Can Tesla's FSD Really Handle China's 300 Million Electric Scooters?

Recently, Li Liyun, the head of autonomous driving at XPeng Motors, commented on social media about the topic of "FSD struggling to deal with 300 million electric scooters in China." He believes that "compared to the United States, road conditions in China are much more complex," and used the road conditions in the central urban area of Guangzhou as an example to predict that "interactions with electric scooters and other non-motor vehicles are indeed one of the headaches that Tesla FSD will face when entering China," and introduced some thoughts of XPeng XNGP for such scenarios. XPeng Motors Chairman and CEO He Xiaopeng immediately forwarded and commented, "From the difficulty of road solutions, India > China > USA, but AI will accelerate these challenges to be leveled."

The Unprecedented Heat of the Intelligent Driving Market

XPeng Motors has been frequently interacting with Tesla on social media recently. On June 20, He Xiaopeng posted on social media that he was in the United States, preparing to test the latest version (12.4.1) of FSD and Waymo; then on June 25, after experiencing FSD and Waymo, he wrote, "We will also learn from FSD their excellent features and user experience, I believe that 2025 will be the moment of fully autonomous driving of Chat GPT!" On July 1, he posted again, "Waymo is more smooth, but I think, FSD has made rapid progress in the past six months, and as the amount of data of FSD is getting bigger and bigger, the whole Infra is getting better and better, I firmly believe that by 2025, FSD will definitely surpass Waymo."

In fact, it's not just XPeng Motors, but also domestic new forces and brands such as AITO and Zhiji, which are supported by Huawei, that have taken intelligent driving capabilities as this year's direction of effort. We have observed many brands announcing the progress of urban NOA, and new cities are constantly joining the "open city list." Artificial intelligence and large models are undoubtedly the hottest keywords in the field of intelligent driving at present. XPeng released the first mass-produced end-to-end large model on the 520 AI DAY, and AITO, Great Wall, NIO, Li Auto and other companies are also accelerating the use of artificial intelligence to enhance the performance of intelligent driving.

Evolution of FSD

As the industry benchmark, the industry has been closely tracking the performance of Tesla's FSD. FSD has led the industry's exploration of autonomous driving in the direction of neural networks, Transformer+BEV, etc., but because Tesla insists on pure visual perception + no map scheme, its actual effectiveness has also been widely controversial in the industry.

However, on August 26, 2023, Musk's live broadcast of the FSD V12 test changed some of the industry's views. Behind this live broadcast is that FSD has integrated the perception, planning, and control links, changed the original rules, directly input the perception results into the neural network for processing, and output the driving instructions by the neural network.

"The industry is very concerned about Tesla's new generation of FSD (V12). The new FSD released by Tesla is a single-vehicle intelligence + end-to-end model based on a large model. With such a new set of schemes, its level of intelligence has made significant progress, so we believe that the new generation of Tesla FSD is an inspiration and motivation for the global intelligent automotive industry." Academician of the Chinese Academy of Engineering, Professor at Tsinghua University, and Chief Scientist of the National Intelligent Connected Vehicle Innovation Center, Li Keqiang said at the recent 11th International Intelligent Connected Vehicle Technology Annual Conference.

"From a large technical route, Tesla FSD V12 can still be regarded as a single-vehicle intelligence. Through the car-mounted sensor hardware that is mainly visual, it can perceive and complete planning and decision-making in real time. On the other hand, in its autonomous driving algorithm development system, its end-to-end model parameter calibration, data loop and other non-real-time development training tasks have utilized the high-performance supercomputer capabilities of the cloud." An intelligent connected vehicle expert analyzed to the reporter.

Obviously, the significant advantage of the V12 version of FSD is that it can reduce the dependence of the self-driving scheme on hardware, such as LiDAR, high-precision maps, etc., and can obtain the global optimal solution, but its weakness is also very obvious. The model requires a lot of training, so the initial performance may not be satisfactory.

"From the current application of FSD, it seems to be acceptable." Professor Zhu Xichuan from the School of Automotive Studies at Tongji University said in an interview with reporters at CICV2024, "The FSD V12 version, with a user scale of about 1.7 million, has not reported very serious traffic accidents after three months. Some safety reports from the industry believe that FSD is safer than human drivers."

Both Sides Have Their Strengths

The evolution of FSD is seen by the domestic industry and has quickly followed.

Including many new forces, autonomous driving solution providers, etc., the industry has realized the huge changes that artificial intelligence can bring to intelligent driving, quickly changing the perception and control scheme based on rules, and large models are more and more invested in application. At the same time, because the pure visual scheme has also been "run through" to some extent by Tesla, it may also affect the domestic car companies' choice of multi-sensor fusion.

"We analyze from the perception perspective, (pure vision and multi-sensor) two schemes each have their advantages, pure vision sensor scheme system structure is simple, hardware cost is low, the overall scheme relies on software and data thresholds are higher; multi-sensor fusion scheme combines the performance advantages of different types of sensors, adaptability to complex weather and other environments is stronger, but the system in the car cost and maintenance cost are higher than the pure vision scheme." The aforementioned expert said, "We guess that under the current intense competition in the industry, many car companies may follow the pure vision scheme, especially in the middle and low-level entry-level models to popularize the application of intelligent driving systems based on pure vision; multi-sensor fusion perception architecture is more used in the middle and high-end models to enhance the experience. If such speculation is established, then the market will have more products for consumers to choose from."

Although Tesla's FSD V12 has not yet had major accidents in the United States, there are many unique problems in China that Tesla faces when entering China, and these problems may have become a "strength" for domestic companies.

The "300 million small electric donkeys" commented by Li Liyun vividly illustrates the complexity of the domestic traffic environment. There are many types of traffic participants in our country, and the traffic congestion in the central areas of many cities is relatively serious. The "Brownian motion" of pedestrians, the rush of electric bicycles, express delivery special electric tricycles, bicycles, and the rush with motor vehicles, and the behavior of violating traffic signals at intersections have greatly increased the difficulty of autonomous driving. This is the "national condition" that Tesla FSD needs to face when entering China.

"End-to-end intelligent driving solutions are driven by data. Tesla FSD technology is more mature, but it is basically still an intelligent driving system version developed and iterated based on data from the North American region, which is more applicable to the application environment of the North American region." The above expert analyzed, "But if it is applied to the domestic traffic environment, our evaluation believes that there are obvious differences in road, car culture, traffic rules and other aspects between the two countries. Tesla FSD indeed must carry out a large amount of high-quality data collection training and model calibration according to the domestic traffic environment; therefore, from this point, there are advantages and opportunities for our domestic car companies' intelligent driving. Our next step should be to accelerate the catch-up and breakthrough in end-to-end technology development."

Compared with this, many models of independent brands are equipped with LiDAR, millimeter wave radar,