Li Auto recently set up an end-to-end large model autonomous driving team of slightly over 200 people, according to local media.
Li Auto (NASDAQ: LI) has reportedly set up a dedicated team to bet on end-to-end AI (artificial intelligence) technology, becoming the latest to do so after Nio (NYSE: NIO).
Li Auto recently set up an end-to-end large-model autonomous driving team of just over 200 people, with some members of other teams providing flexible support, local media outlet 36kr reported today.
The car company isn't alone in doing so, as Nio has previously led the way in making organizational changes by setting up a department dedicated to end-to-end large models, the report noted.
Li Auto's smart driving team is divided into two main groups: algorithm development and mass production development, with a team size of about 800 people, the report said.
Li Xiang, the company's founder, chairman, and CEO, has said that by the end of this year or early next year, Li Auto will launch an end-to-end +VLM (Vision-Language Model) autonomous driving solution that will be trained by over 10 million clips.
Tesla (NASDAQ: TSLA) released FSD V12 earlier this year, which brought good results. This has led to an industry consensus on the end-to-end large AI model used by Tesla, and more car companies in China are starting to experiment with this route.
Nio's smart driving R&D department completed a team restructuring to focus more on end-to-end technology, according to a June 19 report in local media outlet LatePost.
Under the original commonly adopted technology architecture, for smart driving systems, modules such as perception, prediction, decision-making, and control all require engineers in different fields.
The end-to-end smart driving system uses sensor data as input and is used directly in the vehicle's control commands, with all intermediate processes relying on neural network models.
On July 11, Nio started rolling out the Banyan 2.6.5 CN system for vehicles based on the NT 2.0 technology platform, bringing AEB (automatic emergency braking) functionality optimized using end-to-end technology.
It is worth noting that end-to-end technology is demanding in terms of computing power resources.
Li Auto is scrambling to get more training computing power after choosing the end-to-end solution, 36kr's report today said, adding that an insider said Li Auto believes the next point of competition for smart driving is supercomputing centers.
Li Auto purchased more than 300 Nvidia servers last year from Volcano Engine, a cloud service platform owned by ByteDance, and is also currently working with cloud service vendors including Alibaba Cloud and Baidu Cloud, according to the report.
Huawei currently has the largest smart driving training computing power in China, mentioning in June that its computing power would reach 3.5 EFLOPS, the report noted.
Xpeng's (NYSE: XPEV), Nio's, and Li Auto's smart driving centers have computing powers of 0.6 EFLOPS, 1.4 EFLOPS, and 1.4 EFLOPS, respectively, the report said.
Nio reportedly reshuffles smart driving team to focus more on end-to-end tech
據當地媒體報道,理想汽車最近成立了一個由略超過200人組成的端到端大型自動駕駛團隊。
據報道,理想汽車(納斯達克股票代碼:LI)已經成立了一個專門的團隊來押注端到端的人工智能(人工智能)技術,成爲繼蔚來(紐約證券交易所代碼:NIO)之後最新這樣做的公司。
據當地媒體36kr今天報道,理想汽車最近成立了一個由200多人組成的端到端大型自動駕駛團隊,其他團隊的一些成員提供靈活支持。
報告指出,這家汽車公司並不是唯一一家這樣做的公司,因爲蔚來此前曾通過成立一個專門負責端到端大型車型的部門,在組織變革方面處於領先地位。
報告稱,Li Auto的智能駕駛團隊分爲兩個主要小組:算法開發和量產開發,團隊規模約爲800人。
公司創始人、董事長兼首席執行官李翔表示,到今年年底或明年初,理想汽車將推出端到端+vLM(視覺語言模型)自動駕駛解決方案,該解決方案將接受超過1000萬個片段的訓練。
特斯拉(納斯達克股票代碼:TSLA)在今年早些時候發佈了FSD V12,帶來了不錯的業績。這使業界對特斯拉使用的端到端大型人工智能模型達成了共識,越來越多的中國汽車公司開始嘗試這種路線。
根據當地媒體LatePost6月19日的報道,蔚來的智能駕駛研發部門完成了團隊重組,將重點更多地放在端到端技術上。
在最初的常用技術架構下,對於智能駕駛系統,感知、預測、決策和控制等模塊都需要不同領域的工程師。
端到端智能駕駛系統使用傳感器數據作爲輸入,直接用於車輛的控制命令,所有中間過程都依賴於神經網絡模型。
7月11日,蔚來開始爲基於Nt 2.0技術平台的車輛推出Banyan 2.6.5 CN系統,該系統採用端到端技術進行了優化的aEB(自動緊急制動)功能。
值得注意的是,端到端技術對計算能力資源的要求很高。
36kr今天的報告稱,理想汽車在選擇端到端解決方案後正爭先恐後地獲得更多的訓練計算能力,並補充說,一位內部人士說,理想汽車認爲智能駕駛的下一個競爭點是超級計算中心。
報告稱,理想汽車去年從字節跳動旗下的雲服務平台Volcano Engine購買了300多臺Nvidia服務器,目前還與包括阿里雲和百度雲在內的雲服務供應商合作。
報告指出,華爲目前擁有中國最大的智能駕駛訓練計算能力,在6月份提到其計算能力將達到3.5 EFLOPS。
報告稱,小鵬汽車(紐約證券交易所代碼:XPEV)、蔚來汽車和李汽車的智能駕駛中心的計算能力分別爲0.6 EFLOPS、1.4 EFLOPS和1.4 EFLOPS。
據報道,蔚來重組了智能駕駛團隊,將重點更多地放在端到端技術上