Smart Driving Showdown 2024: Lidar may be the biggest winner

The obsolescence of lidar has once become a focus of debate in the autonomous driving industry.

The representative of the abolitionists is Musk. He believes that the purely visual solution of lidar, that is, algorithm + camera, should be cancelled. Musk believes that as long as the algorithm is strong enough, two-dimensional data will be enough. He even publicly said: "Use LiDAR people are all idiots.”

Inspired by Musk, the followers of pure vision are Jiyue Auto, which was born out of Baidu.

As a product of the powerful alliance between Internet giant Baidu and independent brand car company Geely, it combines the intelligent genes from the Internet with the top mechanical qualities of traditional car manufacturers.

According to the promotional caliber of the JiYue press conference, JiYue 01 is the world's first car company to put a large model on the car, the world's first car company to rely on pure visual perception algorithms, and the world's first car company to take over the entire process of running through China's complex urban roads. , and is the first car company in the world to support voice control/parking outside the car. Strength cannot be underestimated. One of the representatives of the pro-support faction is Yu Chengdong of Huawei . Not long ago, he shouted to Musk, "Tesla's FSD solution is very good," but "Huawei's solution is better and safer. We are equipped with 3 lidars. Compared with Tesla, Huawei ADS is far ahead. It can be used with or without high-definition maps.”

Compared with the abolitionist camp, the pro-abolition camp is stronger. All domestic models with urban NOA functions are equipped with lidar. In addition to Xpeng Motors, which focuses on smart driving, the recently launched Zhiji LS6 even comes with lidar as standard. and city NOA. Nowadays, high-end smart driving has become one of the main factors affecting consumers’ purchase intention for pure electric models priced at 200,000 yuan.

This means that whether the leading car companies want to stay ahead or the lower-end car companies want to catch up, lidar has almost become a must-have in this smart high-dimensional competition that has already begun. Not long ago, with the competent authorities passing the " Notice on the Pilot Work for the Access and Road Access of Intelligent Connected Vehicles" and the news that Tesla FSD is about to enter China, the high-end driving battlefield in 2024 suddenly filled with smoke.

It’s unclear who will win or lose in next year’s battlefield, but what we can be sure of is that the once-controversial lidar is about to have its moment of glory.

A terrible question:

Does lidar have a future?

At the beginning of 2020, the domestically produced Tesla Model 3 was officially delivered, stirring up China's new energy vehicle market like a catfish . Only then did BYD and new forces compete for success, and new energy accelerated the replacement of traditional fuel.

The first half of electrification is already a clear sign and is the general trend. The introduction of Tesla FSD will mean the official start of the second half of intelligence. The second half of intelligence begins in 2022, which is the first year that lidar is put into mass production. If the key to success in the electrification stage is scale, the goal is to replace fuel vehicles, and car companies compete for endurance and sales, then intelligence The key to winning at this stage is to differentiate electric vehicles. The way to create them is the software ecosystem, and the ticket to the software ecosystem is autonomous driving.

From an industrial perspective, the autonomous driving industry has always been focused on two hands, one is software algorithms and the other is intelligent hardware. The entire industry history is a history of the continuous evolution and integration of the two. From ultrasonic radar, millimeter wave radar, lidar to pure visual cameras.

The evolution and iteration of these hardware make ADAS, Tesla's pure vision end-to-end solution and high-level assisted driving possible. There are two reasons why lidar is popular: first, due to the characteristics of laser, it has more advantages than cameras and millimeter-wave radar in detecting obstacles; second, it requires low computing power.

Just when car companies were deploying lidar, two major events poured a lot of cold water on it. The first was the intensifying price war for smart electric vehicles in China. Since the cash flow and profits of car companies have been severely squeezed, lidar, which is expensive per unit, has become the first target for cost reduction and efficiency improvement.

Another big thing is that Tesla has provided another autonomous driving solution, a pure vision solution that relies on cheap cameras as the main sensor, and has started the process of commercial use of FSD in the United States. In fact, the first big thing is not a fundamental issue. Taking Hesai as an example, they not only reduced the price of lidar from US$80,000 to less than US$1,000, but also provided various reference designs and reference algorithms for car companies. The volume of lidar The production threshold continues to lower.

The key is the second major thing, which can be described as a "fatal" blow to lidar. For China’s electric vehicle industry, Tesla is really a powerful catfish. In August this year, Musk used a personal live broadcast to

Tesla FSD Beta V12 was cleverly released. From the perspective of artificial intelligence, compared with the previous version, FSD Beta V12 achieved "neural network swallows everything". The biggest difference is that the previous version still has 30 Thousands of lines of code are handwritten by programmers, and FSD Beta V12 is completely a neural network, which is a set of end-to-end autonomous driving technology.

If you don’t understand the terms “neural network” and “end-to-end”, it doesn’t matter. You just need to know that these are the underlying logic of large models such as Open AI’s ChatGPT. The user enters a question, and ChatGPT gives the answer directly. It will not show you the thinking process, but in any case, ChatGPT's answers behave very much like a human being.

The same goes for end-to-end autonomous driving. After inputting driving data into a unified neural network, control signals for the vehicle will be directly output. Since there is no "middleman" card information, end-to-end autonomous driving has a higher theoretical upper limit than the multi-module algorithm equipped with lidar, and it is easier to obtain the global optimal solution. In summary, Tesla has more powerful algorithms and lower costs due to its accumulation of AI technology.

This incident once again triggered the "Tesla Panic", and people kept marveling: Tesla has once again completed the dimensionality reduction attack on China's autonomous driving industry. Does lidar have a future? became a terrible question.

lidar

The biggest winner after compromise

However, as time goes by, the fatal drawback of the pure vision solution has gradually been discovered by the industry, that is, the total cost of the solution is higher than that of lidar.

The plan requires rebuilding a huge system. In imagination, pure visual solutions have ready-made AI algorithms that can be imitated, but in the actual mass production process, there are countless details that need to be improved. In imagination, as long as the algorithm is logically perfect, it will be enough. But in fact the algorithm requires large-scale data feeding.

You must know that Tesla FSD is the result of tilting countless resources. For example, during the development process of FSD, Tesla accumulated more than 9 billion miles of use, which is the world's largest source of autonomous driving data; in order to use this data , Tesla continues to expand its supercomputing cluster, digs top AI engineers everywhere, and develops self-developed algorithms, chips and high-power GPUs. After this calculation, the total cost is higher than the lidar solution.

In other words, unless you have a deep accumulation of AI technology, you will just copy the homework of the top students. The top students will become more and more powerful, and you will become more and more confused.

Finally, after several struggles, at the end of 2023, domestic car companies and their smart driving solution providers finally reached a consensus. We can summarize this as the geometric intersection of impossible triangles.

The impossible triangle is autonomous driving, cost and long-term voice. If autonomous driving has rich functions and can dominate the long-term discourse, the cost of the vehicle will inevitably be high, and it may not be able to survive in the short term. If the autonomous driving function is abandoned, the cost will be reduced in the short term, but in the long term, the right to speak will inevitably be lost.

In this regard, Horizon CEO Yu Kai said: "Currently, there is no need or ability to achieve true driverless driving in the short term. The industry is returning to its commercial nature. The successive implementation of L2+ advanced assisted driving represented by high-speed NOA is creating opportunities for users. value.

The real product goal at the moment is to make the autonomous driving experience on closed roads such as high-speed NOA and loop NOA smooth as silk at a reasonable cost performance. At the same time, we will actively invest money and time to truly achieve NOA in urban areas. Available and truly creating value for users. "

To sum up, the consensus in the industry is that the intermediate solution between autonomous driving and long-term voice is high-speed NOA and urban NOA. No one is betting on a certain technical solution alone anymore, but iterates and optimizes the technology on the premise of satisfying the user experience. The biggest winner of this compromise is lidar.

The urban NOA mentioned by Yu Kai means that car owners can use autonomous driving functions on fixed routes in cities such as highways. Early urban NOA functions generally used high-precision maps to provide more accurate positioning services. However, high-precision map collection costs are high, coverage is low, and updates are slow, making it difficult to meet the rapid and large-scale needs of urban NOA. In the current urban NOA competition, vehicles must have accurate three-dimensional detection capabilities.