Waymo Scales Autonomous Fleet to 11 Cities as Rides Hit 20 Million

Waymo自动驾驶车队扩展至11城，出行服务突破2000万次 Waymo正加速推进全球同步商业化，其混合基础模型使自动驾驶系统安全性达到人类的13倍，每八天即可预防一起严重伤害事故。 Dmitri Dolgov分享了Waymo的发展历程：从2009年源于Google的Project Chauffeur项目，到当前在美国11座城市加速推进全球商业化，并即将进军伦敦和东京等国际市场。他阐述了Waymo如何通过将安全视为不容妥协的使命来规避行业炒作周期的陷阱，从而实现指数级规模扩张，完成超过2000万次自动驾驶行程。此外，Dolgov还详细介绍了多模态Waymo Foundation Model的技术架构，解释了为何结合端到端学习与结构化中间表征的混合方案对于实现超越人类水平的驾驶安全至关重要。 Waymo 早期源于 2009 年谷歌的 Project Chauffeur 项目，一支约 12 人的团队历经 18 个月的高强度研发，成功完成了 10 万英里全自动驾驶以及旧金山湾区 10 条 100 英里路线零人工接管的里程碑。 (3:51) Waymo 部署了多模态“世界-动作-语言”基座模型，协同驱动实际车辆、物理仿真世界和行为评估器，并通过端到端结合结构化中间表征的混合架构来确保运行时的安全校验。 (9:47) 商业化落地步入指数级扩展阶段：Waymo 达成首个 1 亿自动驾驶英里耗时约 16 年，而增至 2 亿英里仅用时 6 个月；累计超过 2000 万次全自动载客中，有 1000 万次是在最近 7 个月内完成的。 (17:00) 依托超 1.7 亿英里实际运营数据，Waymo 在致伤事故安全性上达人类驾驶员的 13 倍，相当于每 8 天避免一起严重人身伤害，目前已在美国 11 座城市提供服务并计划拓展至伦敦与东京。 (19:56)

According to Waymo co-CEO Dmitri Dolgov, the autonomous vehicle developer has shifted from long-term technical de-risking to rapid global commercialization, driven by exponential operational scaling and a multimodal foundation model. Dolgov contends that reaching superhuman safety across real-world edge cases requires a hybrid AI architecture that combines end-to-end learning with verifiable intermediate representations. Hybrid AI Architecture and Safety Verification (9:47) At the core of Waymo's software is a multimodal world-action-language foundation model that powers three interconnected systems: the on-road driver agent, a physical-world simulator, and an evaluative critic. The model integrates inputs from cameras, lidars, and radars alongside visual language models (VLMs) to parse complex social driving dynamics and semantic road scenes. While the model maps raw sensory data directly to driving decisions, Dolgov rejects a purely vanilla end-to-end framework as insufficient for safe, commercial operation. Instead, Waymo augments learned representations with structured, materialized intermediate representations. This hybrid setup enables runtime safety validation, closed-loop evaluation, reinforcement learning reward modeling, and the direct integration of human feedback from remote support operators. Sixth-Generation Hardware and Rider Experience (15:28) Waymo's sixth-generation Driver marks a strategic transition from proving technical feasibility toward cost reduction, system simplification, and high-volume manufacturing readiness. The updated hardware suite is deployed on a custom platform manufactured by Zeekr, designed primarily around passenger utility rather than traditional vehicle ergonomics. Dolgov reports that while the vehicle shares an exterior footprint similar to Waymo's Jaguar I-PACE fleet, the interior provides an expanded cabin featuring sliding passenger doors, extended legroom, and redesigned passenger screens. Currently operational for Waymo employees, the platform is scheduled to roll out to public riders later in the year. Empirical Safety Record and Emergent Perception (19:56) Dolgov frames Waymo's core mission around reducing traffic fatalities, pointing to global figures of 1.19 million annual road deaths—equivalent to one death every 26 seconds. Operating at an active rate of over 4 million autonomous miles per week, Waymo reports that data from more than 170 million autonomous miles shows its driver is more than 13 times safer than human drivers regarding serious injury collisions, preventing an injury every eight days. Dolgov attributes this margin partly to emergent perception capabilities that exceed human sensory limits. In one documented urban incident, a vehicle proactively yielded to an unseen pedestrian behind a bus. System logs revealed the vehicle's lidar had bounced signals underneath the bus chassis, detecting the movement of the pedestrian's feet and accurately predicting their crossing path before they entered direct line of sight. Exponential Scaling and Global Commercial Expansion (17:00) Waymo's operations have accelerated sharply from its 2009 origins as Google's Project Chauffeur, which took 18 months for a roughly dozen-person team to complete 100,000 miles across 10 Bay Area routes. In contrast, while accumulating the company's first 100 million autonomous miles required roughly 16 years, the subsequent 100 million miles took only six months. Out of more than 20 million completed autonomous passenger trips, 10 million occurred in the last seven months alone. Dolgov states that Waymo has advanced from sequential market rollouts to rapid parallel deployment, highlighted by launching four new municipal services in a single day compared to the eight years required for its first four cities. Waymo currently operates across 11 U.S. cities and has announced international expansion targets in London and Tokyo. Waymo's shift to rapid commercialization rests on the claim that its proprietary model architecture and custom hardware have solved the long-tail edge cases of autonomous navigation. The company's expansion into international markets like London and Tokyo will test whether this software foundation and operational model can transfer seamlessly to complex foreign regulatory environments and street topologies.

Waymo联合首席执行官Dmitri Dolgov指出，自动驾驶行业已度过早期的技术试错与验证周期，进入由指数级里程增长驱动的“并行全球商业化”阶段。他强调，要达成超越人类水平的安全性，必须将多模态端到端世界模型与结构化中间表征相结合，并始终以可量化的安全数据作为规模扩张的绝对基石。 算法路径：端到端模型与结构化验证融合 (9:47) Dolgov介绍了Waymo的核心AI架构——涵盖“驾驶员”（Driver）、“模拟器”（Simulator）与“评估者”（Critic）三大支柱的Waymo多模态世界-动作-语言基础模型。该架构不仅融合了视觉语言模型（VLM）对现实语义和社交驾驶场景的常识理解，还直接处理来自摄像头、激光雷达与毫米波雷达的多模态感知数据。Dolgov确认Waymo的基础模型采用端到端方式将传感器输入直接映射到驾驶动作，从而避免了感知与规划之间硬编码接口的脆弱性；但他同时警告，仅依赖“纯粹的原始端到端”（vanilla end-to-end）架构不足以支撑大规模商业化落地所需的极端安全性。 为了在数亿英里的运行中保持超人类级别的可靠性，Waymo在端到端深度学习中引入了“结构化具象化中间表征”。这种混合路径赋予系统关键的安全机制，包括运行时的实时动作安全校验、闭环仿真测试、强化学习的高阶奖励函数设定，以及远程协助人员的人工反馈接入。Dolgov分享了一次涌现感知能力的实证案例：在旧金山一处路口，Waymo车辆通过激光雷达从一辆受阻巴士底部反弹的微弱信号，成功捕捉到底部露出的行人脚步动态，并提前推断其行动轨迹实施避让，展现了系统超越单一传感器物理视线的防御性感知能力。 长期主义安全观与量化数据验证 (19:56) 回顾2016至2017年的资本狂热与随后的行业出清，Dolgov认为AI算法在早期的突破往往容易掩盖自动驾驶长尾问题的极端复杂性，进而产生“短期内即可彻底解决”的行业错觉。不同于硅谷常见的“快速行动、打破常规”文化，Waymo将解决全球每年119万人因车祸丧生（平均每26秒夺走一条生命）的现状作为底线使命，拒绝追求所谓的速效捷径，选择耐受攻克自动驾驶“下一个9”（超高可靠性）所必需的漫长研发周期。 根据Waymo披露的数据，公司目前的自动驾驶行驶里程已超过每周400万英里。在累计逾1.7亿英里的全自动驾驶数据中，Waymo在运营城市涉及严重人身伤害的碰撞事故率比人类驾驶员低13倍以上，按当前运营体量计算，相当于每8天避免一起严重伤害事故。Dolgov举例说明，一辆Waymo在面对前方失控倒地的电动滑板车骑行者时，展现出了超越人类极限的紧急避障与制动反应，确保了当事人的安全。 第六代硬件优化与乘客体验平台 (15:28) 在硬件与整车层面，Waymo推出了第六代Waymo Driver软硬件套件。Dolgov解释，与前几代专注于验证核心自动驾驶可行性的系统相比，第六代系统的核心演进方向是系统简化、大幅削减传感器与计算成本，并满足大批量流水线生产的制造要求。 该套系统已被集成到由极氪（Zeekr）制造的全新定制出行载具中。Dolgov指出，该车型的外部物理尺寸与现款Jaguar I-PACE相近，但内部座舱完全围绕乘客体验进行重新构想：车身配备靠近自动开启的电动侧滑门，车内采用类似“起居室”的开阔腿部空间设计并升级了交互屏幕。该车型目前已面向Waymo内部员工开放日常测试，并计划在年内向公众乘客提供服务。 指数级里程跨越与全球商业化推进 (17:00) Waymo的运营规模已呈现明显的指数级跳跃。从2009年“Chauffeur项目”十余人团队耗时18个月完成10万英里无干预测试起步，Waymo积累前1亿英里纯无人驾驶里程耗费了约16年，而达成第二个1亿英里（累计2亿英里）仅用了约6个月时间。在累计完成的2000万次纯无人载客行程中，有1000万次集中发生在最近的7个月内。在开城节奏方面，Waymo最初拓展前4座城市耗时8年，而今年已实现在单日内同时启动4座新城市的商业运营。 Dolgov表示，由于当前系统的泛化能力大幅提升，新区域落地的工程重心已转变为高精度评估验证以及与当地社区建立信任。目前Waymo已在美国11座城市开展全无人驾驶商业服务，业务重心理论上已完成从“串行技术排险”到“并行全球商业化”的跃迁，并已正式宣布进军英国伦敦与日本东京的国际化落地计划。 Dmitri Dolgov展示了Waymo凭借混合架构AI模型与长期安全投入跨越炒作周期的经验，并确立了快速规模化复制的商业路径。然而，随着运营网络向全美11座城市及伦敦、东京等海外高密度交通环境延伸，如何在复杂异构的异国路况中持续维持13倍于人类的安全水平，并高效适应各地监管规则与社区信任机制，仍是检验其商业模型可扩展性的核心课题。

[3:51] Dmitri Dolgov: probably the most fun I've ever had in my professional life. [7:05] Dmitri Dolgov: today worldwide somebody loses their life to a crash on our roads every 26 seconds. [12:40] Dmitri Dolgov: structured materialized intermediate representations. [17:00] Dmitri Dolgov's children: doggies and it's Waymos [17:00] Dmitri Dolgov: probably similar amounts of cognition between those two. [19:56] Dmitri Dolgov: more than 13 times safer than a human driver [23:17] Dmitri Dolgov: positively surprised by the emerging capability [25:05] Dmitri Dolgov: heads down in execution mode. [25:05] Dmitri Dolgov: rapid parallel global commercialization.

[0:02] We have an unbelievable treat next. [0:06] Uh a founder who's touching a ton of [0:07] lives and um who's been at it for a very [0:10] long time. How many people here have [0:12] been in a Waymo? [0:16] Okay, [0:16] >> [laughter] [0:16] >> that's a relief. [0:18] The chef eats his own food. I do. I do a [0:20] lot of it. All right. And how many [0:22] people love the Waymo experience? [0:26] Rock on. I'm a daily active now. It's [0:28] incredible. Excellent. Thank you. Um we [0:30] have here the creator [0:33] uh a a man who has been at this mission [0:37] for get this founders who've been in AI [0:39] since 2022 [0:41] almost 20 years [0:43] building in the autonomous vehicle [0:45] challenge. And he has not only been at [0:48] this, he's been at it in the great [0:49] times, at the tough times. He's been [0:52] persistent and he has created something [0:54] that is unlike anything else on Earth. [0:57] Truly exceptional. Please join me in [0:58] welcoming DMITRI DOLGOV. [1:00] >> [applause] [1:04] >> THANKS. GREAT TO BE HERE. ALL RIGHT, [1:06] DMITRI. SO we've got about 25 minutes [1:08] together. Uh the goal is to understand a [1:10] little bit about you, what makes you [1:11] tick, what has made you persist since [1:14] the early days of the DARPA challenge 21 [1:16] years ago all the way through Waymo uh [1:18] early days to today and the future. Um [1:21] let's start on you and then we'll get [1:23] into technology very quickly. Sound [1:24] good? [1:25] >> Sounds good. Okay. So Dmitri, you are [1:28] known by your team as technically [1:31] brilliant, [1:32] incredibly intense, but also very kind [1:34] and humble. [1:35] Um you were [1:36] born in the Soviet Union, raised in the [1:38] States, and then chose to go back to one [1:41] of the most prestigious and intense [1:43] physics programs on the planet in [1:44] Moscow. How did those first few years of [1:47] your life shape you and how did it shape [1:49] your character? [1:50] So my parents went to the same school, [1:52] so that to a large degree drove my [1:55] decision to go back. I went to high [1:56] school. I actually traveled around quite [1:58] a bit. I spent a year in Japan, then [2:00] went to uh high school in the States and [2:02] came back to college to do math and [2:03] physics in in Russia. Uh so that was the [2:06] same school that my parents went to and [2:08] I kind of really grew up uh [2:10] hearing stories about what it is like to [2:13] be at that place. I really wanted to go [2:14] back. And I think you know in terms of [2:15] how it shaped me, I think it really set [2:17] the foundation [2:18] uh you know technical foundation and the [2:22] you know in those early days in college [2:24] one of the most important thing is you [2:26] know acquiring the ability to learn and [2:29] independently explore, so I think that [2:30] really helped me in my future career. [2:33] Now, you did this very intense program [2:36] at Moscow Institute of Physics and [2:38] Technology, and then you decided to keep [2:40] going on the AI path. You earned your [2:42] PhD also in AI. And then you pretty [2:46] quickly were attracted to autonomous [2:47] vehicles. In in '05 you were part of the [2:50] DARPA challenge. Can you tell us about [2:51] those early days, what drew you to [2:52] autonomy? [2:55] That was kind of a light switch [2:57] moment for me. Right? I you know in the [3:00] early days when I went to college and [3:01] then grad school it was more you know [3:04] about just the learning of the [3:06] fundamentals and I [3:09] didn't have a at all a clear picture and [3:13] idea of what I wanted to do after. [3:15] Right? And then I think the timing just [3:17] was just incredibly lucky that this was [3:20] when I was finishing up grad school, the [3:22] urban challenges uh the grand challenge [3:24] and then the urban challenge the one [3:25] that I took place in [3:27] were happening and it clicked. It was [3:29] just uh the the the technology is [3:32] incredibly interesting. The mission is [3:36] so powerful that nothing else come close [3:38] and it's a real product there you can be [3:40] hands-on experience it yourself. So it [3:43] really checked all the boxes for me. And [3:45] as you said this has been 20-some years [3:47] ago, you know who's counting, uh and [3:48] then I've I've never looked back and [3:49] that's what I've been doing since. [3:51] Amazing. So Waymo started out of a [3:54] project at uh Stanford Automotive Lab. [3:57] There were two sides of this building. [3:59] There was the autonomy side and then [4:00] there was a solar car side. Fun fact, um [4:04] I was an idealist, I worked on the solar [4:05] car. [4:07] I got that bet very wrong. [4:08] Uh you bet on autonomy. Tell us about [4:11] the first few years of Waymo from from [4:15] '09 to the formative years. [4:18] So we started in 2009. That was at the [4:20] time the Google self-driving car [4:22] project. Um and the first couple of [4:25] years it was all about [4:28] learning the problem space, [4:30] understanding you know what it means to [4:32] try to put a autonomous vehicle on [4:35] public roads. Uh so [4:38] when we started in service of you know [4:40] those goals of learning and [4:41] understanding the problem space, we [4:42] created a couple of goals for ourselves. [4:44] One was to drive 100,000 miles total in [4:47] full autonomy, which at the time was [4:49] kind of not heard of. [4:51] Uh and the second one was to drive 10 [4:53] routes. Each one was 100 miles long. [4:56] They were all over the Bay Area, uh [4:58] chosen to be you know very difficult. [5:00] And we had to do each one from beginning [5:02] to end in full autonomy. You know still [5:04] with a person behind the wheel uh that [5:06] you know could take control, but the [5:08] challenge was to complete each one [5:10] without an intervention. So it was a [5:12] small team of us, it was about you know [5:14] a dozen people. I was the early crazy [5:17] startup days. Yeah, everybody working [5:18] 24/7, you know writing code and building [5:21] hardware during the day and then you [5:23] know doing some testing at night. Uh and [5:25] it took us about 18 months to complete [5:28] uh both of those challenges. [5:32] Incredible. It seemed impossible at the [5:33] time. Now you guys are on hundreds of [5:35] millions of miles. Absolutely. Um okay, [5:37] so early Waymo days, extreme challenge, [5:41] starting to achieve there. Next few [5:43] years, you have a reputation in your [5:45] team [5:46] for grinding really hard. You were [5:48] sleeping at the office. Tell us about [5:50] Dmitri in the first few years of Waymo [5:53] and and how you formed your leadership [5:55] style. [5:59] I got to say those early days was [6:01] probably the most fun I've ever had in [6:04] my professional life. And it is that [6:07] that [6:08] you know [6:09] momentum and that pace of an you know [6:13] early startup days. All right, when you [6:16] are making so much progress you know [6:18] every hour of every day. And you're [6:20] doing everything, right? You are uh [6:22] working on setting up the hardware in [6:24] the cars and then you know configuring [6:26] and calibrating the sensors and the you [6:28] know your pose estimation system. And [6:31] then you are writing software during the [6:32] day and it's everything, right? It's the [6:33] core of the you know the software, the [6:36] algorithms that drive the car. It is all [6:37] of the tools [6:39] uh and UIs and the user experience in [6:41] the car. So you're doing everything. [6:43] You're learning at an insane rate and [6:44] you're making progress at an insane [6:46] rate. So [snorts] uh that's that was uh [6:49] the those were the early days uh of [6:51] project Chauffeur. And then you know in [6:53] those couple of years we've convinced [6:54] ourselves that you know yes this is [6:57] worth pursuing, so we doubled down and [7:00] started actually building towards the [7:02] future of a of a fully autonomous [7:04] product. [7:05] Okay, so exciting first few years, [7:07] intense, fast-paced, technically really [7:10] difficult. Now take us to 2016-17 for a [7:13] moment. This was a period where we [7:15] actually had a hype cycle in AI. Turns [7:18] out there's been a few of them. And [7:21] AV autonomous vehicle was at the center [7:23] of that hype cycle. I mean I remember [7:24] just so many companies going after this. [7:28] And then there was a massive slump. [7:30] And when most people gave up or failed [7:32] or fell apart, you guys persisted. And [7:35] you were a leader in that persistence. [7:37] For all the builders in this room, how [7:40] did you navigate through the hard times? [7:45] So first a comment on you know [7:47] what these cycles kind of you know look [7:49] like to me and you know how I've seen [7:51] them. And you said there there's been [7:52] many some in AV but more generally. And [7:55] kind of the what often leads to a cycle [7:57] like this is some [8:00] rapid some breakthrough that leads to [8:05] very rapid progress in the kind of the [8:08] early parts of the problem and very [8:10] rapid investment flow in the early part [8:12] of the problem. And and in the AVs the [8:14] the problem has always had this property [8:16] that it's very easy to get started, but [8:18] it's you know very difficult to take it [8:20] all the way to a real product, full [8:22] autonomy and superhuman performance. Uh [8:25] so it's somewhat natural given those [8:28] ingredients that whenever there's been [8:29] you know big breakthrough in technology, [8:31] whether it's you know convolutional nets [8:33] or you know transformers or you know [8:35] large language models, it's led to the [8:37] cycle okay now the problem is going to [8:39] be you know very And it kind of reshapes [8:41] the early part of the curve, but it you [8:43] know doesn't change the long tail of it. [8:46] Uh so [8:48] in for us [8:50] I think it was [8:52] understanding [8:53] that it's [8:56] not going to be an easy problem. And but [8:58] it's a very important one. And having [9:02] you know believing in the mission I [9:03] think it's uh today [9:07] worldwide somebody loses their life to a [9:10] crash on our roads every 26 seconds. [9:14] So I guess it's the combination of [9:16] knowing that the mission is [9:18] really really important and then [9:20] understanding what you're up against and [9:21] not looking for a kind of easy wins or [9:24] quick solutions or silver bullets that [9:26] help the team you know have the right [9:28] stamina to go the distance. [9:30] Brilliant. So you guys were in this [9:33] moment where it was really easy to get [9:35] started. A lot of people got there, but [9:38] you guys actually persisted. And got [9:40] through to the other side with a truly [9:41] magical experience. Pretty much every [9:42] hand in this room went up. Truly magical [9:45] experience because of that persistence. [9:47] Let's talk about technology today. [9:49] A lot of people are talking about world [9:50] models. You have had all the components [9:53] of world models for many years. [9:56] How do you think about a world model and [9:58] what is Waymo's version of a world [10:00] model? [10:02] Uh yeah, there are a few things that [10:05] a few terms that people use nowadays. Uh [10:07] people talk about world models, [10:09] world action models, omni models, you [10:12] know, visual language action models. And [10:14] at the core of each [snorts] there's an [10:16] ingredient that is relevant uh and [10:18] really important for Waymo and for what [10:21] we've been building in our AI ecosystem. [10:24] So, at the [10:27] core of our AI ecosystem is what we call [10:30] the Waymo foundation model and it powers [10:33] three main pillars of our AI and our [10:37] tech. So, the driver, [10:39] the simulator, and the critic. [10:42] Right? And those are [10:44] very related but distinct tasks. Uh so, [10:48] at the core of what our foundation model [10:51] needs to be capable of [10:54] uh are things like it needs to [10:55] understand how the world works, the [10:57] physics, the dynamics of the the [10:59] physical world. [11:01] Uh [11:02] and it needs to understand, you know, [11:04] what it is to be a good driver and you [11:06] know, how [snorts] the effects of the [11:08] actions of that driver or our agent on [11:10] other agents in the world. And then we [11:11] need to instantiate those in that [11:13] physical agent that we're putting on the [11:14] roads. So, in a way that foundational [11:16] model that that we've been building over [11:19] the years is [11:20] uh a multimodal world action language [11:23] model. All right, so it's multimodal in [11:25] that it needs to be able to reason about [11:27] not just, you know, images or video but [11:28] also other sensors like clay lidars and [11:31] radars. [11:32] Uh it is a world action model in that, [11:35] you know, it really has to have a deep [11:37] precise understanding of the 3D spatial [11:39] properties of the world, the dynamics, [11:41] the physics, the um [11:43] the [11:43] uh behavioral aspects of other agents [11:45] like cars, pedestrians, cyclists, and so [11:47] forth. Uh [11:49] and we are not just, you know, passively [11:51] modeling those worlds, we're an active [11:53] participant in it. Uh so, we don't not [11:55] only have to, you know, the world model [11:57] has to be controllable but also, you we [11:59] need to have a deep understanding of [12:01] what it means to be a, you know, good [12:02] agent in that world. Uh and finally, [12:05] it's aligned with language and that [12:06] allows us to kind of pull in the general [12:09] world knowledge of a VLM into our model [12:12] that is very, very useful in giving us a [12:14] boost in the uh [12:17] uh in understanding the semantics and [12:20] the deep social aspects of driving. [12:23] Right? [snorts] [12:24] And we've been, you know, uh working on [12:27] productionizing that model, you know, [12:29] for years and it really it um [12:32] requires an extremely high degree of [12:36] performance and accuracy and realism in [12:38] every aspect of what we just talked [12:39] about. [12:40] Brilliant. So, with this driver, [12:42] simulator, critic [12:44] architecture, [12:45] uh there's also been a lot of [12:46] conversation about end-to-end [12:47] architectures. [12:48] Is that the [12:50] appropriate dichotomy? How do we think [12:52] about the approach to getting us to [12:55] extreme performance efficiency [12:57] autonomous vehicles that are [12:59] totally generalizable? Yeah, so to be [13:01] very clear, you know, the world model [13:03] that is described, uh the Waymo [13:04] foundation model, is an end-to-end [13:06] model. [13:07] Uh so, when we talk about an end-to-end [13:08] model, typically we mean that, you know, [13:10] it's one model that goes from sensors to [13:13] decisions or actions. [13:15] >> [snorts] [13:15] >> And uh you know, there are some very [13:17] nice properties of uh such a model. [13:20] One of the most important ones is that [13:22] uh [13:23] you [13:24] it learns the right rich representations [13:27] between different components of the [13:28] system like the encoder and the decoder [13:30] or the perception and the planning part [13:32] of your system. [13:33] As [snorts] opposed to something where [13:34] that interface is engineered, which is [13:36] not sufficient for a, you know, a task [13:38] like driving. [13:40] Uh now, I do think there's a a false [13:42] dichotomy there that's, you know, [13:43] end-to-end or something else. [13:45] Uh really, it my mind has always been [13:48] the question of, you know, it's [13:49] end-to-end [13:51] and then, you know, what else? And what [13:53] else do you need to build if you want to [13:56] have a product that is fully autonomous, [13:59] has superhuman level of safety, and you [14:01] want to deploy it at scale and drive [14:03] hundreds of millions of miles. And there [14:05] it turns out that kind of the basic [14:07] vanilla, you know, end-to-end system is [14:10] insufficient, right? So, it's there's a [14:13] massive difference between using [14:14] end-to-end versus purely relying on it. [14:17] So, at Waymo we've really gone beyond [14:18] that kind of basic vanilla end-to-end [14:20] approach and we've augmented the learned [14:23] representation with structured [14:25] materialized intermediate [14:26] representation. And what that allows us [14:29] to do [14:30] are a few very important things that you [14:32] might not actually need if you are, you [14:34] know, building a different product, if [14:35] you're building a driver assist system [14:37] or, you know, a prototype, a demo, or a [14:39] small scale deployment. But again, those [14:41] things are absolutely critical if you [14:43] want to go all the way to a fully [14:45] autonomous safe system with superhuman [14:48] performance. And those are things like [14:50] having extra validation at runtime in of [14:54] the agent that's running on the car in [14:56] the physical world. It's things like [14:58] richer training and evaluation recipes [15:02] that are very difficult or impractical [15:04] to do in a pure kind of basic end-to-end [15:05] system where this structured [15:07] materialized representation gives you a [15:09] boost in things like closed loop [15:11] evaluation, closed loop training, uh [15:13] rich reward functions for reinforcement [15:15] learning. So, that's been our approach. [15:16] And all the human feedback that you get [15:18] from support and [15:21] drivers dropping in and all of that, [15:23] it's essential to have this type of [15:24] architecture. [15:25] >> Exactly. Exactly. [15:26] Makes perfect sense. [15:28] So, not only have you innovated on the [15:30] software stack but also the hardware [15:32] stack. There's a sixth generation now of [15:34] Waymo driver and you guys have always [15:36] focused on being the driver. [15:38] Tell us about the new sixth generation [15:40] and what was it like the first time [15:42] you've interfaced with it? It was a so, [15:44] the sixth generation uh is our most [15:46] advanced uh hardware suite and sensor [15:49] suite yet. The focus there has been on [15:52] performance but also on simplification, [15:55] drastic cost reduction, and uh high [15:59] scale volume production. Uh [16:02] this is the uh [16:04] the driver that's powering our latest [16:06] vehicle platform, that's the O Hi. [16:08] Uh we [16:10] uh earlier this year started fully [16:12] autonomous operations. Uh it's currently [16:14] only open to employees but coming to [16:16] our, you know, all of the riders later [16:18] this year. [16:19] Uh and yeah, I I had a chance to take a [16:23] ride in one [16:24] as soon as we started [16:26] uh running fully autonomous operations. [16:28] And you know, they [16:30] I spent a lot of my life in various [16:32] generations of our cars. Every once in a [16:34] while there's uh kind of a new first [16:37] moment and that was definitely it. It's [16:39] just that they the cool car is designed [16:41] around the rider experience. It is [16:44] uh even though the external footprint of [16:46] the car is about the same as the i Pace [16:48] but inside you get in and it feels like [16:50] it's a living room. All right, there's [16:51] so much space in the back. We have new [16:52] screens. We have these doors that, you [16:54] know, slide open and will, you know, [16:55] open automatically when you approach the [16:56] car. So, I I I had a blast and I I can't [17:00] wait to have this car in our fleet open [17:02] to everyone. So, you guys are going [17:04] through a period of incredible scaling. [17:06] For for many years you were in the lab [17:08] R&D, surely. [17:10] It took 16 years-ish to get to 100 [17:13] million miles, six months-ish to get to [17:15] 200. Things continue to scale really [17:17] rapidly. 11 cities now, many, many more [17:21] um [17:22] on the horizon. Tell us what is it like [17:24] to scale a new city and then tell us [17:26] about your daily life with a Waymo. How [17:28] do you use it as a creator? [17:33] Well, there's a lot. Okay, so [17:34] exponential scaling. First of all, [17:35] absolutely. It's been a [17:38] phase transition for us in how we're [17:40] scaling. So, [17:42] to give you a couple of additional data [17:44] points, it took us eight years from the [17:46] day when we started our fully autonomous [17:48] operations [17:50] to the day when we had our service, our [17:54] uh our driver providing rides to the [17:57] public in four cities. Earlier this [17:59] year, just a few weeks ago, we launched [18:01] four cities in one day. [18:03] We've uh [18:04] given [18:06] over 20 million fully autonomous rides. [18:10] 10 of those million happened in the last [18:12] seven months. [18:13] Amazing. [18:13] >> So, that's what exponential scaling [18:14] looks like. [18:14] >> Amazing. Uh [18:16] launching new cities, uh [18:19] there's, you know, operational [18:21] components, you know, show up. Uh you [18:23] know, you have to uh collect the data, [18:25] characterize the environment. Um I mean, [18:28] validate the driver. [18:30] Uh a significant part of it is starting [18:32] the conversation with the local [18:33] communities because it's a new thing, [18:35] it's a new product. So, it's on us to [18:36] earn the trust of the, you know, the the [18:38] people there. [18:39] Um and then um more often than not today [18:42] we're seeing that the driver is [18:43] generalizing incredibly well and it's [18:45] just a matter of high [snorts] fidelity [18:48] rigorous evaluate evaluation and [18:49] validation before we deploy the fully [18:52] autonomous product. And then uh we can, [18:54] you know, we grow from there. [18:56] And then the last, what was the last [18:57] part of the question? Oh, what is my [18:58] daily life? Yeah, it was a multi-part [19:00] question, yeah. Uh [19:02] Waymo I mean, Waymo is [19:03] uh my is how I get around nowadays. [19:06] Uh that's how I got here today. Uh it [19:08] was, you know, a a great ride from Palo [19:10] Alto up to San Francisco [19:12] uh on on freeways. [19:13] Uh [19:14] I my family use it. I have three kids. [19:17] They love Waymo. Uh they [19:20] they they I think nowadays they get [19:21] annoyed if on a rare occasion we have [19:23] to, you know, be in a car that's driven [19:24] by you know, myself or my wife or, you [19:27] know, another human being. They're like, [19:28] "Okay, what was going on there?" [19:29] >> same way at this point. Uh [19:31] they they love it. Uh it's been part of [19:34] their lives, you know, for the entirety [19:36] of their lives. There's [19:37] uh when we are [19:39] uh driving around, there's two things, [19:41] there's only two things that get cool [19:42] outs from my kids nowadays. It's doggies [19:44] and it's Waymos. [19:46] >> [snorts] [19:47] >> Nice. [19:48] Um okay. [19:49] Uh probably similar amounts of cognition [19:51] between those two. [19:52] Um, okay, so let let's talk about um, [19:56] uh, safety. You know, one of the most [19:58] meaningful, exciting parts of partnering [20:01] with Waymo has been the fact that [20:03] there's 1.19 million people a year on [20:06] Earth that die in road accidents. All [20:09] right, it is this is life or death. You [20:11] know, not only does it touch everyone in [20:12] this room, but everybody has some [20:15] connection who's been impacted by this. [20:18] Um, you have been about safety from the [20:20] very beginning. [20:21] And it's actually pretty hard, you know, [20:23] in a Silicon Valley where it's move fast [20:25] and break things and see what happens, [20:27] you guys have been incredibly patient [20:30] uh, with safety. Can you tell us about a [20:32] story that made it very real to you and [20:35] how you keep that safety culture at [20:37] Waymo? [20:38] >> [snorts] [20:40] >> So, the numbers you mentioned, [20:43] that's [20:44] what drives all of us at Waymo and [20:47] that's the status quo is not okay. All [20:48] right, we've kind of grown this over [20:50] time, [20:51] uh, but challenging the status quo is [20:54] really important to everyone at our [20:56] company. You're uh, [20:58] absolutely right that how you go about [21:01] building a system like this is different [21:03] from you might do in other areas, in [21:05] other fields, in other industries where [21:07] safety has to be the non-negotiable [21:10] foundation and you have to build that [21:12] into your everything that you do from [21:14] day one. Your model architecture, your [21:18] training and evaluation recipe is the [21:20] mindset of the team. It can be very [21:22] tempting to, you know, focus [21:23] [clears throat] on capability first and [21:25] get to the, you know, 90% [21:27] uh, very quickly, but how you go about [21:31] the first 90% [21:33] is totally different problem for, you [21:34] know, how you go about, you know, [21:36] getting to your next, you know, Nines. [21:38] So, get keeping that in mind and [21:40] focusing on safety as the non-negotiable [21:43] fundamental layer from day one is super [21:45] important. And then, uh, you know, today [21:47] we're driving more than uh, 4 million [21:50] miles in full autonomy per week. [21:53] Uh, and you see [21:56] a lot of [21:58] uh, events from the field and uh, today [22:01] we have the data [22:03] uh, over 170 million miles fully [22:05] autonomous miles where we see that the [22:07] Waymo driver is more than 13 times safer [22:11] than a human driver when it comes to [22:13] serious injury causing collisions in the [22:15] cities where we operate. And you see [22:17] that [22:18] sort of superhuman safety behavior [22:20] manifesting itself on the roads, [22:23] you know, daily. All right, I see [22:24] examples of, you know, recently there [22:26] was a little while ago there's an [22:27] example uh, that I saw of um, uh, uh, a [22:31] person [22:32] uh, I think it was a young woman on [22:34] electric scooter [22:36] on the road and then she lost control [22:38] and tripped and fell right in front of [22:40] the Waymo and the Waymo [22:42] uh, driver showed superhuman [22:44] uh, accuracy and reaction time and was [22:46] able to, you know, swerve and brake and [22:48] everybody walked away. So, it's uh, [22:51] things like this that, you know, uh, [22:53] my my uh, I myself personally and the [22:55] team, the whole team find very rewarding [22:58] uh, in terms of actually having a real [23:00] impact on the safety of our roads. And [23:02] the scale that we're operating, that 13x [23:04] reduction means that we are preventing a [23:06] serious injury [23:07] every 8 days. [23:09] And that impact will just grow as we [23:11] scale up. Wow. [23:12] We're going to open the room to audience [23:14] questions, a couple in in just a moment. [23:17] Before we jump in, I heard a story about [23:19] uh, the lidar detecting or the radar [23:22] detecting the footsteps of somebody [23:25] behind a bus. Did that happen and how [23:28] does that work? Yeah, yeah, this was one [23:30] of those moments where uh, [23:32] I was [23:34] positively surprised [23:36] by the emerging capability of our [23:38] system. And the situation was [23:41] uh, this was I think in San Francisco. [23:43] Uh, the Waymo driver was at an [23:44] intersection. There was a bus that [23:47] crossed uh, and we're, you know, sitting [23:49] there waiting at a red light. Uh, so the [23:51] bus crossed and stopped partially [23:53] blocking the intersection. So, our light [23:55] intersection. So, then our light turned [23:57] green. The Waymo driver started to [23:58] proceed and as it's proceeding, [24:01] it detects a pedestrian on the other [24:05] side of the bus. [24:07] And you know, you can't see through the [24:09] bus. It's, you know, uh, not through [24:12] lidars, not radars, not cameras, you [24:13] know, the the windows are reflecting the [24:15] people inside the bus. [24:17] Uh, and then, you know, starts to [24:19] react defensively and sure enough, a [24:21] pedestrian emerges from behind the bus [24:23] and then we're able to nudge around them [24:25] and everybody, you know, goes on their [24:27] way. So, when I saw that, I was I have a [24:31] blew my mind. I wasn't sure what was [24:32] going on. Um, I guess capable as [24:35] superhuman as the Waymo driver is, it [24:36] doesn't see through solid objects. So, [24:39] actually what turned out was happening [24:40] is that our lidar [24:43] had was bouncing the signal, you know, [24:45] under the bus and got a little bit of a [24:48] sparse return from the movement of the [24:51] person's feet under the bus and that was [24:53] enough for the Waymo AI to not only [24:56] detect that there's a pedestrian there, [24:57] but also make a prediction about what's [24:59] going to happen in the future and keep [25:01] everyone safe. Mind-blowing. Pretty uh, [25:04] unbelievable. We've got time for one [25:05] question from the group. [25:07] Uh, if anybody has [25:09] a key [25:10] Jim, sorry, no free codes. Not at this [25:12] one. No. Yes, Jim, please. Thank you. [25:15] Does this work? Yes. I was just saying [25:17] congratulations on all you've achieved. [25:19] It's really mind-blowing. If you think [25:21] about the next 5 to 10 years, really [25:24] focus on the business model, what are [25:26] the milestones, what happens in major [25:29] cities, what's going to be different [25:31] than where we are today? Just kind of [25:33] walk us through through your vision of [25:34] the future. [25:37] So, we're heads down in execution mode. [25:39] Uh, [25:39] >> [snorts] [25:39] >> we've transitioned from an intentional [25:43] sequential de-risking of the driver and [25:46] key parts of the business to rapid [25:49] parallel global commercialization. Uh, [25:52] that means uh, deploying the Waymo [25:54] driver [25:55] uh, in more places uh, across the United [25:57] States. Today we're in 11 cities uh, [26:00] operating fully autonomous and serving [26:02] uh, our riders. That uh, we're going to [26:04] expand in those existing places. We're [26:06] going to add new geographies, new [26:08] cities. We're also expanding [26:09] internationally. Announced that this [26:11] year we'll plan to offer a service in [26:12] London and in Tokyo. So, uh, you will [26:15] see us uh, just accelerating that [26:17] deployment all in service of our [26:18] mission. [26:20] Good news to our team in London. Well, [26:22] we covered a lot, Dmitri, from the very [26:24] early days where you could get a lot of [26:27] distance with not a lot of technology to [26:29] then persisting through extremely hard [26:31] times in autonomous vehicles and getting [26:34] that extra mile. We talked about world [26:36] models, uh, driver simulator critic [26:38] architecture. We got into the hardware, [26:41] the sixth generation hardware, safety [26:44] and scaling. Uh, but most of all, I hope [26:47] that we learned a little bit more about [26:49] Dmitri, the man who's brought the magic [26:51] that is Waymo to to so many of us. And [26:54] as I've gotten to know you more and [26:55] more, I'm constantly struck not only by [26:56] your your brilliance and your [26:58] persistence and performance, but also [27:00] your humility. [27:01] Uh, it says a lot for accomplishing this [27:04] much. Thank you, Dmitri. Please join me [27:05] in thanking Dmitri for all he does. [27:07] Thank [applause] you. Thank you. Thank [27:08] you. And the many lives saved. Thank [27:09] you.