Li Gen from the Vice-Pilot Temple
Smart Car Reference | Official account AI4Auto
"Do you still have contact with your L4 friends?"
Zhou Guang, CEO of Yuanrong Qixing, did not answer positively.
Three years ago in Futian, Shenzhen, Zhou Guang led the team to build the Yuanrong Qixing Robotaxi test car, took me through the downtown evening rush hour, and later sent me all the way to Shenzhen Airport, the whole process of automatic driving, even in the temporary parking traffic of the Shenzhen Convention and Exhibition Center commercial building, Yuanrong Robotaxi also saw the stitches and needles out - no downtime, no remote assistance, no takeover.
As the CEO of Yuanrong Qixing, all the topics that Zhou Guang talked about three years ago have their own contextual limitations: L4 autonomous driving. Yuanrong Qixing's external introduction is also a L4 autonomous driving company.
In the autonomous driving industry terminology, L4 is a watershed, different from L3, L4 completely gets rid of human driver dependence; Unlike L5, L4 can only be used in limited scenarios and ranges, the so-called ODD (Operational design domain) area, so that L5 can truly achieve complete unmanned driving.
But three years later, in 2024, Zhou Guang corrected and repeatedly emphasized in the image of a betrayer and evangelist:
L4 is a pseudo-concept, or L4 doesn't exist anymore.
Because the new end-to-end AI paradigm is reshaping all industries, including autonomous driving in transportation scenarios.
End-to-end, no graph, AI 2.0
Zhou Guang's emphasis on end-to-end refers more specifically to the surging paradigm of AI algorithms.
To put it simply, it is that in the past, autonomous driving may be divided into perception module, control module, positioning module, and planning and decision-making module according to different functions...... However, from the 2023 CVPR Best Paper Award to the end-to-end award, a new paradigm has emerged to a new height.
Under the end-to-end paradigm, the entire autonomous driving system has only one module, or only one neural network - you can also understand it as the AI brain, perceiving the data signal output, and finally making the data signal output, just like a human driving a car.
In the field of autonomous driving, Zhou Guang is now an end-to-end evangelist and absolute believer.
He said that the development of modules and stacks for autonomous driving involved the transfer of parameters between modules, and artificially defining interfaces would inevitably filter some information, which ultimately led to limited processing scenarios.
At the same time, because of the "non-derivative" between each module, each module can only achieve the local optimal on the module, but cannot achieve the overall global optimum.
To put it more bluntly, once in the autonomous driving scenario, the algorithm still relies on human rules, the machine will always encounter a long-tail scenario that cannot be eliminated. Autonomous vehicles are hardly "human-like", that is, they can handle scenario challenges like humans on the fly.
The scene experience that Zhou Guang took as an example is lane change - four-lane lane lane change, human driving is basically a line to the end to complete the lane change, but the machine driven by the rule algorithm will only change lanes like a carp, lane by road.
The difference in product experience is also very obvious in the technology behind it.
Zhou Guang explained that the end-to-end advantage is that when you make some decisions that require a high level, it may be very perfect, like traditional autonomous driving, you may only need to look at 1 or 2 seconds, and your decision may be based on these 1 or 2 seconds, and the decision will be inaccurate if it is a little longer. But end-to-end, you can make an accurate judgment of about 10 seconds, and 10 seconds is much worse than 1 or 2 seconds.
Zhou Guang believes that the AI 1.0 era of autonomous driving and even the entire robot industry is over.
What era is the 1.0 era?
In the era when people are programmed to write rules, neural networks are taught to drive, and then Coner Case is constantly defined and processed, even if the pursuit of completely unmanned autonomous driving requires an era of ODD, constraints, and areas at a certain stage, that is, the era of L4.
What is the era of AI 2.0?
The entire AI driver's brain is a large neural network model, which is not divided into modules or stacks, it receives perception signals and then outputs decision-making signals.
A more well-known practice is the V12 version of Tesla's FSD, where Musk ordered the elimination of almost all human rule codes, and more than 99% of the decisions were handed over to neural networks.
Zhou Guang has the same opinion as Musk.
But it is said that Zhou Guang shared this cognition earlier than Musk, in March 2023, Zhou Guang met with a certain bigwig and talked about this, and then the bigwig met with Musk in Silicon Valley in May, and Musk talked about advancing towards an end-to-end route - which also made the bigwig realize that there is something about the young CEO of Yuanrong Qixing.
Zhou Guang was very excited and emotional about this, and his statement was that if you choose the right technical judgment in a row, it means something...... This continuity, connected to the end-to-end front, the former fusion advocated by Zhou Guang and Yuan Rong Qixing's team, and no map.
No map means not relying on high-precision maps. As a kind of "redundancy", high-precision maps allow intelligent driving to make rapid progress on high-speed and fast roads, but once because of high-precision maps, the speed of mass production of intelligent driving was greatly restricted, especially the speed of entering the city. High-precision maps are also a must-have for L4 landing.
Zhou Guang revealed that when Yuan Rong went to mass production and qualification review this year, he directly took the navigation map, and where there is a navigation map, he can drive intelligently, but found that this is not the case with other players, such as at some intersections, there will be high-precision maps to assist in patching.
Therefore, the CEO of Yuanrong Qixing emphasized that his plan is "true no map", which is the embodiment of technological advancement, because only end-to-end can be truly without a map, there is no end without a map, and only end-to-end and true no map can achieve global coverage faster and can be opened across the country.
In fact, there was a scandal between Yuan Rong Qixing and Huawei, and the plan that Zhou Guang talked about is similar to the latest release of Huawei's intelligent driving.
But when asked about the real situation with Huawei, Zhou Guang chose not to answer.
There is a saying on the rivers and lakes that behind Huawei's "far ahead" is actually the supply of Yuanrong Qixing; Another theory is that Yuan Rong has only had contact and communication with Huawei at the level of intelligent driving.
Either way, the facts are clear:
First, both Huawei and Yuanrong Qixing have chosen a mapless, end-to-end intelligent driving promotion route.
Second, Huawei's intelligent driving solution, which is now publicly available, is very clear: it is completely self-developed.
So where is the mass production plan of Yuanrong Qixing?
Zhou Guang once again said that he could not comment because of "customer first".
But at the interview site, he seemed to be waiting for a customer who might come to the booth at any time, and he told his colleagues many times: the door is open, don't close it, X always comes, I need to go out immediately.
And Mr. X is the president of a giant car factory.
Intelligent driving is also divided into high-level and low-level
I asked Zhou Guang if he cares too much about the distinction of means - if the end point of autonomous driving is the same, the final thing is to realize the takeover of various road conditions and scenarios by AI drivers, whether there are pictures or not, end-to-end or not, and whether it is considered AI 2.0...... Isn't it all essential?
Zhou Guang disagrees, he believes that cognitive and conceptual caves are very important, which determines the process and the result.
One of his repeated remarks is: programming is not equal to AI, and general purpose cannot be achieved by rule base...... Only end-to-end can achieve universal autonomous driving, or artificial general intelligence in real-world driving scenarios.
So Zhou Guang's view is, what is the direct impact of end-to-end, AI 2.0? It is the concept and stage that has been divided into L4, the AI 1.0 era, that has been falsified.
End-to-end will break the bounded regions and conditions, and L4 will no longer be meaningful and non-existent.
This point of view, which Zhou Guang expressed publicly earlier, caused an uproar, and was later thrown out as a question at the exchange site of a well-known L4 unicorn company.
The answer of the CEO of the L4 unicorn is to the effect that L4 is difficult, and players who can't make it naturally think that L4 is a pseudo-concept.
And Zhou Guang didn't intend to respond to this kind of response, he raised his hands like surrender and applause, "I have nothing to say, I bless, others are willing to run as far as they want on the road he insists on, and the farther he runs, the better." ”
His current understanding is very clear, mass production of autonomous driving (L2+\L3), and then completely unmanned driving (L5).
Because no map will accelerate the popularization of intelligent driving, Zhou Guang particularly emphasizes the popularization and implementation of second- and third-tier cities, which is the real test of technical strength.
Then the data snowball brought about by popularization will achieve qualitative change, or emerge, in the end-to-end network and iteration.
Zhou Guang is an analogy with the iPhone, he believes that end-to-end means the beginning of AI 2.0, which means that the iPhone has been invented, and then the data snowball will accelerate the realization of qualitative change of the generation of iPhone, the general trend.
Zhou Guang also emphasized that the impact of end-to-end is not only at the level of autonomous driving, but also in the intelligent driving track, which will also usher in a new pattern reshaping.
At present, in the field of intelligent driving, the echelon pattern is realized according to the technology axis and the mass production axis, and the NOA in the urban area is considered to be the first echelon in technology, and the mass production progress of the intelligent driving scheme is regarded as the core variable in the data, and the technology and data are combined to constitute the overall competitiveness.
Zhou Guang agrees with the two dimensions of technology and data, but his view is that the implementation and experience of urban NOA will become the most critical variables.
Those that cannot achieve urban NOA, including high-speed NOA, can only be called "low-level intelligent driving". Low-level intelligent driving can rely on high-precision maps and rule bases...... But high-end intelligent driving is not good, only end-to-end.
And Yuan Rong Qixing, now a player in the high-level intelligent driving track - at least got the ticket.
In Zhou Guang's view, the race for high-end intelligent driving has just begun, and if he is asked to make a list of suppliers, including Yuan Rong, he thinks there are only three in China.
In fact, it is the end-to-end impact on Yuanrong Qixing that is fundamental.
In the five years since Yuanrong Qixing was founded, or earlier, the company's vision was still L4 and autonomous driving.
But recently, Zhou Guang led the team to make adjustments: "to create general artificial intelligence in the physical world".
This is a much more ambitious goal than autonomous driving, which includes autonomous driving, robots, or more fashionable embodied intelligence.
However, Zhou Guang also stressed that the car is the only way to realize this vision of general artificial intelligence in the physical world, because end-to-end systems require massive amounts of data, and there is no other form of robot that can provide and realize it today, except for cars.
So when asked "Do you still have contact with L4 friends", the CEO of Yuanrong Qixing gave this answer:
"I communicate a lot with people at AGI now, and I like to communicate with great people."