closed-loop efficiency has been greatly improved, which is reflected in the efficiency improvement of the entire chain of data collection, model training, deployment, and simulation.

In the simulation phase, it will be possible to simulate data based on real scenarios in 2022, and it will have the ability to use AI to generate extreme scenarios and integrate them into massive training data in 2023. With the support of such data closed-loop capabilities, Xiaopeng's software release speed has been significantly improved.

Starting from the second half of 2023, the iteration speed of intelligent driving function software versions has significantly accelerated. In the fourth quarter of 2023, the city NOA landing goals set by the leading OEMs at the beginning of the year have begun to be fulfilled in a concentrated manner. At the same time, commuting NOA is also becoming popular. Commuting NOA, also known as memory driving or AI driving, refers to point-to-point single-path navigation assisted driving on a specific route set by the user after learning on the vehicle side. The algorithm technology stack of commuting NOA is exactly the same as that of urban NOA, but the scope of the scene is greatly reduced, and a lightweight, high-precision map of a single path is "rebuilt" by learning the same route multiple times. In terms of hardware cost, because the working conditions under a single path are relatively simple and controllable, the generalization requirements of the algorithm model are also relatively low, and the demand for chip computing power and sensors is also much lower than that of urban NOA.

From the actual application effect, highway + commuting NOA has covered 85% of users' travel scenarios. Therefore, for car companies, in the early stage of large-scale promotion of urban NOA functions, taking the lead in implementing commuting mode can not only meet user needs to the greatest extent possible under limited capabilities, and gradually cultivate users' usage habits of advanced intelligent driving functions, but also provide data accumulation for OEMs to upgrade or promote urban NOA functions in the future. Under the demonstration effect of leading car companies, many vehicle manufacturers such as BYD, Zhiji, and Leapmotor have put the launch schedule of commuting NOA functions on the agenda.

04. Obstacles to L3

Unlike L2/L2+, L3 is no longer considered assisted driving, but conditional autonomous driving. The driving task of the vehicle will be mainly the responsibility of the intelligent driving system itself, and the driver does not need to be ready to take over at all times. However, judging from the current development situation, it is difficult to commercialize L3 on a large scale. In addition to technical factors, regulations and ethics are insurmountable obstacles. In particular, regulatory issues are quite complicated. Autonomous driving technology involves driving safety and life. Especially for L3 and above autonomous driving technology, the driving responsibility is more borne by the vehicle itself, and the uncertainty of driving is further increased.

Therefore, governments around the world are cautious about the implementation of high-level autonomous driving, and the pace of the introduction of relevant supporting laws and regulations is relatively slow, which to a certain extent restricts the development of high-level autonomous driving technology. The main operations of L3 and above autonomous driving vehicles are completed by the vehicle itself. Therefore, traffic accidents that occur when the autonomous driving system is operating normally should be the responsibility of the vehicle manufacturer. However, judging from the current traffic policies of various countries, L3 technology has not been widely recognized, and the first person responsible for traffic accidents is mostly identified as the driver.

Taking the current "Road Traffic Safety Law of the People's Republic of China" as an example, it clearly stipulates that "during driving, the driver shall not engage in any behavior that affects safe driving", indicating that the driver still needs to be responsible for driving tasks at all times and will be the first person responsible when an accident occurs. The current US federal traffic regulations have made clear division of responsibilities for traffic accidents involving self-driving cars, stipulating that "when a traffic accident occurs with a self-driving car, the human driver as the backup driver must bear responsibility", and the regulations also add that car manufacturers do not evade responsibility for traffic accidents.

The "Autonomous Driving Law" promulgated by Germany in 2021 stipulates: "Level 3 autonomous driving cars can be driven on 13,200 kilometers of highways throughout Germany at a speed not exceeding 60 km/h. The hands can be freed but the driver cannot sleep. The driver is not allowed to look back continuously or leave the seat. The driver is still required to take over the vehicle when necessary." If a vehicle that meets the above conditions is involved in a traffic accident, the responsibility will belong to the vehicle manufacturer. Relevant laws in Japan stipulate that when an L3 autonomous driving vehicle in its territory has an accident, the driver shall bear the responsibility in principle, and the manufacturer's responsibility is limited to cases where there are clear defects in the vehicle system. Accidents caused by system hacking are subject to the government relief system.