news

"We are determined not to make driverless taxis. We believe that 'freeing up energy and reducing accidents' is the real social value created by intelligent driving." Recently,NIOCEO Li Bin's remarks at a media communication meeting have gone viral again. He believes that "the value of autonomous driving is not to eliminate the jobs of private car drivers and taxi drivers. This is really not exciting at all."

Li Bin's remarks touched some people's hearts. Not long ago, news about Wuhan taxi drivers losing orders due to "LuoBoKuaiPao" appeared on social media platforms. Some Wuhan taxi drivers said that their business was taken away by LuoBoKuaiPao, and they could hardly pick up passengers in places with a large number of taxis.

A single stone stirs up a thousand waves. Some people cheer for the new technological innovation, while others are sad that drivers may face unemployment. Will autonomous driving take away the jobs of human drivers? On this topic, everyone has different positions and opinions.

What I want to discuss is why Robotaxi has become popular again at this point in time after years of silence? Is there a better way or rhythm for the popularization of autonomous driving technology?

Robotaxi, or unmanned taxi, is an important application of autonomous driving technology in taxi scenarios. Unlike passenger car scenarios, taxi scenarios are inherently "separated from people and vehicles", which prompts Robotaxi to directly explore L4 level highly autonomous driving.

According to official information from Baidu, "LuoBoKuaiPao" has opened manned test operation services in 11 cities, and has carried out fully unmanned autonomous driving travel service tests in Beijing, Wuhan, Chongqing, Shenzhen, Shanghai and other places.

Looking back at the history of industry development, Robotaxi has experienced a wave of "expectation inflation" followed by a "return to rationality."

Since Google launched its autonomous driving research project in 2009, the world’s interest in autonomous driving has increased. Around 2015, Robotaxi startups emerged in large numbers, and the capital market became more enthusiastic about autonomous driving investment and financing. Cruise, Baidu’s L4 autonomous driving division, Pony.ai, AutoX, WeRide, etc. were all established during this period.

But withFord, Argo AI, which was invested by Volkswagen, went bankrupt, and Ibeo, the pioneer in the field of lidar, filed for bankruptcy. The number of autonomous driving startups has dropped significantly since 2019. The main reasons behind this are the unclear commercialization path and the tightening of funds, and the industry has entered a period of pain.

Why is Robotaxi popular again? There are several major factors behind it.

First, consumers have a deeper understanding of Robotaxi. In the past two years, there was still a safety officer sitting in the driver's seat of Robotaxi, but now it is truly unmanned. Citizens only need to swipe the last four digits of their mobile phone number on the car window with their fingers, and the door will open. Then they can fasten their seat belts and click OK on the screen to start the journey.

It is reported that the Robotaxi put into use in Wuhan completed an average of about 20 orders per day within one month, which is about the same as the number of ordinary online car-hailing drivers, but the average price per order is 60% of that of ordinary online car-hailing drivers. Although the traffic accidents that everyone is worried about have occurred, they were all caused by humans violating traffic regulations. According to the statistics of insurance companies, the accident rate of the unmanned car of the Carrot Express is 1/14 of that of humans. Therefore, compared with the experience of taxis refusing to pick up passengers and drivers not turning on the air conditioner in summer, citizens have become more enthusiastic about Robotaxi.

Secondly, the cost of a Robotaxi vehicle has dropped significantly compared to the past, which has raised the market's expectations for its commercialization. At the Apollo Day 2024 press conference in May this year, Baidu launched the sixth-generation unmanned vehicle "Robot Run" equipped with Baidu's sixth-generation intelligent system solution. The new car is priced at only 204,600 yuan, and the cost of the whole vehicle has dropped by 60% compared to the fifth-generation car, which is much lower than the Robotaxi launched by other autonomous driving companies.

In addition, by reducing platform fees and driver costs while providing 24-hour service, autonomous driving platforms can significantly reduce the unit cost of services, making the unit price per mile lower than that of traditional taxis.

The last point is actually the most important——TeslaIt said that Robotaxi will be launched this year. Recently, Tesla FSD has made significant progress. On March 31, Tesla pushed the new version of FSD V12 to some users in the United States. The new version removed the word "Beta" for the first time and replaced it with "Supervised", which means "supervised".

Musk called it "Airbnb on wheels", and car owners can choose to add their vehicles to Tesla's online car-hailing fleet and share revenue with Tesla. Tesla is a leader in smart driving technology, and if it can release Robotaxi as scheduled, it will be tantamount to an industry earthquake.

Right now, intelligent driving technology is undergoing a milestone transformation, which is called "end-to-end". The new version of Tesla FSD uses end-to-end technology, and Carrot Run's intelligent driving also uses an end-to-end big model.

What is "end-to-end"? It is a concept in deep learning. An AI model only needs to input raw data to output the final result, without introducing additional modules for processing. For example, ChatGPT, you talk to it and it answers you, and only a large language model is involved in the middle.

But intelligent driving is much more complicated. In addition to recognizing human commands, it is also necessary to recognize objects on the road and observe their texture, color, movement speed, etc. In the past, these links required different modules, and the degree of intelligence achieved by these modules varied.

For example, the module responsible for route planning and the module responsible for vehicle control are not intelligent enough, so they need to manually input a large amount of road conditions and rules and act in accordance with the rules.

As intelligent driving evolves towards end-to-end, it can judge the road conditions and react on its own. This is why many domestic new energy vehicle companies have laid off a large number of programmers in the intelligent driving department in the past six months. Because in the end-to-end era, there is no need for so many programmers to write code bit by bit and formulate every driving rule bit by bit.

Simply put, before driverless cars replace taxi drivers, the first people to be replaced are actually a large number of programmers who write code.

Musk said, "FSD's realization of autonomous driving will be Tesla's 'ChatGPT moment'". This also accelerated the development and application of large end-to-end models of autonomous driving in China.Xiaopeng, NIO, Li Auto and other new energy vehicle companies are all conducting research and development and deployment of end-to-end model technology. Industry chain companies such as Huawei, Pony.ai, NVIDIA, Yuanrong Qixing, SenseTime, and QINGZHOU Zhihang are also making efforts in the end-to-end field.

Figure: Xiaopeng R&D deploys end-to-end large model technology

Before autonomous driving is realized, all companies are working towards advanced intelligent driving with better experience. Different from traditional intelligent driving, end-to-end technology system driving is more like an experienced human driver and is not too rigid and dogmatic.

"The core advantage of end-to-end technology is that it can significantly improve the implementation efficiency of high-level intelligent driving while reducing costs to a certain extent. The success of this technical path is due to the combination of large models, large computing power and big data, which has promoted a qualitative leap in artificial intelligence algorithms." Liu Xuan, vice president of Shenzhen Yuanrong Qixing Technology Co., Ltd., pointed out in a communication with Auto Market Story.

In terms of reducing costs, end-to-end technology does not rely on high-precision maps, which is particularly important for achieving autonomous driving across the country and even the world. Liu Xuan explained that human drivers do not form a high-precision map in their minds when driving, but rely on intuitive perception of roads and road signs. As an intermediate result, although high-precision maps are helpful for machines to understand the environment, they also bring a series of problems such as maintenance and updates, timeliness, and data security. Therefore, reducing dependence on high-precision maps is the key to realizing end-to-end algorithms.

On the other hand, Tesla's FSD performance breakthrough has also led to discussions in the industry about whether LiDAR is really necessary. If Tesla can achieve unmanned driving with just cameras, will this be a blow to the LiDAR industry?

In this regard, Yu Qian, co-founder and CEO of Qingzhou Zhihang, told Auto Market Story that "the end-to-end approach is applicable to both pure vision and combined with lidar. However, although lidar provides the necessary safety redundancy under current technological limitations, in the long run, visual technology is fully capable of achieving high-level autonomous driving. With the advancement of technology, pure vision systems are expected to reach or even exceed the level of human drivers."

Li Bin also expressed his own opinion that LiDAR is like a car airbag, and its greater role lies in safety value, which can provide a backup for the 1% of extreme scenarios that pure vision cannot handle. "Whether a company uses LiDAR is a business issue, which depends on whether it pays the cost for that 1%."

Although new technology iterations and cost reductions have raised expectations for autonomous driving, Tesla's development experience shows that end-to-end intelligent driving technology has very high requirements for computing power, algorithms, and big data, and it is difficult to achieve without sufficient economic and technical strength. Tesla also finally achieved FSD end-to-end autonomous driving after a lot of testing.

Musk said that Tesla plans to invest $10 billion in artificial intelligence infrastructure and computing power in 2024. If other companies' investment intensity fails to reach this level or their capital utilization is inefficient, they will not be able to compete.

Liu Xuan admitted that Tesla's advantage lies in the fact that it has more than 6 million vehicles and a huge amount of real data to form a data closed loop. However, the specific implementation path models of different end-to-end models are not the same, and the implementation process methods are also different. "We can train our own models in a targeted manner. The CBD and congestion scenarios in China's complex urban areas have accumulated considerable experience for us. I believe these can help the end-to-end algorithm model to quickly iterate to a competitive level."

Pony.ai CTO Lou Tiancheng said that at this stage, the amount of data for autonomous driving is no longer a problem for training an end-to-end model of general performance. However, to train a high-performance end-to-end model, the quality requirements for data may increase by several orders of magnitude. This is a challenge that the autonomous driving industry will face.

As the AI ​​modeling level of autonomous driving systems becomes higher and higher, the demand for training computing resources is increasing.

Tesla has continuously increased its investment in training computing power in recent years. Tesla revealed in its Q2 2024 financial report that its AI training computing power has reached 35,000 H100 GPU equivalent computing power, and is expected to increase to about 90,000 H100 GPU equivalent computing power by the end of the year. Compared with the end of 2023, the year-on-year growth is about 500%, reaching the same echelon as Google and Amazon. Previously, Tesla also deployed a larger A100 GPU training cluster, and its actual training computing power investment is far ahead in the autonomous driving industry. Although domestic OEMs and autonomous driving companies are also investing in training computing power, few can reach the scale of Tesla.

However, there is no clear timetable for when Tesla's Robotaxi product will be put into operation. Tesla's driverless taxi Robotaxi, which was originally scheduled to be released on August 8, will be postponed to October 10. As for the reason for the postponement, Musk said that some design changes will be made to the vehicle. Some industry experts believe that Tesla postponed the release of Robotaxi mainly because its technology is not yet mature, and some bugs may appear if it is released on schedule.

In addition to technical issues that need to be improved, the Robotaxi business model is even more difficult.

Take Robotaxi for example, with 500 driverless taxis, even if each car can take 20 orders a day and earn 10 yuan for each order, the daily income of Robotaxi is only 100,000 yuan. This number is a lot for ordinary people, but for a technology company engaged in autonomous driving, it is estimated that it cannot even cover the salary of a few algorithm scientists.

Li Bin also gave his own reasons for not making Robotaxi. "Transportation is a social issue, not a separate technical issue. Technically, Robotaxi will become more and more mature, but it may not be a sustainable business. The number of cars that a city can accommodate is limited, and Robotaxi will not expand without boundaries like cloud services."

The capital market is cautiously optimistic about Tesla's Robotaxi plan. Morgan Stanley said that autonomous driving technology is affected by a series of unpredictable factors such as laws, regulations, and ethics, and its market penetration may change in an extreme curve, and the industry is still a long way from the turning point.

Although Robotaxi is still a long way from large-scale application, the end-to-end intelligent driving solution can significantly improve the user experience and make more consumers willing to accept autonomous driving, which may accelerate the overall penetration rate of high-level intelligent driving. Industry experts predict that the domestic high-speed NOA penetration rate will exceed 30% by 2026, and the urban NOA penetration rate will exceed 10%. 2024 will be an important time window for the transformation of autonomous driving from awareness to purchase.