"It is necessary to have an objective understanding of the urban NOA. ”
This is the statement of Yu Kai, founder and CEO of Horizon Robotics, at the Automobile 100 Forum, and it is also the voice of many intelligent driving workers in the industry.
It is generally believed that the current high-speed NOA on the market has just reached the "easy to use", and has not yet let consumers "love to use", and the current situation of urban NOA, I am afraid that even "usable" is a problem.
It can be seen that the shirtless battle of high-level intelligent driving still has an "endgame" without winning or losing. At this stage, track players are concerned about solving the problems of "usable" and "loving", and then solving the "easy-to-use" problems. The pass rate, more anthropomorphic experience, and higher traffic efficiency of standard scenarios have become the "pain points" of intelligent driving that all enterprises focus on.
Judging from this year's Beijing Auto Show, the technical hotspots emphasized by intelligent driving companies and car companies have switched from last year's BEV+Transformer to "end-to-end". It can be seen from the press conference of the auto show media day that most executives will mention the "end-to-end" strategic deployment when they mention the intelligent driving business——
In the next few years, it seems that whoever really gets through the "end-to-end" will likely get the ticket to victory in the endgame of intelligent driving.
However, from the perspective of Tesla, the originator of "end-to-end", based on the company's investment intensity, the end-to-end large model is actually very money-burning, and it may have to start with tens of billions.
He Xiang, a data intelligence scientist at Momo Zhixing, once mentioned in an interview that the end-to-end model is a pure data system, and in principle, the model parameters are large enough and the data distribution is better, so it can be continuously improved. Burning money and time-consuming, but also with high ceilings.
High-level intelligence competes for beauty
Huawei
On the eve of the Beijing Auto Show, Huawei unveiled its new "Qiankun" brand with intelligent driving as the core. Among them, ADS 3.0 adopts a new end-to-end architecture, based on the GOD (General Obstacle Recognition) network, which realizes the progress from simple "obstacle recognition" to in-depth "understanding of driving scenarios".
The new architecture of Qiankun ADS 3.0 uses the PDP (Predictive Decision Regulation) network to achieve pre-decision-making and planning, so as to achieve human-like decision-making and planning, with a more human-like driving trajectory, higher traffic efficiency, and a >96% pass rate at complex intersections.
Xiao-hsien
It is reported that based on the end-to-end large model, the XNGP high-end intelligent driving assistance system of the car will complete the upgrade of the perception model and the regulation of the large model on the car.
He Xiaopeng said that at present, Xpeng Motors has begun public testing of the end-to-end solution, and will officially share the end-to-end actual data performance in May. Through the industry's first mass-produced 2K pure vision occupation network model, the world is reconstructed with more than 2 million high-precision meshes, and every detail of dynamic and static obstacles can be clearly identified.
SenseTime
As the first company to propose a universal model of autonomous driving with integrated perception and decision-making, SenseTime showcased the road test of UniAD (Unified Autonomous Driving), an end-to-end autonomous driving solution for mass production, for the first time at this year's auto show.
The scheme can rely only on the actual road test results of visual perception without high-precision maps. Whether it is a complex urban road or a rural road without a center line, the vehicle can efficiently and accurately complete a series of difficult operations including turning left on the bridge at a large angle, avoiding vehicles occupying the road and construction areas, and bypassing running pedestrians.
Bosch
Although it does not emphasize the end-to-end layout, Bosch, the traditional supply chain giant, has become a bright spot for Tier1 in high-end intelligent driving. Bosch's intelligent driving technology and products brought to the show showed a leap forward.
Its latest cabin and driving integrated solution, which can handle intelligent driving and intelligent cabin with only one chip, is expected to be mass-produced in 2026, compared with the current intelligent cabin and intelligent driving dual-chip solution, the cost can be reduced by 30%.
Bosch China's high-end intelligent driving solution, developed locally by Bosch Intelligent Driving & Control Systems in China, will be mass-produced in 2023. Among them, the city pilot assistance function will be gradually pushed from May.
horizon
Horizon Robotics exhibited SuperDrive, a full-scene intelligent driving solution, which is a high-end urban intelligent driving model room built by the company. If Journey 6 is the culmination of hardware in the field of Horizon hardware and software, SuperDrive is a moat at the software level.
With the end-to-end perception architecture of dynamic, static, and OCC (Occupancy Network) triple play, and the data-driven interactive game algorithm, SuperDrive can take into account the scene passing rate, traffic efficiency, and behavior anthropomorphism in any road environment.
It is reported that SuperDrive can flexibly handle complex traffic flows like old drivers, and the success rate of lane change in congestion scenarios is increased by 50%, and the pass rate at intersections is increased by 67%.
A new money-burning pit
At the end of 2022, SenseTime and its joint laboratory proposed the industry's first universal autonomous driving model with integrated perception and decision-making, UniAD, which won the best paper at the 2023 International Conference on Computer Vision and Pattern Recognition (CVPR) the following year.
If SenseTime was not able to make the industry feel the potential value of end-to-end in 2022, then in 2023, the paper award won by this company finally made the academic circle and the automotive industry realize the broad prospects of end-to-end in the autonomous driving track.
At the practical level, Tesla tried the live broadcast test of Tesla's FSD V12 last year, so that more enterprises can see the possibility of end-to-end implementation, and in March this year, the FSDV12.3.1 version was pushed to North American car owners, introducing an end-to-end neural network, which instantly set off a great wave in the industry.
Many people default to the fact that Tesla is the originator of end-to-end. Because of this, Tesla has also become a pioneer case in the industry to peel back the cocoon and understand the end-to-end in detail. Previously, a consulting company analyzed that referring to Tesla's investment intensity in recent years, the end-to-end model is actually very money-burning, and it may start with tens of billions.
As mentioned above, He Xiang, a data intelligence scientist at Momo Zhixing, once mentioned in an interview that the end-to-end model is a pure data system, and in principle, the model parameters are large enough and the data distribution is better, so that it can be continuously improved.
The ceiling is high, but it burns money and takes time.
In addition, end-to-end model training is very dependent and consumes computing power, and in 2024 alone, Tesla plans to invest more than $1 billion in Dojo supercomputing.
In early 2023, Tesla had said that it had analyzed 10 million video clips collected from Tesla customers' cars. And Tesla judged that it would take at least 1 million diverse, high-quality clips to complete an end-to-end autonomous driving training to work properly.
Not only is it a huge hole in burning money, but end-to-end is also a long-term project, and it is difficult for many companies to see the results in the short term. This also means that end-to-end is a game for a few enterprises, and only leading companies with sufficient capital reserves and willing to invest for a long time can have the ability to support the huge data and computing power required end-to-end.
SenseTime said at the auto show that many companies in the industry are laying out end-to-end solutions, but they are generally divided into three types -
The first is the cloud computing project where the decision-making layer and the perception layer are independent of each other, the second is the combined end-to-end, which means that the connection between the perception layer and the decision-making layer requires a strong intermediary and is easier to implement, and the third is the real end-to-end, which is the solution currently developed by SenseTime, which can enable enterprises to meet the needs of intelligent driving at the lowest cost.
However, SenseTime did not mention how much money has been invested in this long-standing end-to-end plan.