Waymo CEO回应L2升维L4：有可能，但只靠端到端还不够

量子位

量子位2026年5月14日

Waymo CEO on L2 to L4 Transition: Possible, But End-to-End Alone Isn't Enough

8.7Score

TL;DR · AI Summary

Waymo CEO Dmitri Dolgov confirms L2 can evolve into L4, but pure end-to-end is insufficient; a hybrid architecture with world models, simulation, and evaluation is required, while hardware costs have dropped to ADAS levels.

Key Takeaways

Waymo uses a cloud-based multimodal world action language model distilled to veh
The sixth-gen autonomous kit cost is now comparable to ADAS systems, supporting
Waymo handles 500k orders per week, operates in 11 cities, and plans expansion t

Outline

Jump quickly between sections.

§Introduction: Can L2 Evolve to L4?
Waymo co-CEO Dmitri Dolgov addresses the possibility of upgrading from L2 to L4, emphasizing fundamental technical differences.
·End-to-End Is Not the Only Solution
Pure end-to-end black-box systems hinder efficient simulation and evaluation, requiring structured semantics for better interpretability and training efficiency.
·Three Core Tasks of the Cloud Base Model
The multimodal world action language model performs driving instruction, simulation generation, and decision assessment.
·Hardware Cost Reduction Achieved
The multi-sensor fusion system now costs as little as ADAS kits, with LiDAR proving critical in weak signal detection.
·Commercial Acceleration and Global Expansion
Waymo has surpassed 20 million cumulative orders, processes 500k weekly, operates a 3,000-vehicle fleet, and is expanding to London and Tokyo.
·New Business Direction: Deployment on Private Vehicles
Autonomous driving services may soon be available on private cars, not just via Robotaxi hailing.

Mindmap

See how the topics connect at a glance.

查看大纲文本（无障碍 / 无 JS 友好）

Waymo L4技术与商业化路径
- 技术架构
  - 端到端+结构化语义
  - 云端世界模型
  - 蒸馏至车端小模型
- 硬件进展
  - 第六代多传感器套件
  - 激光雷达关键作用
  - 成本降至ADAS水平
- 商业化布局
  - 周订单50万单
  - 落地11城，出海英日
  - 私家车部署可能性

Highlights

Key sentences worth saving and sharing.

Pure end-to-end is like a black box, making simulation training difficult.
— Paragraph 4
⬇︎ 下载 PNG 𝕏 分享到 X
The cost of the sixth-generation kit is no longer high—it's comparable to that of ADAS systems.
— Paragraph 8
⬇︎ 下载 PNG 𝕏 分享到 X
2026 will become the 'Global Robotaxi Year'.
— Paragraph 10
⬇︎ 下载 PNG 𝕏 分享到 X
Moving from L2 to L4 is a qualitative leap, not a linear progression.
— Paragraph 5
⬇︎ 下载 PNG 𝕏 分享到 X
Based on a faint LiDAR signal, the system detected a pedestrian and proactively avoided them.
— Paragraph 7
⬇︎ 下载 PNG 𝕏 分享到 X
In the future, you won’t need to hail a ride to experience autonomous driving—your own car could offer it.
— Paragraph 11
⬇︎ 下载 PNG 𝕏 分享到 X

#Waymo#Autonomous Driving#End-to-End#World Model#Robotaxi

Open original article

Waymo CEO回应L2升维L4：有可能，但只靠端到端还不够 – Quantum Bit

[](http://www.qbitai.com/)

[](javascript:void(0))

扫码关注量子位

[](https://weibo.com/qbitai?is_all=1)

< img id="wx_img" src="https://www.qbitai.com/wp-content/uploads/imgs/qbitai-logo-1.png" width="400" height="400">

Waymo CEO回应L2升维L4：有可能，但只靠端到端还不够

Image 7: 一凡 _[一凡](https://www.qbitai.com/author/yifan "由一凡发布")_ 2026-05-14 10:12:22 来源：量子位

云端基模蒸馏上车，用语言对齐世界模型

一凡发自副驾寺forever
Autonomous Vehicles Reference | Official WeChat Account AI4Auto

Where are we with autonomous driving?

It depends on how far Waymo, the global Robotaxi leader, has gone:

Technologically, it's moving towards end-to-end, with a cloud-based multimodal world behavior language model as the foundation, distilled into smaller models for on-board use.

Commercially, they're running 500,000 rides per week in the US, deploying in 11 cities and launching in London and Tokyo, marking the start of the "Robotaxi Globalization Year."

These are recent updates from Waymo's Vice President of Engineering, Dmitri Dolgov.

Dolgov is a veteran in the industry, an engineer who has witnessed the entire journey from the Desert Challenge to global deployment over the past 21 years.

In his interview, he discussed the evolution of autonomous driving technology over the past 21 years and also looked ahead to how society will change when autonomous driving becomes widespread.

End-to-end is necessary but not sufficient

Dolgov revealed that Waymo is also moving towards end-to-end, but this does not mean abandoning the previous technical paradigms. He believes this is a false dichotomy.

Because for Waymo, pure end-to-end is like a black box, where sensors input images, and the system uses image pixels to directly output vehicle trajectories. This makes simulation training difficult.

This forces the simulation system to generate a complete pixel world, which is already challenging to create, let alone training and evaluation processes which are inefficient.

Waymo's solution is to build a base model on the cloud, which Dolgov calls the Multimodal World Action Language Model (MVLAM). It still follows the architecture of a world model, but the linguistic dimension is introduced through VLM, incorporating general world knowledge and semantic understanding.

This base model isn't solely used for driving; it's primarily adapted for these three categories of tasks:

Driving: Cloud teacher model learns how to drive, then distills to the student model on board.
Simulation: Generates realistic virtual environments for autonomous systems.
Evaluation: Evaluates driving decisions and teaches drivers how to operate.

To Dolgov, if you only do ADAS, you can rely on the driver model to solve 90% of the problems in autonomous driving systems.

However, to leap over the gap to Level 4, you need to use simulators and value judgmenters.

This shows that Dolgov acknowledges that Level 2 players could enter the Level 4 camp. But he doesn't believe that Level 2 and Level 4 are linearly related. The technologies they address have fundamental differences, and the transition from Level 2 to Level 4 is qualitative, not quantitative.

Waymo's sixth-generation suite costs comparable to ADAS systems

Apart from the software architecture, Dolgov shared some hardware progress at Waymo.

He stated that Waymo's sixth-generation autonomous driving suite still employs multi-sensor redundancy perception, with laser radars, millimeter-wave radars, and cameras all onboard, and they've been unified and simplified.

Dolgov believes that while laser radars still offer advantages over pure vision, such as detecting weak signals like pedestrians hidden by buses.

For example, once Waymo's autonomous vehicle was about to start, a bus passed in front, blocking its view. However, using the laser radar, the system detected a weak signal on the other side of the bus, which turned out to be a pedestrian walking.

Based on this weak signal, the autonomous system judged there was a pedestrian and predicted they would walk around the bus. Thus, it slowed down and turned slightly to the side, creating space to avoid the pedestrian.

While still using multiple sensors, Dolgov revealed that the cost of the sixth-generation suite is not high and can now compete with smart-assistive driving suites.

This is surprising because in China, the cost of laser radars can reach "thousand-dollar" levels due to mature automotive electronics supply chains and economies of scale.

In the US, Tesla's commitment to pure vision has limited the production of laser radars, yet Waymo managed to bring the cost down to current levels.

Dolgov also mentioned some unique challenges with sensors, such as cleaning, heating, and controlling wet and slippery roads in cold weather.

New technologies and new suites are helping Waymo accelerate autonomous driving.

The acceleration of autonomous driving

In his interview, Dolgov disclosed several data points. He stated that Waymo's cumulative orders have surpassed 20 million rides, with half coming from the last seven months.

They're currently growing at a rate of 50,000 rides per week, with a fleet size of 3,000 vehicles, and weekly autonomous driving mileage reaching 6.4 million kilometers.

This pace is expected to continue, with Dolgov noting that Waymo has already deployed in 11 cities in the US and plans to launch in London and Tokyo this year. This means 2026 could be the "Robotaxi Globalization Year."

However, Dolgov acknowledged that deploying Robotaxis overseas is challenging. Although the model's generalization capability is improving, adapting to new cities still requires customization, data collection, and optimization, followed by validation.

Besides technical issues, there's also communication with regulatory bodies to gain local trust.

One More Thing

When discussing commercial progress, Dolgov also touched upon the "corner cases" in autonomous driving commercialization.

He noted that in remote areas with few people taking taxis, deploying Robotaxis wouldn't make sense since they wouldn't get many rides and wouldn't be profitable.

But this doesn't mean that remote regions can't adopt autonomous driving.

Dolgov revealed that Waymo's system can also be deployed on private cars in the future.

In other words, in the future, it might not be necessary to hail a ride to experience autonomous services; owning a car could suffice.

As Waymo moves from Level 2 to offering Robotaxi ride-hailing services, the company is considering entering the private car market.

From this perspective, the business models of Level 2 and Level 4 are converging.

Waymo

[](https://www.qbitai.com/2026/05/417469.html#)

![Image 13: 一凡](https://www.qbitai.com/2026/05/417469.html#)[一凡](https://www.qbitai.com/author/yifan "由一凡发布")

[燃油SUV车主熬出头了！华为乾崑智驾加持，全新奥迪Q5L率先实现智能化](https://www.qbitai.com/2026/04/406960.html "燃油SUV车主熬出头了！华为乾崑智驾加持，全新奥迪Q5L率先实现智能化")_2026-04-25_
[做了10年Robotaxi，小马智行首次入局RoboVan](https://www.qbitai.com/2026/04/409648.html "做了10年Robotaxi，小马智行首次入局RoboVan")_2026-04-28_
[腾讯智慧出行：单纯大模型上车无意义，要落地场景智能体](https://www.qbitai.com/2026/04/409550.html "腾讯智慧出行：单纯大模型上车无意义，要落地场景智能体")_2026-04-28_
[Momenta曹旭东：规模L4要百亿美元投入，现金流业务是物理AI门票](https://www.qbitai.com/2026/04/407485.html "Momenta曹旭东：规模L4要百亿美元投入，现金流业务是物理AI门票")_2026-04-26_

扫码分享至朋友圈![Image 14: Scan me!](blob:http://localhost/7de33e8630cd36d089023f6d33ac9e1e)

[](https://service.weibo.com/share/share.php?url=https://www.qbitai.com/2026/05/417469.html&title=Waymo%20CEO%E5%9B%9E%E5%BA%94L2%E5%8D%87%E7%BB%B4L4%EF%BC%9A%E6%9C%89%E5%8F%AF%E8%83%BD%EF%BC%8C%E4%BD%86%E5%8F%AA%E9%9D%A0%E7%AB%AF%E5%88%B0%E7%AB%AF%E8%BF%98%E4%B8%8D%E5%A4%9F&appkey=4017757111&searchPic=true&ralateUid=6105753431 "分享到新浪微博")[](https://www.qbitai.com/2026/05/417469.html)

Hot Articles

[](https://www.qbitai.com/wp-content/uploads/2026/05/d8ffe2484658e11e004c471083a7da01.webp) #### 00后下场整顿Agent：啥都不学就能用好AI，这才是正确打开方式 2026-05-07

[](https://www.qbitai.com/wp-content/uploads/2026/05/811aee789e849c6f9d434975400917d9.png) #### 两项AI政策发布，范式智能战略布局与产业方向高度契合 2026-05-09

[](https://www.qbitai.com/wp-content/uploads/2026/05/74d2cfd2cfa84f89e3ebc6dd19c6de17.jpeg) #### 太初元碁携龙虾一体机亮相北京科博会 2026-05-09

[](https://www.qbitai.com/wp-content/uploads/2026/05/c207c5513bcee7794c5c29707a3f19cf.png) #### 阶跃最新语音模型位列 Artificial Analysis 评测榜中国第一 2026-05-09

[](https://www.qbitai.com/wp-content/uploads/2026/05/eb4c11f6178ac1452acfe0054b9e313d.png) #### 不更新参数就能强化学习！OpenAI翁家翌提出新范式：决策只需AI手搓一个.py 文件 2026-05-09

[](https://www.qbitai.com/2026/05/417469.html#)追踪人工智能新趋势，报道科技行业新突破

Waymo CEO on L2 to L4 Transition: Possible, But End-to-End Alone Isn't Enough

TL;DR · AI Summary

Key Takeaways

Outline

Mindmap

Highlights

Waymo CEO回应L2升维L4：有可能，但只靠端到端还不够 – Quantum Bit

Waymo CEO回应L2升维L4：有可能，但只靠端到端还不够

相关阅读

Hot Articles