A few days ago, lidar company Luminar Technologies said in a letter to shareholders that Tesla was "our largest lidar customer in the first quarter, accounting for more than 10% of our revenue in the quarter."
According to the data, Tesla purchased more than $2 million worth of lidar from Luminar Technologies in the first quarter of this year.
Image courtesy of Luminar
This move not only contradicts Tesla's current pure vision self-driving scheme, but also directly "slaps in the face" Musk, the CEO of the company that has publicly denigrated LiDAR many times.
As a "pioneer" in reducing the cost of autonomous driving, Musk has publicly dised lidar more than once: October 2015: Lidar is meaningless and unnecessary for self-driving cars; April 2017: Lidar is bad, they'll ditch it, remember my words, it's my prediction; February 2018: Lidar is expensive, ugly, and unnecessary; In April 2019, lidar was like a person with a bunch of appendices, and the existence of the appendix itself was basically meaningless, and it would be ridiculous if there were a bunch of them. Any company that relies on lidar can end up with nothing. Fools only use lidar, and now whoever (does autonomous driving) still has to rely on lidar, then it is doomed! Don't believe it, let's see!
However, the move to spend a huge amount of money on lidar in the first quarter of this year will inevitably make the outside world speculate that with the continuous upgrading of lidar technology in recent years and the low price, will Musk begin to change his mind to adopt lidar?
Especially since 2021, Musk has uncharacteristically lost his tough attitude towards lidar. Musk once appeared in the Clubhouse's audio chat room, admitting that he was "inappropriate" about lidar and said that his rocket company, Space X, has developed its own in-house lidar. Musk said at the time: "Obviously, if I hated lidar, I wouldn't do it." ”
However, Musk still insists that solving the visual problem is the first priority in terms of autonomous driving, and in the current earnings call, Musk also said that Tesla is moving away from radar.
Therefore, Tesla's spending of 2 million yuan to purchase lidar in the first quarter is still worth discussing.
LiDAR, a necessity for high-end intelligent driving?
Before 2021, the high-cost model of "LiDAR + high-precision map" was also recognized as a feasible solution for high-end intelligent driving in the industry.
However, with the introduction of BEV (Bird's Eye View) technology in Tesla's FSD beta that year to carry out the concept of environmental perception, and the integrated information of vehicle sensors through the improved Transformer model, BEV technology has quickly become a new development consensus in the industry.
Because the coordinate system of information feedback is the same, image data, millimeter-wave radar data, lidar data, and high-precision map data can be fused in the space constructed by the system.
However, the divergence between the use of lidar and the purely visual route also began to emerge from this stage.
At the technical level, if a car wants to perceive its surroundings, it needs to "see" and "recognize" its surroundings, which requires labeling traffic participants in the development stage to help cars gradually understand what cars, trucks, pedestrians, and bicycles are, which is the so-called whitelist mechanism. However, the actual road environment is very complex, and falling stones, illegally occupying and parking vehicles will cause confusion in vehicle identification, which is the reason why intelligent driving vehicles cannot recognize piles and barrels and other items in the early years.
In order to solve this problem, the practice of domestic manufacturers is to use lidar to judge the volume of unknown obstacles through laser point clouds, and then select braking or lane change bypass based on map data. However, the new problem lies in the fact that such vehicles can combine map data to achieve detours and other operations after identifying obstacles, which requires OEMs to purchase and update high-definition maps from map operators, and at the same time bear the problem of slow update of high-definition maps.
What's more, the price of a lidar is often thousands of dollars, which is unattractive to Tesla, which is extremely demanding on the cost side.
From the perspective of marketing, as Tesla is selling to the global market, it is not realistic to obtain high-definition maps of all the countries on sale, so at AI Day 2022, Tesla introduced the new OCC (Occupancy Network) occupancy network.
Functionally, OCC can output the occupancy probability of each 3D position around the vehicle to the on-board computer through the high-speed information acquisition of the camera, and mark the semantics of the objects in the picture and distinguish them by color.
To put it simply, OCC divides the real world into small squares, so as to jump out of the inherent thinking of object recognition, so as to facilitate the vehicle to make quick and accurate judgments.
Judging from the technical demonstration announced by Tesla two years ago, with the blessing of a variety of deep learning models, the Model 3 with only 144 TOPS computing power and 8 surround view cameras is enough to match the detection capability of lidar, and can even cooperate with fully automatic 4D annotation technology to complete the real-time drawing of high-precision maps of current roads on the vehicle side.
In other words, Tesla's FSD (Full Self-Driving) can be separated from radar and can meet the vehicle's needs for autonomous driving with only a pure vision solution.
For pure vision solutions, the role of lidar in BEV is more similar to a kind of safety redundancy, and whether car companies adopt it depends on their control of vehicle costs.
In fact, the most influencing factor in the lidar market is not the control of thousands of yuan of hardware by the car companies themselves, but the boundaries of the software development capabilities of the car companies.
Last year, Li Xiang, the founder of Li Auto, said at the spring media conference that to make a BEV-based urban NOA, it basically needs to start at 2 billion US dollars, not to mention the additional cost of breaking through the format barrier between camera data and radar data. This is still a lot of investment for new energy vehicle companies whose net profits are still negative.
Therefore, when most car companies are building high-end intelligent driving, it is obviously a good choice to use lidar with light map routes to achieve urban NOA, but in terms of hardware cost, it is destined to be more expensive than the same level of pure visual solutions.
The reason why Musk was able to give up lidar is precisely because of his early profitability in the automotive business and the result of continuous R&D investment in AI and intelligent driving technology.
Today, there are only two car companies in the world that can run through pure vision technology based on AI large models and OCC occupancy network technology, Tesla and Jiyue Automobile, which is backed by Baidu Apollo.
Image source: Tech Planet
In fact, compared with most of the models equipped with high-end assisted driving technology in the market, Tesla, which has a lower vehicle price, has a price of 64,000 yuan on FSD that far exceeds that of competing products, which undoubtedly shows Tesla's desire to put money into its own pocket through software technology research and development.
Pure vision still buys lidar?
Since he is determined to make money on the road of pure vision, Musk still spends 2 million yuan to purchase lidar, which is bound to have other plans.
In fact, as early as May 2021, foreign media revealed that Tesla signed a contract with LiDAR company Luminar for testing and development using laser sensor technology. As part of the deal, Luminar sold the sensors used in the Model Y to Tesla, people familiar with the matter said.
However, some analysts at the time believed that Tesla's engineers may have benchmarked their camera-centric self-driving system with lidar.
Earlier Guidehouse Insights analyst Sam Abuelsamid said that any speculation that Musk will change his path and adopt lidar is unrealistic, "If they make such a change, it will actually scrap their entire fleet, and they will not modify 1 million cars for it." ”
Two years later, Luminar's CFO Tom Fennimore's statement on the earnings call also seemed to confirm Musk's speculation that he hadn't changed his technology course: "It's not the first time they've ordered lidar from us, but I want to say it's not recurring, it's intermittent." ”
Obviously, in the past two years, from the beta version of FSD to the official landing now, although Tesla has purchased lidar for a long time, it has not changed its insistence on FSD.
Since 2024, in the case of declining electric vehicle sales and increasingly fierce competition from China, Musk has just bet Tesla's future on the Robotaxi to be launched in August, and he is also actively visiting China to promote the implementation of FSD in China.
At present, Tesla has begun to push the FSD V12 version of the beta test name to North American users on a large scale OTA, with the imminent launch of its Robotaxi and the imminent landing of FSD in China, Tesla's large-scale purchase of LiDAR may only be the final sprint for the verification of visual recognition accuracy before FSD landing.