(Report Producer/Author: Huajin Securities, Li Hui, Dai Zhengzheng)
The development of Tesla's FSD - simplifying
FSD is the most complete product in Tesla's assisted driving solution
Tesla's autonomous driving solutions include Basic Autopilot (AP), Enhanced Autopilot (EAP), and Full Self-Driving (FSD). Among them, FSD stands for Full Self-Driving, which is the most complete product in Tesla's Autopilot portfolio.
In terms of function, in addition to the basic active cruise control and lane maintenance centering, Tesla FSD can also achieve: 1) automatic assisted navigation driving, including automatically driving into and out of highway ramps or overpass forks, overtaking slow-moving vehicles; 2) automatic lane change, including automatic lane change on highways; 3) Automatic parking, including parallel parking and vertical parking; 4) Intelligent summoning is in the appropriate scenario, the vehicle parked in the parking space will respond to the call, drive out of the parking space and go to the owner's location; 5) Traffic light and sign recognition; 6) Automatic auxiliary steering in urban areas, detecting lanes, vehicles and obstacles, and operating vehicles for steering; 7) Automatic speed offset adjustment, which can adjust the driving speed of the vehicle independently according to different environments and scenarios.
Tesla, as a leader in pure vision solutions, has cameras at the core of FSD's intelligent driving
Different from the multi-sensor fusion solutions of most domestic manufacturers, Tesla's FSD autonomous driving is a pure vision solution with cameras as the core. The original design of the pure vision scheme was inspired by the study of human vision; That is, after the information collected by the human eye reaches the retina, it passes through multiple areas of the cerebral cortex and the nerve layer, and finally forms biological vision and generates images in the brain. Tesla's goal is to design the car's visual cortex through algorithms, software and hardware, and build a vision-based computer neural network system like the human brain. First, while the Tesla car is driving, the vehicle collects environmental image information through the camera. Tesla's HW 2.0/2.5/3.0 hardware is equipped with 8 cameras that monitor different directions, including three front cameras (including 1 main camera, 1 wide-angle camera, and 1 narrow-view telephoto camera), 2 front cameras, 2 rear cameras, and 1 rear camera.
Iteration on the algorithm side
The algorithm modules of autonomous driving are usually divided into "perception", "decision planning", and "motion control". Among them, the "perception" module is the core of autonomous driving, and most of the technical upgrades are concentrated in the perception module, the purpose of which is to make the vehicle's "perception" of the driving environment reach the level of human perception; The "decision planning" is based on the output of the "perception" module, through the planning of car behavior and driving path, so that the car can reach the specified destination, and as far as possible to ensure driving safety, efficiency and comfort.
At the perception level, Tesla has undergone a transformation from "Feature Extraction Network RegNet" to "BEV+Transfomer" and then to "BEV+Transfomer+Occupancy Network", and the decision-making and planning level has gradually tilted from "Rule-based" to "Machine learning-based" since 2021. Until January 2024, Tesla has become the world's first company to mass-produce "end-to-end" neural networks by launching FSD V12 Beta, realizing the integration of perception, decision-making, and planning.
Algorithm 4.0: The application of Occupancy Network reduces the computational complexity, and the introduction of time series information pushes image recognition to 4D
Decision-making level: In 2021, Tesla began to add elements of neural networks to the path planning level, and launched the "Monte Carlo Tree Search Algorithm" to output decisions through path selection probability and situation evaluation. However, only a small number of neural networks are used at this stage, and most of them are still artificial rule codes. In 2022, the newly launched "interactive search network" combines the Monte Carlo algorithm into the Occupancy network, and each trajectory calculated will have a cost function to optimize the tree search to give more candidate targets, which depends on the four factors of collision probability, comfort, intervention possibility and human operation similarity. The interactive search network successfully reduced the computation time from 1 to 5 milliseconds to 100 microseconds. But the function part is still rule-based code.
Software: In 2021, Musk announced that the FSD Beta number will start with V8.1. From the perspective of major version upgrades, the rhythm of annual updates is basically maintained; In terms of update frequency, there is a more obvious acceleration of iteration, from one update every two months at the beginning of the period, to two to three updates per month, or even four updates per month. With the expansion of the test scope and the increase of driving data, the V9/V10/V11 versions mainly focus on the optimization of functions and continuously deepen the anthropomorphism of intelligent driving.
HW3.0→HW4.0: 50% more cameras and mmWave radar returns in a high-precision 4D version
The reasons for the previous abandonment of millimeter-wave radar are: (1) on the one hand, the low resolution of traditional millimeter-wave radar causes the degradation of fusion sensing performance; Tesla's director of artificial intelligence has said that for low-resolution radar, when passing through a scene like an overpass, it is difficult to distinguish between the overpass and the parked vehicles below due to the low elevation resolution of the radar, which can easily lead to a collision. (2) on the other hand, the number of millimeter-wave radar channels limits the improvement of its perception ability. In contrast, cameras are capable of generating large amounts of data, and software improvements can make the most of this data.
As for the return of millimeter-wave radar, the main reasons are: (1) the resolution of high-precision 4D millimeter-wave radar has been greatly improved; Since 2021, radar chip solution providers such as NXP and TI, as well as radar system suppliers such as Continental ZF and Bosch, have been accelerating the mass production of 4D imaging millimeter-wave radar. The new 4D millimeter-wave radar has greatly improved resolution performance, with point cloud output (better fusion with vision or lidar, and possible classification and recognition capabilities) and all-weather performance, making it one of the options for high-end solutions. (2) be able to compensate for the risks of purely visual solutions; Tesla's "ghost brake" problem (braking without warning) criticized by the public is mainly caused by the sensor's perception defects, due to the poor performance of the camera at night or in bad weather conditions such as fog or rain; At the same time, the camera reaction time is usually longer than that of millimeter-wave radar, and it often takes a few frames to identify changes in the velocity of an object. On the other hand, millimeter-wave radar measures distance, relative speed, and direction based on the difference between the transmitting frequency and the receiving frequency, and can be used at night, in backlight, fog, rain, and snow.
Parsing Tesla FSD
Benefiting from the leading position in terms of data volume, computing power, and hardware adaptability, Tesla's FSD has strong technical competitiveness
(1) The amount of data: The effect of the AI model depends on the quantity and quality of the input data, and the more excellent driving data is entered, the more suitable and excellent driving decisions the AI model can make. FSD has been tested in North America since October 2020, and with the increase in FSD push regions and users, Tesla's driving data will increase exponentially; In terms of data volume, it is difficult for domestic manufacturers to catch up. In April 2024, Tesla announced that its Full Self-Driving (FSD) technology-powered cars have driven more than 1 billion miles, equivalent to 1.61 billion kilometers, while no domestic manufacturers have reached this mileage.
2) Huge computing power center: AI models are trained on data and are built on computing power platforms, and their computing power and computing power investment are key indicators. 1) In terms of computing power, Xpeng is based on the "Fuyao" intelligent computing platform built by Alibaba Cloud, with a computing power of up to 600PFLOPS (6 billion floating point operations per second), and in 2022, the computing power of Tesla's computing center has reached 2 EFLOPS (20 billion floating point operations per second). In the future, after Tesla's self-developed super computing platform Dojo is put into use, its computing power will rise to a new level; According to Tesla's computing power development plan released in June 2023, Dojo will reach 100 EFlops computing power by October 2024. 2) In terms of computing power investment, at Tesla's AI Day 2022, Musk said that Tesla currently has a super computing center with more than 14,000 GPUs; In August 2023, Tesla launched a new training cluster of 10,000 H100 GPUs, which are five times faster than the previous generation A100 and are expensive at nearly $40,000 each. At the same time, the later maintenance cost of the computing platform is much higher than the hardware cost; Musk has publicly stated that Tesla will spend more than $2 billion on expanding the computing power of training computing power in 2023, and said that he will take the same action in 2024; In comparison, there is still a big difference between the computing power investment of domestic manufacturers and Tesla.
(3) Self-developed hardware solutions with high adaptability: Tesla set up a chip team in February 2016 and successfully launched the FSD chip in April 2019, and it took three years to launch the HW hardware solution. The self-developed HW3.0 is the first autonomous driving hardware solution entirely from a car company, and it is also the most powerful solution for deep learning theory on mass production models, and at present, HW has evolved into the 4.0 era. The advantages of self-developed hardware are, first of all, high cost performance and high utilization, which greatly reduces the hardware cost of FSD; The second is that it has a high degree of development freedom, which can better support Tesla's innovative algorithms and other related technical solutions. In contrast, most domestic manufacturers use purchased chip solutions, and there is a certain gap between them and Tesla's self-developed hardware solutions in terms of adaptability and utilization.
Tesla's FSD progress in entering the Chinese market
Tesla's FSD entry into China may become a key issue
On April 28, 2024, at the invitation of the China Council for the Promotion of International Trade, Musk arrived in Beijing, met with relevant leaders of the Ministry of Foreign Affairs, the Ministry of Industry and Information Technology, the Ministry of Commerce, and the China Council for the Promotion of International Trade, and went to the Tesla Gigafactory in Shanghai; Speculation or a slowdown in adoption of Tesla's core EV business due to economic uncertainty has left Tesla's core EV business struggling with the hope of generating recurring, high-margin revenue by adopting its FSD software suite. Prior to this, Musk had said on Tesla's Q1 earnings call that "we plan to release it as a regulated autonomous system to any market where we can get regulatory approval, and we believe that this includes China, subject to regulatory approval." ”
So far, Tesla has launched two EAP (Enhanced Self-Driving) subscription options in China, including a monthly package of 699 yuan and a quarterly package of 1,399 yuan, and a one-time purchase of 32,000 yuan, which will help existing Tesla owners increase the use of software and services and prepare for the upcoming FSD and robo-taxi business.
On May 30, it was reported that Tesla was about to register its full self-driving software FSD in China; If Tesla successfully registers its FSD software with China's Ministry of Industry and Information Technology, Tesla employees will be able to conduct internal testing of FSD on China's public roads, with plans to upgrade it to Chinese users in the coming months.
Or due to the lack of redundant design, the FSD autonomous driving rating stays at the L2 level
It can be seen that at present, the domestic car companies that have passed the L3 test basically adopt the redundant design of multiple sensors, that is, the system has a backup system, when the main system fails, the backup system can be started and continue to run, so as to ensure that the vehicle can continue to run when it fails, and assist the vehicle to drive to a safe area, but also make up for the semantic uncertainty of a single sensor; Car companies such as Wenjie, Mercedes-Benz, and BMW mainly focus on lidar design, while companies like Changan Automobile and Deep Blue tend to have high-precision map solutions in the absence of lidar.
Without the use of lidar and high-precision maps, Tesla's HW 3.0 version of the autonomous driving function may only reach the L2 level, and is only used as an autonomous driving assistance function. However, it is worth noting that Tesla is considering better solutions to improve the safety and reliability of autonomous driving: 1) Millimeter-wave radar has returned as a high-precision 4D millimeter-wave radar in FSDHW4.0 (reserved interface, not yet mass-produced), or to a certain extent, it can replace lidar to make up for the risks of pure vision solutions and ensure the normal use of autonomous driving at night or in severe weather conditions such as fog or rain. 2) At the same time, in recent years, Tesla has successively purchased Luminar lidar for reliability testing; Although it is not known whether LiDAR will return for the time being, it can be seen that Musk's attitude towards LiDAR has changed.
Excerpts from the report:
(This article is for informational purposes only and does not represent any investment advice from us.) To use the information, please refer to the original report. ）
Selected report source: [Future Think Tank]. Future Think Tank - Official Website