--- title: "End-to-end autonomous driving: Who's all in and who's on the fence" type: "Topics" locale: "zh-HK" url: "https://longbridge.com/zh-HK/topics/22260282.md" description: "'End-to-end' is not a panacea. At first glance, this view from Bai Yuli, head of Nio's AI platform and senior R&D director, may easily lead outsiders to misunderstand that Nio is wavering in its commitment to the end-to-end approach. In fact, this is a reaffirmation of Nio's strategy. Nio plans to model the code for planning and control before building a more integrated 'end-to-end' framework. Currently, the 'end-to-end' approaches of XPeng, Li Auto, and Nio are all similar—'modular' to 'disassembled and reassembled' to 'end-to-end' framework. On June 8, at the 2024 China Auto Chongqing Forum..." datetime: "2024-07-05T02:39:03.000Z" locales: - [en](https://longbridge.com/en/topics/22260282.md) - [zh-CN](https://longbridge.com/zh-CN/topics/22260282.md) - [zh-HK](https://longbridge.com/zh-HK/topics/22260282.md) author: "[汽车之心](https://longbridge.com/zh-HK/profiles/3726156.md)" --- > 支持的語言: [English](https://longbridge.com/en/topics/22260282.md) | [简体中文](https://longbridge.com/zh-CN/topics/22260282.md) # End-to-end autonomous driving: Who's all in and who's on the fence "End-to-end is not a panacea." At first glance, this view from Bai Yuli, head of Nio's AI platform and senior R&D director, could easily lead to the misunderstanding that Nio is wavering in its commitment to the end-to-end approach. In reality, this is a reaffirmation of Nio's strategy. Nio plans to model the code for planning and control before integrating it into a more comprehensive "end-to-end" framework. Currently, the "end-to-end" approaches of XPeng, Li Auto, and Nio follow a similar pattern: **"modular" → "disassembled and reassembled" → "end-to-end" framework**. On June 8, at the 2024 China Auto Chongqing Forum, Li Xiang, chairman and CEO of Li Auto, shared new insights on autonomous driving technology: "**End-to-end + VLM** (Visual Language Model) **\+ generative verification systems will form the most critical technical architecture and system for future physical-world robotics**." Li Xiang believes that relying solely on end-to-end solutions cannot address corner cases; instead, capabilities must be enhanced. Using VLM, vehicles can react promptly to intersections, traffic lights, etc. As the first domestic automaker to implement "end-to-end" in vehicles, XPeng announced on May 20 that its end-to-end framework consists of three components: the neural network XNet (focused on perception and semantics), the planning and control model XPlanner, and the large language model XBrain (focused on scene cognition). The strategies of XPeng, Li Auto, and Nio **differ from Tesla's approach of relying entirely on neural networks for end-to-end solutions—neural networks are just one part of the system**. In fact, few in China's autonomous driving industry fully understand how Tesla achieved this. "No one dares to claim that end-to-end is entirely neural networks," said He Xiaopeng during a media interview after the "end-to-end" launch. "It’s completed within a system, just like how brakes operate within a rule-based framework. Our advantage lies in refining the algorithm sandbox for brake controllers." Wu Xinzhu, vice president of Nvidia's automotive division, believes **end-to-end is the final chapter in the trilogy of autonomous driving**. Facing this decisive battle, Tesla initiated commercialization of its end-to-end framework in February, prompting several Chinese automakers to set timelines for their own "end-to-end" deployments. Halfway through 2024, will this year mark the "first year" of end-to-end adoption? From XPeng's pioneering implementation to Tesla's "classical" approach, how should domestic automakers proceed? Mapping a complete end-to-end framework from academia to industry might reveal their positions. _**01.**_ **After XPeng, Who Will Scale End-to-End Autonomous Driving Next?** In August 2023, Tesla's FSD V12 became the first mass-produced "end-to-end" framework. In February 2024, Tesla rolled out FSD V12 to some users, marking its commercial debut. FSD V12's smooth and impressive performance quickly stood out. In May 2024, XPeng announced its "end-to-end" framework was vehicle-ready. Broadly, "end-to-end" adoption spans three groups: automakers, autonomous driving firms, and academic institutions. Some academic and industrial efforts predate Tesla. **Automakers like Nio, Li Auto, XPeng, Xiaomi, Jiyue, IM Motors, GAC, Great Wall, and ZEEKR are among China's first movers**. Recently, Nio established a dedicated division for end-to-end model R&D, merging its perception and planning teams. Post-restructuring, Nio's autonomous driving efforts are split into "cloud" (large models) and "vehicle" (deployment architecture), abandoning the previous functional silos. The "cloud" team develops foundational models to support future vehicle-side iterations. **"Cloud" signifies potential breakthroughs in computational bottlenecks**. Nio's edge-computing capabilities now deliver 287.1 EFLOPS, **equivalent to 100 distributed training clusters**, "roughly matching Tesla's 100,000 H100 chips." Nio adopts a gradual "end-to-end" approach. Ren Shaoqing, Nio's VP of autonomous driving, notes that end-to-end requires fully modelized functional modules and **robust engineering support**: "Without rapid training and validation, it’s useless—or even toxic." In late 2023, Li Auto formed a dedicated end-to-end team under its algorithm R&D division, targeting supervised L3 autonomy by late 2024 or early 2025. Li Auto's framework combines **end-to-end + VLM + generative verification**. XPeng claims its end-to-end model will enable 30x capability gains in 18 months, with iterations every two days. Li Liyun, XPeng's autonomous driving lead, clarified: "Our AI framework integrates XBrain, XNet, and XPlanner—interconnected yet specialized." **This modular approach underpins XPeng's mass-production strategy**. Other notable moves: - Xiaomi's December 2023 car launch featured its self-developed "end-to-end" model, touted as a **global first in mass production**. Ex-TuSimple CTO Wang Naiyan joined Xiaomi, cautioning against narrow interpretations of "end-to-end." - Jiyue CEO Xia Yiping: "**End-to-end is our next R&D priority**." - IM Motors co-CEO Liu Tao: "We're **fully committed to end-to-end deployment** for human-like driving." Partnering with Momenta. - GAC R&D: "Exploring end-to-end solutions with **preliminary success**." - Great Wall (Haomo): Its 2023 "Snow Lake·HaiRuo" model focuses on decision-making, with end-to-end as the ultimate goal. ZEEKR remains cautious, treating end-to-end as pre-research due to data and safety concerns. Most automakers, like ZEEKR, acknowledge end-to-end as inevitable. Autonomous driving suppliers—Huawei, Momenta, DeepRoute, and Sensetime—plan end-to-end deployments for 2024-2025. Academically, Shanghai AI Lab's UniAD (CVPR 2023 best paper), Huazhong University of Science and Technology, and Wayve lead in research. Credit: Gongjin Lan & Qi Hao's paper tracks 2022-2023 industry/academic end-to-end projects. End-to-end marks a technological culmination but remains nascent. _**02.**_ **Post-Models: Will End-to-End Split into High/Low Tiers?** Diverging approaches are emerging. Tesla's FSD V12, praised for neural network performance, eliminated 300K lines of C++ rules. Shanghai AI Lab's UniAD researcher Li Hongyang was inspired by Openpilot's low-cost efficacy, realizing "autonomous driving could be this simple." If Tesla represents **high-end**, Openpilot-like solutions are **low-end**—different but illustrative. Two evaluation methods exist: closed-loop (feedback-enabled) and open-loop (modular comparisons). UniAD is open-loop validated. **Which end-to-end performs best?** He Xiaopeng cites disengagement rates: 1,000 km/highway vs. 10-100 km/urban. Tesla's FSD V12 improved from 267 km to 537 km. **End-to-end relies on BEV+Transformer**, combining 2014's bird's-eye-view with 2017's Transformer. Tesla merged them in 2020-2022, enabling neural "end-to-end" optimization. This **perception→prediction→planning→decision** shift replaces rules with neural networks—Tesla's "classical" approach. Skepticism persists. Momenta's 2016 critique cited inefficiency, inflexibility, and opacity—issues still partly unresolved. **03.** **AI-Driven Revolution: Is Compute the Key to End-to-End?** Tesla's FSD V12 and Sora reignited end-to-end discussions in 2024. Sora's video-generation capability—vital for end-to-end training—validates the approach. Musk boasts Tesla's "world-best real-world simulation." Yet he admits: "**FSD training is compute-bound**. We’ll use non-car video data when spare compute exists." Compute scarcity shapes AI progress. Domestic players must scale up. Alibaba's 2022 Zhangbei Super Compute Center (12 EFLOPS) outpaced Google (9 EFLOPS) and Tesla (1.8 EFLOPS). XPeng's 600 PFLOPS "Fury" cluster accelerated training 170x. Current compute comparisons: - Tesla: 10 EFLOPS (2023), targeting 100 EFLOPS by October 2024. - Li Auto: 2.4 EFLOPS (June 2024). - Nio: 1.4 EFLOPS (September 2023). - Sensetime: 12 EFLOPS (18 EFLOPS by 2024-end)—China's largest. Tesla's planned 100 EFLOPS (≈300K A100 GPUs) dwarfs XPeng's 600 PFLOPS (≈30K A100s). Most Chinese firms operate at thousand-GPU scale. Haomo claims 2,000-5,000 GPUs suffice for nationwide deployment. Despite gaps, domestic compute could near Tesla's 1-2-year-old levels by 2024-end. Jiyue's Xia Yiping asserts: "While Tesla leads, our localization surpasses theirs in China." As urban NOA and end-to-end converge, the competition intensifies. ### 相關股票 - [NIO Inc. USD OV (NIO.SG)](https://longbridge.com/zh-HK/quote/NIO.SG.md) - [NIO-SW (09866.HK)](https://longbridge.com/zh-HK/quote/09866.HK.md) - [NIO Inc (NIO.US)](https://longbridge.com/zh-HK/quote/NIO.US.md) - [Systems Software (IN00356.US)](https://longbridge.com/zh-HK/quote/IN00356.US.md) - [XPENG-W (09868.HK)](https://longbridge.com/zh-HK/quote/09868.HK.md) - [Global X Autonomous & Electric V (DRIV.US)](https://longbridge.com/zh-HK/quote/DRIV.US.md) - [iShares Self-Driving EV&Tech ETF (IDRV.US)](https://longbridge.com/zh-HK/quote/IDRV.US.md) - [XPeng (XPEV.US)](https://longbridge.com/zh-HK/quote/XPEV.US.md)