[TechWeb] if electrification was the key word of last year’s Shanghai auto show, then intelligence is one of the most important trends of this year’s Beijing auto show.
With the rapid development and popularity of AI large models, the deep intelligence of cars is gradually becoming possible, and the boarding of large models is accelerating this trend.
With the blessing of the large model, the full-scene intelligent driving NOA is speeding up landing.
in the cockpit, by constantly learning and improving the perception ability of voice and vision, the intelligent system is becoming more humanized.
At the scene of this auto show, Xiao Peng announced that the AI Tianji system, which is based on the in-depth application of AI models, will be the first in the world, and the car model product NETA “Qiankun Circle” will get on the bus for the first time.
Even Tencent and Baidu, which have not participated in the exhibition, have recently released solutions for boarding large models and signed cooperation agreements with foreign car companies to help them achieve a breakthrough in the level of intelligence.
What is the application of the big model to get on the bus? At present, car owners on large models should be used in two fields: intelligent driving and intelligent cockpit.
In the field of intelligent driving, autopilot used to write the corresponding solution in advance according to a specific scene, and once the corresponding situation was perceived on the actual road, the vehicle would deal with it according to the previously set way.
With large models, vehicles have the ability to perceive, predict and make decisions.
For example, in the perceptual side, the most commonly used scheme is BEV+Transformer, in which vehicles can identify all kinds of objects around the vehicle and build dynamic real-time maps through cameras and lidar.
After perceiving the situation around the vehicle, the large model will quantify and model the risk behavior of other traffic participants and predict the trajectory of other traffic participants for some time in the future.
Finally, the system will plan the optimal driving path according to these information, and directly output the corresponding driving instructions such as acceleration, deceleration and steering to the vehicle side.
In addition, the advantage of the large model lies in the continuous learning ability.
Through the training of a large number of actual road videos, the large model can be continuously trained and tested in the cloud to help the system iterate and optimize quickly.
At present, navigation-assisted driving in Xiaopeng urban area has entered 243 cities across the country, and it is planned to achieve point-to-point intelligent driving throughout the country this year.
the “official version” of ideal full-scene intelligent driving NOA will cover highways and ring roads across the country and 100 cities.
Huawei’s non-map commercial assistant driving is known as “nationwide open”, applicable to large and small cities and township cities, with 99% of the available distance.
With regard to the development prospect of the big model “getting in the car”, Baidu founder Li Yanhong said that with the progress of technology, the car robot will realize the experience of driving a “new version” every day, just like buying a new car, bringing continuous surprises to users.
Compared with autopilot, it is easier for users to perceive the application of large models in the field of intelligent cockpit.
For example, in the cockpit of Ji Yue car, Baidu Wenxin one-word model is integrated, and many operations on the screen can be done “can be seen.
” In addition, the voice assistant does not have to say the command system before it can be executed, but now with the blessing of the big model, the system seems to be able to better understand the user’s mind.
For example, when you say “I’m a little hot”, the system will automatically turn down the air conditioner.
Even if there are multiple instructions, continuous dialogue and so on, the large model can understand the different needs of each person and meet different operation instructions.
Xia Yiping, CEO of Jiyue, said that the current utilization rate of intelligent voice system by Jiyue owners is 98%, with an average of more than 60 times per person per day.
In addition to the voice assistant, the current application of the large model in the cockpit also includes the function of in-cabin text diagrams, which can transform users’ simple text into pictures, and even achieve more complex functions such as generating travel strategies for users.
The big model of Beijing Auto Show accelerated to get on the bus.
From the observation of this year’s Beijing Auto Show, many car companies have announced their plans to get on the AI model.
At the press conference of Xiaopeng Automobile, Xiaopeng CEO he Xiaopeng announced that he Xiaopeng will fully access the AI model with Tianji system as the entrance.
In terms of intelligent cockpit, based on the in-depth application of the AI model, Xiaopeng’s life assistant, AI Xiao P, will have a higher level of understanding and be able to combine more than 100 cockpit functions through personalization.
In terms of intelligent driving, based on the end-to-end large model, Xiaopeng’s XNGP high-order intelligent driving assistance system will complete the upgrade of the perceptual large model and control the large model to get on the car.
Specifically, Xiaopeng car will occupy a large network model through 2K pure vision, reconstruct the world with more than 2 million high-precision grids, and clearly identify every detail of dynamic and static obstacles.
The theme of this year’s auto show is “ushering in a new era of AI model definition”.
Guan announced that the large model product NETA “Qiankun Circle”, which was jointly developed with 360th, will get on the bus for the first time through OTA in June.
Officials say that after the big model gets on the bus, Nezhan L will be able to play “five roles in a car”, including voice secretary, cockpit babysitter, extended-range housekeeper, all-round bodyguard and smart driver, which will also improve the market performance and influence of Nezhan L.
The day before the opening of the auto show, Lantu held a separate technical communication meeting and announced cooperation with Huawei in the areas of smart cockpit and smart driving.
With the support of the VCOS GPT model, the interaction and perception of AI voice in the car has more personality and emotion, and the car control scheme can be given according to the user’s feelings rather than instructions.
According to the official data given by Lantu, in terms of voice interaction, the AI voice interaction of Lantu new generation intelligent cockpit has a speech recognition rate of 98%, the conversation response time is only 550ms, and the vehicle control response has reached the industry-leading 1s level.
In addition, both Baidu and Tencent have recently released solutions for getting on the bus with large models.
For example, Tencent released a “global intelligence” plan for a large model of the automobile industry, not only for the field of intelligent cockpit, but also covering all aspects from research and development, production to marketing.
In addition to domestic car companies, multinational car companies and joint venture brands are also very positive about the boarding performance of AI models.
At the auto show, Toyota announced that it would work with Tencent on AI models, cloud and digital ecology around “software-defined cars”.
Nissan also signed a memorandum of understanding with Baidu to conduct a feasibility study on strategic cooperation in the field of artificial intelligence and intelligent vehicles, and future Nissan models in China will also be equipped with Baidu AI solutions.
At the end: there is no doubt that it is bigModel boarding is becoming an industry consensus.
However, in the long run, it remains to be seen whether the large model can bring practical value to users and sales to car companies.
At least for now, if car companies want to get on board large models, they need to invest heavily in hardware computing power, software research and development, etc.
, and have a long payback period.
For many users, the improved voice interaction capabilities of the large model car are indeed worthy of recognition, but are many other more complex functions really needed? For example, the text-generating function has almost no usage scenarios in the car, and these functions cannot be a decisive factor in whether a user places an order.
It will take time to test whether large models can redefine automobile intelligence.
(Xiao Jian), return to the home page of First Electric Network>.