On August 30th, the 27th Chengdu International Auto Show (hereinafter referred to as "Chengdu Auto Show") kicked off. As the first A-level international auto show in the second half of 2024, intelligent and AI models are still the highlights of many car companies' exhibitions at this year's auto show.
A reporter from China Business News learned from an interview at the Chengdu Auto Show that brand car companies such as Jiyue and Jike have been intensively deploying AI big models this year, and have started a "race" in end-to-end AI big models. After the use of end-to-end AI big models, intelligent driving will become more intelligent, covering areas including urban areas from the previous coverage of highways to now.
In fact, as early as before the opening of the Chengdu Auto Show, new models equipped with end-to-end AI big models had already been released. On August 28th, Jiyue released the Jiyue 07 in Chengdu. According to the official introduction of Jiyue, this model is equipped with the first autonomous driving model, Apollo Self Driving, which supports unmanned driving. At the same time, the native AI model of Jiyue has also been upgraded from SIMO to SIMO2.0.
Coincidentally, Jike also announced the update of Haohan Intelligent Driving 2.0 this year. According to official information from Jike, thanks to Nvidia's 7-nanometer process, the chip with a total power of 508TOPS, and the Haohan Intelligent Driving 2.02025 model, Jike 001 has officially entered a new end-to-end era.
However, several industry insiders have stated in interviews with reporters that the large-scale entry of truly intelligent vehicles into the market still requires progress and development in various aspects such as technology, policies, and consumer awareness.
Bet on end-to-end AI big model
The hottest concept in current intelligent driving cars is undoubtedly the end-to-end AI (End to End) big model. According to relevant information, end-to-end AI big models refer to AI models in the field of autonomous driving that can directly process raw input data and output final results without going through multiple intermediate processing steps.
The end-to-end model integrates the three major modules of perception, planning, and control together, eliminating the boundaries between modules, making the system more concise and efficient, reducing information transmission losses between modules, and thus improving the overall efficiency and response speed of the system.
In the field of autonomous driving, the application of end-to-end models enables the system to process data faster, improve system response speed, and reduce reliance on high-precision maps and LiDAR, thereby lowering costs.
Taking the extremely advanced ADFM model as an example, there have been further improvements in driving ability, parking ability, safety ability, and other aspects. Taking driving ability as an example, the ADFM large model can achieve super strong obstacle avoidance, control the switches, brakes, and steering wheel simultaneously, and avoid high-risk vehicles while changing lanes.
It is understood that the above-mentioned intelligent driving functions of Jiyue will be implemented through users purchasing ASD intelligent driving packages. In the ASD intelligent driving package, the high-speed elevated navigation assistance function is a free standard. The buyout price of ASD intelligent driving package is 29900 yuan.
The staff at the Jiyue booth told reporters that after the segmented end-to-end AI model got on board, the response of intelligent driving was faster and safer than before. Using less code and increasing safety redundancy is a major trend in future intelligent driving. Moreover, the upgrade cost of intelligent driving is also lower, and it can be upgraded through OTA push notifications
In addition to Extreme Yue, Extreme Krypton is also working on end-to-end AI big models. According to the information released by Jike, the 2025 Jike 001 is equipped with Haohan Intelligent Driving 2.0, marking Jike's official entry into an end-to-end new era.
According to Jike, the end-to-end model of the entire scene in Haohan Intelligent Driving 2.0 includes a scene cognition model and an interactive regulation model. In the scene perception part, the model has added the ability to understand 1000 traffic scenarios, improving cognitive accuracy by over 120%; In terms of interactive regulation, it maximally reproduces the driving style of human drivers, and can predict the behavior of surrounding vehicles and pedestrians 1.2 seconds in advance, greatly improving the system's interactive game and path planning capabilities.
After being equipped with an end-to-end AI model, intelligent driving can be used in urban areas. For example, from point A to point B, it can record 20 routes and drive 100 kilometers on its own. "The staff at the Jike booth told reporters that the Nvidia chip equipped with the Haohan Architecture 2.0 has high computing power and stronger versatility. The end-to-end AI model will also be upgraded through OTA in the future, covering the vast majority of car models.
According to a research report by CITIC Securities, Chinese companies such as Huawei and Ideal Auto are actively following up on end-to-end models and are committed to promoting data collection and training of computing components to provide a foundation for end-to-end AI model training. With the gradual application of end-to-end and large model technologies, it is expected that the parameter count of intelligent driving algorithms will significantly increase, and the demand for computing power will also show an upward trend. The upgrading trend of intelligent driving hardware will continue.
Intelligent driving has become a 'must win' area for car companies
The end-to-end big model is becoming a "battleground" for car companies, as it is the main path to achieving advanced intelligent driving.
According to a research report by Guolian Securities, end-to-end neural networks can fully simplify computational steps, reduce the need for manual feature engineering, and identify correlations in data, thereby significantly improving computational efficiency. Benefiting from the increase in the scale of effective behavior trajectory data, the end-to-end large model of intelligent driving is expected to become a high-level intelligent driving solution. Starting from 2023, the modular end-to-end planning mode will accelerate and gradually become the main direction representing the iteration of high-level functions in intelligent driving
The research report related to Minmetals Securities also analyzed that the end-to-end model combines multiple modules into one model, reducing the cumulative error of multi task training. In an end-to-end solution, prediction, decision-making, planning, and control can learn the driver's driving behavior data, directly provide planning or control signals, and replace manual labor with training data
Originally, rule-based algorithms required a lot of manpower to write complex rules that listed all possible scenarios and coping strategies that vehicles could encounter. However, in actual environments, obstacles and scenarios that did not appear in the training data will always appear, and it is impossible to solve endless corner cases manually, "said Minmetals Securities.
But new technologies mean increased investment, and currently major car companies are investing heavily in end-to-end technology research and development.
Jiayue CEO Xia Yiping stated at a press conference on August 28th that Baidu has been researching and developing AI and autonomous driving for over a decade, investing over 150 billion yuan.
In addition, companies such as NIO and Leapmotor are focusing on end-to-end business. According to relevant media reports, Zero Run Motors has recently established an end-to-end team, and the person in charge of the research and development of intelligent driving algorithms is about to take up their position. In June of this year, NIO restructured its intelligent driving research and development department and abandoned the long-standing "perception decision regulation" technology route, exploring more clearly the use of end-to-end large models to achieve advanced intelligent driving.
However, in the eyes of industry insiders, domestic car companies still need to strive to catch up in the end-to-end aspect. The current mass production of new domestic car makers is to replace the planning module with neural network rules, and each module is still independent and segmented end-to-end. Minmetals Securities stated that it is expected that it will take about six months to a year for new domestic car makers to truly integrate end-to-end production.
Large scale market entry still takes time
In fact, in the eyes of industry insiders, AI intelligent vehicles are currently facing a series of challenges including technology, policies, etc., and there is still a long way to go before they enter the market on a large scale.
In the view of Zhang Lingling, Associate Professor of Marketing at the China Europe International Business School, the interaction between humans and vehicles may become one of the key areas for the expansion of intelligent vehicles. In many consumer scenarios, consumers hope that their needs can be easily met. If the interaction between cars and people becomes increasingly intelligent, understanding consumers' scene demands and decisions, and forming a certain frequency of interaction on this basis, it will become an entry point for many service platforms. Then there will be a very broad market space in the future, "said Zhang Lingling.
Zhao Chunzhang, the head of the automotive industry and senior consultant at Lis Category Innovation Strategy Consulting, told reporters that AI big models will be the most important core of future automobiles and an element that reflects the core competitiveness of various automotive brands. From 'someone driving' to 'no one driving', it has changed people's traditional mode of transportation. The AI big model has brought real changes to the automotive industry
But Zhao Chunzhang also stated that the arrival of truly intelligent cars depends on three aspects: technology, policies, and cognition. On the technical side, with the integration of AI models and the true breakthrough of intelligent driving, the experience has almost completely replaced human drivers. Currently, industry-leading car companies such as Tesla, Xiaopeng, and Huawei have reached or are close to this ability, and technology will no longer be a limitation
On the policy side, when governments around the world will open up autonomous driving and provide corresponding supporting infrastructure will affect the speed of popularization of smart driving cars, "Zhao Chunzhang further stated.
The reporter noticed that the national standard information public service platform recently showed that three mandatory standards related to intelligent connected vehicles, including "Technical Requirements for Vehicle Information Security", "General Technical Requirements for Automotive Software Upgrades", and "Intelligent Connected Vehicle Autonomous Driving Data Recording System", were released on August 23, 2024 and will be enforced from January 1, 2026. This means that the supervision of intelligent connected vehicles in China is being strengthened.
But in addition, the popularity of AI intelligent cars also depends on consumers' acceptance of new product categories. Although AI models can technically replace human drivers, most consumers still have certain concerns, and the promotion and education of product categories will take some time, "said Zhao Chunzhang.