The Tesla Robotaxi is called the Cybercab, and it doesn't have a steering wheel or pedals. I just arrived at the event site on the Cybercab, and there are still 20 similar cars on the road. "At the" We, Robot "launch event on October 11th, Tesla CEO Elon Musk said that Tesla will produce the Cybercab before 2027, and the cost of the car is expected to be less than $30000.
Although Tesla did not disclose specific technical details of its perception solution at this press conference, judging from the Cybercab on site, this car is different from the vast majority of L4 Robotaxis currently on the market. It not only does not have a steering wheel and pedals, but also does not have a complex sensor layout on the exterior of its body. Previously, overseas analysts have stated that Cybercab mainly relies on cameras and artificial intelligence technology for driving, and may not use LiDAR.
In addition, Musk also announced that the Model 3 and Model Y will achieve FSD functionality without human supervision in the future, both of which use pure visual perception solutions without millimeter wave radar.
In addition to differences in perception schemes, during the Q2 earnings conference call this year, Musk made it clear that Robotaxi's intelligent driving solution will not rely on high-precision maps. At present, most L4 Robotaxis in China are still based on high-precision map solutions.
In addition to Tesla, Xiaopeng Motors in China has also announced that it will launch Robotaxi products in 2026. At the same time, Xiaopeng Motors also claims to implement the L3+autonomous driving experience based on L2 intelligent driving assistance hardware. In the recently released new car, Xiaopeng Motors has cancelled the laser radar and adopted a visual perception solution.
What gives confidence to Tesla and Xiaopeng is the application of end-to-end big models in the field of autonomous driving. At present, both Tesla and Xiaopeng Motors have applied end-to-end large models in their intelligent driving assistance systems. With the support of this technology, the iteration speed of related systems has been greatly upgraded, and the user experience is also better than previous rule-based intelligent driving systems.
Xiaopeng stated that after the end-to-end large-scale model is mass-produced and put into use, Xiaopeng Motors' OTA updates will achieve "version iteration every 2 days and experience upgrade every 2 weeks"; Tesla has been exposed to have at least three versions undergoing training iterations at the same time, and once one version is approved, it will be released immediately, from FSD V12.5.2 to 12.5.3, in just two days. At the end of this month, Tesla will push the V13 version of the FSD system.
Xiaopeng Motors Chairman and CEO He Xiaopeng believes that the end-to-end model's lower limit capability is expected to improve rapidly next year. Once improved, it will take less than two years to achieve the ability to exceed L4 autonomous driving standards globally.
Although executives from car companies such as Tesla and Xiaopeng have vigorously promoted the enormous benefits of end-to-end big models for L4 autonomous driving and Robotaxi, there is currently no industry consensus on whether this technological roadmap can truly make L4 Robotaxi a reality.
Different from L2 driving assistance, automatic driving above L3 level will transfer the responsibility subject of the accident to the vehicle, which puts forward high requirements for the stability and safety of the auto drive system. The inexplicability of the black box of the end-to-end large model brings certain risks to the auto drive system. L4 Robotaxi requires 100% security and cannot accept end-to-end& quot; Black Box& quot; The inexplicable and uncertain nature it brings. Meanwhile, due to the black box nature of end-to-end large models, the training results are not controllable. When problems arise in the new version of the system, developers cannot directly modify them and can only inject new data to enable the large model to self train and produce a better solution as much as possible.
&Amp; quot; After Tesla launched its end-to-end FSD, there were some issues where the car collided with the road shoulder, especially at night. Sometimes there were scratches, and other times it directly collided with the road shoulder, causing the tires to deflate. Hou Cong, President of Qingzhou Zhihang, told First Financial reporters that Waymo, also in the United States, has not adopted an end-to-end large-scale model, but has been able to achieve unmanned Robotaxi operations in multiple cities, and user feedback has been quite good.
Some practitioners of L4 Robotaxi believe that end-to-end big models have certain value for L4 autonomous driving, but this does not mean that the current L2 system can evolve to the L4 level through end-to-end big models.
Robotaxi needs to achieve L4 level autonomous driving capability, with the ultimate goal of providing a safety net for the system. Therefore, it needs to surpass the level of human drivers and perform 10 times better than normal humans. Therefore, it is necessary to continuously strengthen various safety redundancies and enhance safety net capabilities. Relatively speaking, the challenge at L4 level is greater, and with the support of end-to-end models, there is more value from 0 to 1 Zhang Ning, Vice President of Xiaoma Zhixing and Head of Robotaxi's Autonomous Driving Travel Business, told First Financial reporters that L2 level assisted driving only needs to perform like a human, and ultimately be backed by a human driver. The greater value of end-to-end lies in cost reduction.
Apart from controversies over software and large models, L4 Robotaxi practitioners generally believe that pure visual solutions or hardware based on L2 intelligent driving assistance systems cannot achieve L4 autonomous driving.
Zhang Ning stated that for non motorized vehicle objects, pure visual solutions have inherent flaws in detection, especially in domestic urban road scenes with a high number of pedestrians and bicycles. If there is a situation where objects on the road cannot be seen, the autonomous driving mode will not slow down or avoid them in a timely manner, resulting in safety hazards; The fusion sensor solution, including laser radar, cameras, millimeter wave radar, etc., can ensure that the vehicle can see clearly, see far, and walk steadily, with higher safety and reliability.
Musk's view that autonomous driving does not require high-precision maps is also highly controversial.
Technically speaking, high-precision maps and lane level auxiliary maps, as a type of over the horizon sensor, are helpful for driving efficiency and safety; However, high-precision maps require high freshness, and the current update speed cannot meet the needs of enterprises for city opening and expanding usage areas; For car companies, high-precision maps are also an additional cost expenditure.
For L4 level Robotaxi, end-to-end large models can reduce the system's dependence on high-precision maps in the initial stage of operation, allowing the company to expand its operational scope faster; However, in the middle and later stages of operation, high-precision maps still have an important impact, which can further improve the reliability, safety and smoothness of the auto drive system.
At present, all L4 level autonomous driving solutions on the market require high-precision maps. In the advanced assisted driving market, Huawei and Tesla also rely to some extent on high-precision maps or lane line level high-precision auxiliary maps. In the analysis report on Tesla FSD, it can be seen that it also has lane line level map information; while Huawei has a map annotation team of about a thousand people, "Zhang Ning told reporters.
Although Tesla claims to launch a competitive Cybercab by 2026, Musk's previous statements on autonomous driving have had a "black history" of multiple delays. On the other hand, there are still significant differences in the industry regarding how to implement L4 Robotaxi and which technology solution to adopt. Whether L4 Robotaxi can truly be implemented in just 2-3 years remains to be observed.