news

the new era of autonomous driving is here.

although there are still three or four months to the end of 2024, the annual keyword of the automotive industry seems to be no surprise, that is, "end-to-end" technology in the field of autonomous driving. 2024 is also called a new era in the field of autonomous driving by the industry. from tesla's release of fsd v12 in august last year, to sensetime's release of uniad in april this year, baidu's release of apollo adfm in may, xiaopeng motors' release of xnet, xplanner, and xbrain in the same period, and ideal auto's release of end-to-end + vlm in july, almost all major manufacturers have their own end-to-end systems.

end-to-end (e2e) autonomous driving technology is an emerging technology path that breaks the boundaries of perception, decision-making, and control modules in traditional autonomous driving systems and integrates these three links into a unified neural network system. the core idea of ​​this technology path is to use deep learning models to directly automate the entire driving process from sensor input (such as cameras, lidar, etc.) to control output (such as steering wheel rotation angle, throttle and brake force).

this technical path simplifies the complexity of traditional autonomous driving systems and improves the overall efficiency and response speed of the system. however, end-to-end technology also faces challenges such as large data requirements, high computing resources, and system security and explainability. the industry generally believes that as the technology continues to mature, end-to-end autonomous driving will gradually become mainstream.

however, as things stand, due to the differences in end-to-end algorithms and application development speeds among various companies, for most users, end-to-end seems more like an advanced theory rather than a truly feasible function.

so, who can be the first to let users experience the "end-to-end" technology dividend? ideal auto is obviously at the forefront of the industry.

on september 14, ideal auto's new generation of intelligent driving "end-to-end + vlm" pushed version 6.2.0 (e2e-vlm beta4) of the intelligent driving system to thousands of users, creating the industry's first "parking space to parking space" product landing experience.

ideal industry's first end-to-end + vlm dual system solution, first-class user experience

ideal is not the first company in the industry to develop end-to-end technology, but why is ideal developing so fast?

this is to mention that ideal auto has demonstrated a unique innovative idea in autonomous driving technology: end-to-end + vlm dual-system architecture strategy. according to ideal's official information, this architecture is inspired by the dual-system theory proposed by nobel prize winner daniel kahneman, and it is the first dual-system intelligent driving solution in the ideal world. the design of this dual-system architecture aims to improve the stability and reliability of the autonomous driving system, and make the system more flexible and efficient by processing common scenarios with a fast system and complex scenarios with a slow system. ideal's end-to-end, unlike the traditional segmented type, adopts one model integrated end-to-end, which can ensure "high efficiency" in most scenarios, enable intelligent driving to have the driving ability of "old drivers", and support the rapid launch of its map-free noa function. the vlm visual language model is the world's first large model successfully deployed on a car-side chip, with logical thinking and decision-making capabilities to deal with complex scenarios. based on this technical solution, ideal auto has brought users a new intelligent driving product and experience form, namely the "one-button intelligent driving" function from parking space to parking space.

the world model supports the large-scale and high-speed iteration of the new generation of ideal intelligent driving, and provides an automated ai capability evaluation system. through reconstruction technology, the problem scenarios encountered by users are turned into "wrong question sets", and through generation technology, the user's real driving scenarios are applied as "simulated questions". these two technologies ensure that wrong questions will not be answered wrongly during model evaluation, while also having excellent generalization capabilities.

in terms of the extremely challenging data closed loop, ideal auto effectively verified the function of noa by combining the world model in the cloud with the shadow mode on the vehicle side. at the end of august, at the chengdu auto show, lang xianpeng, vice president of intelligent driving r&d of ideal auto, said: "thanks to the huge training data, powerful computing power and ai capability evaluation system, the iteration speed of the new generation of ideal intelligent driving is amazing. ideal auto's current accumulated training mileage has exceeded 2.2 billion kilometers, and it is expected to exceed 3 billion kilometers by the end of 2024; the current training computing power has reached 5.39eflops, and it is expected to exceed 8eflops by the end of 2024. ideal auto invests more than 1 billion yuan in training computing power each year. we believe that the training computing power required to ultimately achieve autonomous driving must reach the order of 100eflops, which is equivalent to an annual investment of more than 1 billion us dollars."

whether users like it or not is determined by data

it is this series of technical layouts that has rapidly increased the penetration rate of ideal auto's autonomous driving technology among users, and significantly improved the penetration rate of urban noa models. user reviews and daily average activity also verify users' recognition of ideal's autonomous driving technology.

on july 15, ideal auto officially launched noa without map. it no longer relies on high-precision maps and other prior information, and can be used on all navigable urban roads across the country. according to the ideal auto intelligent driving satisfaction survey, users gave ideal intelligent driving a high rating of 9 out of 10 in the survey. after ideal stores nationwide launched noa without map test drive, the store noa test drive rate doubled, and the sales of ad max models above 300,000 yuan accounted for nearly 70%.

in less than two months from the launch in july to the end of august, ideal auto's urban noa daily average mileage increased by 3 times, and the urban noa daily average activity increased by 8 times, which shows that ideal smart driving has crossed the gap faced by new technologies and entered the late mass stage. the growth of ideal auto's user volume, user daily average mileage, and user daily activity has driven the rapid growth of the total accumulated mileage of smart driving.

from 2022 to the end of august 2024, the total cumulative mileage of intelligent driving will achieve a leap-forward growth from 400 million kilometers to 2.2 billion kilometers, and is expected to reach 3 billion kilometers by the end of this year, which is expected to rank among the leading levels globally.

the industry's first "parking space to parking space" product was launched, opening a new era of one-click intelligent driving

at the end of july, ideal auto's new intelligent driving technology architecture based on end-to-end model, vlm visual language model and world model was successfully implemented on the vehicle side and started internal testing with 1,000 people. the number of users in the 1,000-person group reached 1,029, covering 270 cities. in less than a month of internal testing, the total urban noa mileage of users in the 1,000-person group reached 211,000 kilometers, the longest single-day urban noa driving mileage was 391 kilometers, and the longest single zero-takeover urban noa mileage was 81.6 kilometers.

after the great success of the 1,000-person group, ideal auto started recruiting 10,000-person experience groups for end-to-end models + vlm visual language models at the end of august. on september 14, the ideal auto end-to-end + vlm 10,000-person experience group was officially launched, which will continue to accelerate the iteration and evolution of ideal's intelligent driving technology, and the safety will also be greatly improved, providing users with a smarter, safer and more convenient car experience.

on september 14, ideal auto also officially pushed the "parking space to parking space" one-click smart driving to a group of thousands of people, marking the industry's first "parking space to parking space" product launch. currently, the parking space to parking space function only supports underground garage scenarios, and requires completion of a learning process. functions include: one-click activation of navigation and functions (no need to set up navigation additionally, go directly to the destination parking space), direct activation of starting in p gear in vertical parking spaces in underground garages, parking in parking spaces, travel within the park, travel within the underground garage (including spiral passages), entering/exiting the park on public roads, identifying gates and start-stop and other user-frequent vehicle usage scenarios.

in 2023, the market size of china's autonomous driving system will be 330.1 billion yuan, a year-on-year increase of 16.09%, maintaining a high-speed development trend. at present, china's autonomous driving technology is in a critical period of rapid development and profound changes, and high-level intelligent driving has become the focus of industry competition. ideal auto's industry-first "parking space to parking space" product has a huge impact on the field of high-level intelligent driving. it not only indicates that ideal's end-to-end + vlm intelligent driving technology is being implemented for all users, but also guides the industry to choose to develop better technical frameworks and underlying algorithm models, injecting new vitality into the future of autonomous driving.

(china news network)