Head of Great Wall MotorsWei Jianjunalso went live in person!
It is the most difficult of all subjects.Unpictured City NOA.
Advertisement
What’s even more challenging is that the route chosen by Lao Wei not only includes the city’s main roads, but also narrow roads, construction roads, or extremely complex irregular intersections in the old urban area of Baoding, a “second-tier city”.
Without the rules and regulations of Beijing, Shanghai, Guangzhou and Shenzhen, “chaos” is better, but the more chaotic it is, the more it tests your wisdom.
Wei Jianjun said directly:
Advertisement
Juanzhijia comes to Hebei and welcomes friends to challenge our urban NOA routes.
Wei Jianjun’s live broadcast of measured NOA without pictures Overview
The system configuration measured by Wei Jianjun is based on the Orin-X high-computing domain control platform, equipped with 1 lidar, 3 millimeter-wave radars, 11 high-definition cameras, and 12 ultrasonic radars.
The time is 10:30-12:00 in the morning.
The live broadcast route starts from Baoding Great Wall Haval Technology Center and passes through a short section of highway:
The high-speed scene is not difficult, and the system can complete the up and down ramps independently.
Wei Jianjun commented:
Great Wall's high-speed NOH has been put into mass production a few years agoit was the industry leader at that time, and now there is no challenge.
Old friends who are familiar with smart car reference all know that the high-speed NOH mentioned by Wei is the autonomous driving company of Great WallHao Mo Zhi XingThe product,Back in 2021It is mass-produced in Wei brand, Tank, Euler and other models.
At that time, the mainstream of the industry was still competing to see whether the L2 Family Bucket was standard.
The challenge really began when we got off the highway and entered Baoding city. Wei Jianjun also focused on Baoding's road conditions.
Compared with first-tier cities such as Beijing, Shanghai and Guangzhou, the characteristic is that there are many irregular and asymmetric intersections:
Violations of traffic regulations occur frequently:
And, when the road is narrow, it is necessary to complete the diversion of turning and changing lanes:
The old city is being renovated, and the construction on the road is numerous and irregular:
And this set of unpictured NOA also successfully challenged the hell-level difficulty of Baoding’s unique “Liudaokou”:
The intelligent driving system made zero errors throughout the entire process, with no emergency exits or downgrades, and it perfectly completed various maneuvers, detours, and avoidances.
Wei Jianjun said,Great Wall has been working on autonomous driving for 10 yearsbut everythingTaking “safety” as the keynote, we will never show off or brag if we are not sure.this time I participated in the live broadcast in person, which is because I have enough confidence in the progress of my own products.
In fact, there is a way to test smart driving in the lap test, and the “gold content” can be divided into three levels.
The third level is the “zero takeover” demo that is often seen at press conferences and official accounts. You know… I can't say that it doesn't show any strength, but it is definitely the result of careful “polishing” by the official, and the degree of reference is limited. .
The second level is what the government proactively provides for everyone to experience, such as media test drive sessions and 4S stores’ smart driving experience routes for ordinary users. The selection of the route was “refined over time” and run through countless times in the backstage. It was an open-book examination. Some basic skills can be seen, but in the face of complex and changeable open roads, versatility and stability cannot be guaranteed.
The first gear is to throw it directly into the rolling traffic without any special adjustments or presets. This is the best way to test the true level of intelligent driving. Spontaneous tests by users and fans are common.
So we can see the two key factors of the smart driving test, one isIs there any optimization and debugging for specific routes?;The second isDo you use high-definition maps in the background?.
Wei Jianjun's live broadcast measured the NOA of the Great Wall. Which category does it belong to?
Great Wall Motor revealed to Smart Car Reference that the NOA just measured by Wei Jianjun,There is definitely no high-precision map.. Completely real-time perception, and OCC, which is much hyped in the industry, is also on board.
Whether it was specially prepared or not, the Great Wall did not deliberately hide it.becauseWei Jianjun suddenly decided to broadcast the live broadcast last Monday.without any notice or plan.
So in the past seven days, the Great Wall Intelligent Driving team has been optimizing the software algorithm, but this optimization is aimed at the vehicle-side algorithm itself, not the test route.
In fact, the road conditions in Baoding are too complex and unexpected, and the road renovation construction changes day by day.It’s not impossible to have an “open book exam”.
This is why in the live broadcast barrage,Some users bluntly said that Wei Jianjun is a “real person”.
Because he appeared in a live broadcast for the first time, in addition to Great Wall's demand and desire for traffic (who doesn't have it~), the core purpose is to showcase Great Wall's technical strength and product progress in smart driving.
And the Great Wall is by no means just talk. The urban NOA tested by Lao Wei live today is still the engineering version, but Great Wall revealed that the mass-produced version will be launched in the middle of this year, and the first model will be the new Wei brand Blue Mountain.
In fact, the interior of the new Blue Mountain has been “leaked” in the live broadcast. I wonder if you have noticed it?
Great Wall NOA, what kind of technical system
Great Wall Motors is the first among the established car companiesDemonstrate an accurate understanding of intelligent driving technology and AI systematization capabilitiesplayer.
It is also the only player in the traditional automotive camp that truly understands and is involved in AI, rather than trying to solve the problem once and for all by purchasing “turnkey” solutions from suppliers.
In fact, this does not solve the problem at all. The “first brother” of independent new energy has suffered greatly from this. To this day, it can't get rid of the “unintelligent” label. It can only blur the concept of electrification and package it as intelligent.
So, what kind of understanding does the Great Wall have and what kind of technical system does it have?
The NOA function that Wei Jianjun tested today is behind Great Wall’s latest SEE integrated large model.
The key point isIntegration. The core connotation is link integration that realizes integrated perception and decision-making.
Not a strictly end-to-end large model because this “big black box” lacks interpretability.
There are still sub-models such as perception, planning, decision-making, etc., but they are no longer independent modules. First, the perception model based on the Transformer architecture extracts features from the input data of each sensor.LosslessInput to subsequent planning and perception models.
Integration is reflected in the fact that Great Wall uses data-driven instead of rule-driven.Instead of concretely defining what good driving behavior is by rules, the model is allowed to directly learn “how to drive” like a human driver. The complexity of the road conditions it can handle is greatly increased, and the decision-making driving behavior is closer to humans..
“Lossless” is reflected in the OCC occupying the network and subverting the previous “whitelist” mechanism of perception and recognition. The system does not need to know what obstacle is blocking the road, it only needs to understand where it is, coupled with lidar. The dual redundancy is enough to output highly accurate perception results.
The planning, decision-making and other modules behind perception all use data-driven models, and such a large model with a massive amount of excellent human driving case data is large enough in itself.
Returning to the core of smart driving safety that Great Wall emphasizes most, in engineering practice, the algorithm will inevitably have false detections and missed detections, or some extreme driving behaviors that are inconsistent with human habits. Some are safety hazards, and some are not conducive to safety. Build user confidence.
Therefore, Great Wall specially added a layer of human-written rule “shell” to the data-driven algorithm as a protective measure in extreme scenarios.
Therefore, the meaning of SEE is actually this: multiple sensor redundancy at the hardware level, OCC at the software level, “safety” represented by the pre-fusion algorithm, and data-driven “efficiency” efficiency under the large model framework.
Finally, it is the “Experience” experience of being able to drive anywhere (with navigation) and park easily (various difficult parking spaces, memory parking, etc.).
If you still question the intelligence capabilities of Great Wall, there is probably one last angle:
What about computing power?Without computing power infrastructure, no matter how much you talk about it, it’s all nonsense.
The Great Wall also announced the latest situation:Kyushu ——The total computing power scale reaches 1.64 EFLOPS. Huawei, which was “far ahead” in the past two days, announced a figure of 3.3 EEFLOPS.
1 EFLOPS is equivalent to 10 trillion floating point operations per second.
In addition, Great Wall's Jiuzhou supercomputer can also have high-performance storage of 5T/s and high-performance network with communication bandwidth of 3.7TB/s. It can capture hardware exceptions in real time and recover from exceptions at the minute level. Kcal can continuously train without failures for several months, and the training cost of a 100-billion-parameter model can be reduced to the level of 100 Kcal per week.
Great Wall takes the lead in AI transformation, and the “Five Constants” for smart driving emerge
From the first live broadcast of smart driving by Wei Jianjun, the head of Great Wall Motors, we can see the approximate level of “leading” smart driving in the industry.
First of all, the functional experience no longer depends on how many L2 functions are standard, because it is useless no matter how many standard configurations there are. A bunch of ODD functions do not overlap, and the user experience is fragmented. In 2024, product cars like this will no longer be able to make it to the table.
Must be high-speed NOA entry standardfull coverage of smart driving from toll station to toll station, including macro navigation and entry and exit ramps, as well as micro game lane changes, overtaking detours, etc.
Secondly, by the end of 2024, urban NOA must also be mass-produced and put on the road, and it must be an experience that does not rely on high-precision maps – you can drive wherever there is a road.
From the construction of the technical system shown by the Great Wall, we can also see what understanding and strength the leading players need to have.
The first is the construction of computing power infrastructure. Taking Huawei as an example in China, the graph-free NOA must be based on at least several EFLOPS.
The second is the framework system. There are two key points here. One is the BEV perception model based on the Transformer architecture, and the other is the data-driven AI model that penetrates into all aspects of intelligent driving.
Data drive is very important, and it is directly related to whether smart driving products can be used and whether they are easy to use.
The last step is to turn demo testing into engineering capabilities for mass production, which includes verification and testing of technical solutions, sensor calibration, vehicle functional safety certification, software and hardware adaptation, etc.
Judging from such product progress and the systematic capabilities behind it, domestic independentDriving “Wuchang” IntelligentlyThe situation is gradually taking shape:
Representatives from major technology companiesHuawei, DJIrepresentative of new forces in car manufacturingXpeng Motorsrepresentative of autonomous driving startup company Momentaand the current only representative of the transformation of established car companies,Great Wall Motors.
And for smart driving, Great Wall has advantages that other players based in Beijing, Shanghai, Guangzhou and Shenzhen do not have – Baoding.
Momenta is also one of the partners of Great Wall Smart Driving. CEO Cao Xudong recently told Great Wall:
“More valid smart driving data collected in one week in Baoding is more than in one quarter in Shanghai.”