"Based on the mid-level computing power computing platform of Journey 6M and Journey 6E, the cost of inclusive urban NOA and high-speed NOA systems will be reduced by 40%, truly promoting intelligent driving to make it within everyone's reach. The cost of the system is within 5,000 yuan, and within 5,000 yuan, comrades. ”
On the eve of the Beijing Auto Show, at the 2024 Intelligent Driving Technology Product Launch Conference, Horizon Robotics released the Journey 6 series of in-vehicle intelligent computing solutions based on the new BPU Nash architecture.
(Dr. Chen Liming, President of Horizon Robotics, and Zhang Hongzhi, Vice President, attended the launching ceremony of the strategic cooperation between the two parties)
During the Beijing Auto Show, many intelligent driving algorithm manufacturers also successively released mid- and high-end intelligent driving solutions based on Journey 6.
Yihang Intelligent released the "Duxing" urban intelligent driving platform, using the mass-produced BEV "Lingmu", including the urban intelligent driving standard version, the urban intelligent driving performance version and the urban intelligent driving all-round version of three series of solutions, among which the first urban intelligent driving standard version based on the horizon journey 6E covers high-speed NOA, urban memory navigation and other high-end intelligent driving functions, the whole set of prices is not higher than 5000 yuan.
This is what Yu Kai said at the Horizon press conference "within 5,000 yuan" high-end intelligent driving program. After the press conference, Leifeng.com had an exchange with Chen Yuxing, founder and chairman of Yihang Intelligence.
The NOA function is an important node on the mass production route of autonomous driving
According to Chen Yuxing, Yihang completed the actual vehicle deployment of the Journey 6E in only 2 weeks, plans to deliver high-speed NOA within the year, and will further upgrade high-end functions such as urban memory navigation.
In his view, the NOA function is an important node on the mass production route of autonomous driving. Before the advent of NOA, all assisted driving functions were fragmented, but NOA implements point-to-point autonomous driving functions, thus bringing users an immersive driving experience.
In 2022, Yihang Intelligent will mass-produce high-speed NOA based on high-precision map and visual convolutional neural network technology. This year is also generally regarded by the industry as the first year of high-speed NOA mass production.
On the basis of high-speed NOA technology, the NOA function has two major development directions, one is to expand high-speed NOA scenarios, and the other is to improve the product performance of existing high-speed NOA.
In 2024, Yihang Intelligent will successively add urban application scenarios on the basis of high-speed NOA, and develop the urban memory navigation function; At the same time, the BEV algorithm is introduced for high-speed NOA to improve the perception performance of high-speed NOA.
According to Chen Yuxing's judgment, with the accumulation of data and the continuous iteration of algorithms, it will enter the city-wide NOA mode in 1-2 years, that is, the BEV plus no map scheme, laying the foundation for the ultimate autonomous driving solution. And around 2028, the industry can reach the ultimate solution for end-to-end autonomous driving.
At present, by adding a road memory module on the basis of the high-speed NOA algorithm framework, Yihang Intelligent has realized the urban memory navigation function. Among them, more than 85% of the new functional algorithms are derived from the reuse and fine-tuning of high-speed NOA algorithms, that is, the mass production of high-speed NOA algorithms is equivalent to having the backbone algorithms of urban intelligent driving.
Chen Yuxing revealed that high-speed NOA has been mass-produced in SAIC, BAIC and other OEMs.
In terms of perception performance improvement, through the iteration of the high-speed NOA algorithm, the performance of high-speed NOA is improved in the dimensions of traffic efficiency, perception ability, and safety, so that high-speed NOA can truly be used from usable to easy to use.
For example, by optimizing the multi-lane path planning and traffic flow monitoring mechanism, the efficiency of automatic overtaking and merging is improved, so as to achieve higher traffic efficiency. By improving the detection range of road boundaries, the temporary construction scenarios of roads can be identified more accurately, and the technical scheme of heavy perception and light map can be realized.
If you want to mass produce BEVs, you can't do without traditional perception algorithms
Whether it is high-speed NOA or urban memory pilot, it is a point-to-point autonomous driving realized in a specific scenario, and it cannot cover the entire city. With the development of the industry, there are still some shortcomings in the detection process of traditional algorithms, such as the fusion mechanism of sensors. Therefore, in order to further expand the urban intelligent driving scenario, it is necessary to introduce BEV perception algorithms.
BEV can solve the problem of synchronization and registration between multiple cameras, and can also solve the problem of light map. However, the mass production of BEVs is inseparable from traditional perception algorithms. In order to promote BEVs to meet the conditions for mass production at the vehicle specification level, it is necessary to combine the core modules of traditional perception algorithms as a supplement to achieve necessary capabilities such as remote focus perception and traffic sign recognition.
Chen Yuxing believes that BEV is still deficient in some specific perception problems for long-distance target processing, and traditional perception can make more precise detection for long-distance objects. In addition, in the BEV training process, it is also possible to use the traditional perception that has been proven in mass production to develop a shadow mode-like method, and even to train some BEV data.
Therefore, the mass-produced BEV "Eye" released by Yihang is actually a BEV plus a traditional perception module to supplement.
In addition, the mass production of traditional perception algorithms is also an important definition of the mass production conditions of BEVs, and only after mass production of traditional perception algorithms can the BEV meet the mass production conditions.
Chen Yuxing said, "If we don't have the experience of realizing perception into mass production, it may be difficult to distinguish when BEVs will reach mass production conditions. The second level, from the perspective of BEV, also needs to be supplemented by traditional perception in terms of technology or accuracy. ”
"Therefore, in the "Eye" solution we released, traditional perception and BEV are equally important, and this is a perception accuracy that we finally obtained after mass production of many projects, a large amount of data collection, calibration, and iteration. ”
Autonomous driving in cities is the second half of autonomous driving
According to Leifeng.com, the urban intelligent driving platform released by Yihang Intelligent this time - Duxing, is divided into three different product lines.
The first is the standard version of urban intelligent driving, the main functions are urban memory navigation, high-speed NOA with extreme performance, and BEV perception;
The second is the urban intelligent driving performance version, which will add a lidar for safety redundancy, and the performance will be improved by more than 50%, and finally a lightweight urban all-scene NOA will be realized.
The third is the all-round version of urban intelligent driving, which realizes the NOA function of all urban scenarios, applies the large model of autonomous driving, and successively transitions to the end-to-end solution.
On the Duxing platform, the main recommendation is based on the Horizon Journey 6E scheme, which has four different configurations (the main push configuration is 7V1R configuration), and the functions that can be achieved include urban memory pilot, high-speed NOA, memory parking and L2 full-function scenarios.
Chen Yuxing said, "It is difficult to reach the price of 5,000 yuan for mass production of high-speed NOA, and we can achieve less than 5,000 yuan in 7V1R, and increase the memory driving function." If there are car manufacturers who are more cost-sensitive, we have also launched a 5V1R lower-cost solution to achieve high-speed NOA functions with extreme performance. ”
In addition, Yihang has also launched 9V1R and 11V1R configuration schemes.
(Based on the mass production rhythm of the J6E solution)
Chen Yuxing told Leifeng.com, "We basically completed the test deployment of the journey 6E in two weeks, and it is expected that the mass production of high-speed NOA can be carried out by the end of this year, and the memory navigation assistance can be realized in May and June next year (2025)." ”
Returning to the development status of the intelligent driving industry, new and old car companies including Wei Xiaoli, Huawei, Great Wall, GAC, Chery, BYD, Geely, SAIC and other new and old car companies have entered this high-end intelligent driving battle.
The user experience upgrade brought about by high-end intelligence has become an important selling point for car companies. Except for a few new forces that insist on full-stack self-development, most of the remaining traditional car companies have adopted the form of self-development + external cooperation, which is also the window period for the fierce competition of intelligent driving algorithm companies.
Many players, including Huawei, DJI, Momenta, Yihang Intelligence, Momo Zhixing, SenseTime, Pony.ai, Qingzhou Zhihang, etc., need to achieve greater compatibility of high, medium and low configuration hardware through more adaptable algorithm models on the basis of systematic cost reduction of software and hardware.
In the exchange session after the press conference, Chen Yuxing said that urban autonomous driving is the second half of autonomous driving, because it can solve the high-frequency needs of end users. However, urban autonomous driving is still in its early stages, and it depends not only on the competition between autonomous driving companies, but also on the development of "teammates" - car companies.
If the sales of car companies are not good, the overall cost of car companies will be further compressed, and it will also be a round of survival test for intelligent driving companies.
One thing worth noting is that one of the buzzwords at this year's auto show is "end-to-end".
Chen Yuxing said in the communication after the press conference, "End-to-end is the ultimate intelligent driving solution that we believe is relatively clear, and it should be based on BEV mass production, because there is not so much experience and data, and end-to-end needs to be accumulated." ”
Chen Yuxing believes that an end-to-end enterprise cannot only focus on algorithms, but only have real advantages with algorithms, data, experience, and iteration.
However, the end-to-end problem is that, in addition to the so-called "black box", there is also the reality of huge consumption of computing power.
Tesla CEO Elon Musk, who arrived in Beijing on April 28, said on social media, "Tesla will invest about $10 billion this year in AI training and inference, and inference is mainly used in cars." ”
Musk also added that any company that doesn't spend $10 billion a year or can't deploy efficiently can't compete in the market. End-to-end autonomous driving requires strong computing power, and the end-to-end realization depends on whether the cash flow and sales volume of car companies in the industry can support R&D costs.
Therefore, car companies need to have more algorithms and cost considerations for suppliers. Yihang is also developing lightweight BEVs that implement functions on relatively inexpensive chips to meet the needs of car manufacturers. Leifeng.com