Waymo Rejects Vision-Only Driving as it Plans Global Expansion

Waymo计划全球扩张，拒绝纯视觉自动驾驶方案 Waymo副总裁Srikanth Thirumalai表示，出于乘客安全考量，传感器冗余仍是不容妥协的底线，而更低成本的硬件与车队合作正助力公司推进全球规模化扩展。 在这段18分钟的对话中，Waymo的Srikanth Thirumalai解释了为何公司始终坚持采用涵盖lidar、radar和摄像头的多传感器技术方案而非纯视觉系统，并强调乘客安全是重中之重。他详细介绍了Waymo如何利用EMMA以及Google的Gemini等多模态视觉-语言模型推进自动驾驶，同时通过其第六代自动驾驶套件降低运营与硬件成本。此外，Thirumalai还探讨了冬季气候下的车辆部署、与Uber及Avis的战略车队合作、对驾驶员劳动力的长期影响、消费者信任度的变化以及国际扩张计划。 Waymo 坚持采用摄像头、激光雷达（LiDAR）与毫米波雷达互补的多传感器冗余方案以保障全天候安全，并向纯视觉开发者发起挑战，要求其在 1 亿英里商业运营规模下拿出可比的客观安全数据。 (0:35) Waymo 正在基于 Google Gemini 研发类似 EMMA 架构的多模态视觉-语言-动作（VLA）模型，并预计完全简化的端到端自动驾驶架构距离全量生产部署仍需数年时间。 (5:31) Waymo 第 6 代自动驾驶硬件通过精简传感器数量并提升精度与性能大幅削减系统成本，同时 AI 模型的跨地域泛化能力使其能将积雪等恶劣天气作为可攻克的工程问题予以解决。 (8:04) 通过与 Avis 和 Uber 建立战略车队合作，Waymo 充分利用合作伙伴在车辆运营与场站维保方面的专业经验，加速在达拉斯和奥斯汀等重点城市的商业化扩张。 (12:03)

Srikanth Thirumalai argues that Waymo's path to safe, commercial autonomous driving relies on non-negotiable multi-sensor redundancy, multimodal foundation AI models, and strategic operational partnerships. While competitors pursue vision-only systems, Waymo prioritizes empirical safety benchmarks over hardware dogma and relies on established fleet operators to scale sustainably. Multi-Sensor Redundancy and Empirical Safety Verification (0:35) Waymo rejects vision-only autonomous driving architectures, contending that public trust requires combining cameras, lidar, and radar. Thirumalai explains that Waymo remains stubborn about passenger safety while staying flexible on technical execution, treating these three sensor types as complementary. When severe weather or unexpected edge cases degrade camera vision, radar and lidar provide the necessary safety net to maintain vehicle control. Addressing competitors like Tesla and Wayve that argue cameras alone are sufficient, Thirumalai notes that Waymo is not dogmatic about its hardware stack. However, he challenges vision-only proponents to present empirical proof, stating that any alternative approach must objectively demonstrate equivalent safety over 100 million miles of driving before Waymo would consider abandoning redundant sensors. Multimodal Generative AI and Architectural Evolution Waymo is advancing its AI stack through vision-language-action foundation models, evidenced by its published research on the End-to-End Multimodal Model for Autonomous Driving (EMMA). Thirumalai describes vision-language models (VLMs) as an evolution of classic encoder-decoder concepts. While current models primarily process video streams and text prompts, Waymo and the broader industry are actively integrating radar and lidar signals into these foundation models. Thirumalai confirms Waymo relies on Google's Gemini models for this research, dismissing the need to test external foundation models. Invoking the "bitter lesson"—the principle that general methods leveraging massive computation and data ultimately outperform specialized human-engineered heuristics—he forecasts that fully unified, end-to-end driving architectures remain a few years away from commercial deployment. Hardware Economics and Fleet Management Partnerships (8:04) Although taking a Waymo in San Francisco currently costs roughly the same as an Uber, Waymo is attempting to compress operating expenses. Thirumalai acknowledges that first-generation driverless economics are inherently high, but highlights that Waymo's sixth-generation hardware suite reduces capital expenditure by using fewer total sensors with upgraded fidelity. Additional cost cuts stem from optimizing depot facilities, fleet maintenance, and remote response operations. To expand into markets such as Austin and Dallas without absorbing excessive ground-level overhead, Waymo divides labor with commercial transportation leaders. Avis provides vehicle maintenance and depot infrastructure, while Uber supplies passenger demand. In Austin, Thirumalai notes that Waymo accounts for over 99% of driverless trips on Uber's network. Furthermore, he treats adverse conditions like snow as solvable data-collection milestones rather than fundamental technological barriers. Passenger Sentiment and Gradual Labor Transition (13:43) Thirumalai addresses concerns regarding driver displacement by framing autonomous deployment within the vast scale of global mobility. Because global transportation comprises trillions of vehicle miles, autonomous fleets will take significant time to capture meaningful market share. In his view, this extended timeline gives society, cities, and human drivers sufficient room to adapt to shifting labor demands. Regarding public acceptance, Thirumalai points to San Francisco's two years of commercial ride-hailing as evidence of shifting consumer attitudes. While first-time riders often board with a mixture of excitement and hesitation, Thirumalai observes that passengers routinely normalize the situation within seconds, forgetting that the driver's seat is unoccupied. This psychological adjustment has driven customer sentiment steadily higher as driverless trips turn into routine transit. Waymo continues to prepare international road-testing sites for future commercial launches, yet key industry uncertainties remain. Long-term progress depends on whether competitors can ever validate vision-only systems over 100 million commercial miles, and whether fully end-to-end multimodal models can safely replace modular perception and planning stacks.

在本次对谈中，Waymo高管Srikanth Thirumalai阐述了公司的核心发展逻辑：坚持以乘客安全为中心的多传感器冗余策略，依托第六代硬件迭代与外部车队运营伙伴降低单位经济成本，并借助多模态生成式AI推进端到端架构演进与规模化扩张。 坚守多传感器冗余与亿英里安全基准 (0:35) 针对特斯拉和Wayve等竞争对手倡导的纯视觉方案，Thirumalai表示Waymo恪守“对乘客安全与体验绝不妥协，对技术实现路径保持灵活”的基本原则。Waymo认为，在建立公众信任的关键阶段，传感器冗余至关重要。摄像头、激光雷达（Lidar）和雷达（Radar）构成互补模态，当某一类传感器在极端恶劣天气或边缘场景下面临感知挑战时，其他模态能够协同补位，形成完整的安全防护网。 面对行业内关于纯视觉路线的争论，Thirumalai指出Waymo在硬件配置上并不持教条立场，但强调方案替代必须依赖客观实证。他提出，若有竞争对手能够拿出客观数据，证明其系统在超过1亿英里的实际商业化行驶规模下达到同等安全水准，Waymo才会重新评估其硬件架构选择；在此之前，多传感器融合依然是其保障规模化安全的核心基石。 多模态生成式AI与端到端模型演进 围绕生成式人工智能的行业趋势，Thirumalai认可视觉语言模型（VLM）在自动驾驶领域的潜力，将其视为经典编码器-解码器架构的技术升级。Waymo已公开基于EMMA（端到端自动驾驶多模态模型）架构的研究成果；目前行业主流VLM虽主要处理视频流与文本提示，但技术演进方向正逐步将雷达与激光雷达等空间传感器信号整合至多模态基础模型中。 在底层基础模型选择上，Thirumalai透露Waymo直接复用谷歌母公司的Gemini模型，其综合能力已足以支持研发需求，无需试验其他模型。但他同时预计，高度简化的端到端自动驾驶架构实现工程化部署尚需数年时间，当前技术路线正遵循“惨痛教训”（The Bitter Lesson）所揭示的客观规律，即极简架构配合海量训练算力与数据最终将胜出。 硬件迭代与第三方运营降本路径 (8:04) 针对目前旧金山无人驾驶行程收费与Uber等传统网约车接近的现状，Thirumalai解释称，突破性技术在应用早期的单位经济成本天然偏高，Waymo正从全平台各环节推进降本措施。其第六代无人驾驶硬件套件（6th-Generation Driver Hardware）通过减少传感器搭载总量并提升单个器件精度，实现了硬件层面的显著降本；同时，Waymo也在逐步压降后台场站运营、车辆维保以及远程协助所产生的运营开销，以期将前期平台投资转化为长期规模效应并让利于乘客。 为快速进军达拉斯和奥斯汀等新市场并避免自建重资产设施，Waymo选择与传统租赁及出行平台建立战略协作。其中，Avis凭借数十年的大型车队管理经验承担地面维保与场站运营，显著加快了开城进度；在出行服务网络端，Uber首席执行官此前透露Waymo占其奥斯汀网络无人驾驶行程的99%以上，Thirumalai表示尽管当前体量在整体网约车大盘中比例较小，但早期的运营指标与车辆利用率反馈非常积极。 场景泛化应对恶劣天气与劳动力转型 (13:43) 针对拓展纽约等复杂城市及应对降雪等恶劣气候的质疑，Thirumalai表示Waymo的大型AI模型在不同地理区域间展现出良好的泛化能力，天气带来的感知挑战属于可通过持续收集数据逐步解决的工程问题，而非技术本质瓶颈。谈及对人类网约车和出租车司机就业的潜在冲击，Thirumalai指出全球出行市场规模庞大，达数万亿英里里程，自动驾驶的大规模渗透需要相当漫长的时间，这为社会机制与劳动者提供了充分的适应过渡期。 此外，在商业运营批准落地旧金山两周年之际，Thirumalai观察到公众心理已发生显著改变：多数乘客在首次乘车时虽伴有紧张与好奇，但在车辆起步数秒内便迅速习惯无人驾驶状态，推动整体用户满意度持续走高。目前Waymo工程团队亦在同步推进国际化扩展，其已公开测试的海外地点将作为未来商业落地的首批据点。 Thirumalai系统阐释了Waymo在保障多模态冗余安全底线的前提下，依靠软硬件升级与生态合作改善单位经济效益的发展路径。不过，完全端到端的简化模型商用仍需数年时间，全球出行市场的全面渗透受制于漫长的时间周期，其海外市场的具体落地时间表也仍待明确。

## Background 该视频由频道 Ai4 于 2025 年 8 月 29 日发布在 YouTube 平台上。内容记录了《财富》杂志资深撰稿人 Jessica Mathews 与 Waymo 工程副总裁 Srikanth Thirumalai 在 Ai4 大会现场展开的炉边对话，围绕 Waymo 的技术路线、降本进展及行业发展展开交流。 ## TL;DR 对话中，Srikanth Thirumalai 深入阐释了 Waymo 坚持多传感器融合方案背后的安全底线与工程哲学，强调必须依靠数亿英里规模的实际验证数据说话。针对前沿技术，他探讨了多模态大模型在自动驾驶中的整合前景以及遵循“苦涩的教训”的演进方向。此外，他还详细介绍了第 6 代 Waymo Driver 的硬件降本路径、与 Avis 和 Uber 等伙伴的运营协同，并回应了公众信任建立过程以及 AI 对人类司机就业影响的长期演进逻辑。 ## Atmosphere & style 整场炉边对话节奏明快、务实严谨且不失幽默。面对主持人关于纯视觉方案与劳动力替代等尖锐议题，Thirumalai 展现出资深工程专家的沉稳与坦诚，既有对安全标准毫不妥协的原则坚持，也有对技术演进规律的客观洞察，整体传递出一种基于扎实落地与数据验证的笃定与从容。 ## Key takeaways - [00:04](https://www.youtube.com/watch?v=y40-CMBbRVE&t=4s) Waymo 坚持“对安全体验极度坚持、对实现细节保持灵活”的工程哲学，根据软硬件发展持续迭代技术栈，但绝不以牺牲安全性为代价。 - [00:04](https://www.youtube.com/watch?v=y40-CMBbRVE&t=4s) LiDAR、毫米波雷达与摄像头构成关键的多模态冗余安全网，在极端天气和恶劣光照等单模态受限场景下能够实时互补托底。 - [00:04](https://www.youtube.com/watch?v=y40-CMBbRVE&t=4s) 评价自动驾驶路线拒绝空谈概念，必须以数亿英里（100M miles）级真实道路行驶规模下经过验证的客观安全数据为准绳。 - [05:31](https://www.youtube.com/watch?v=y40-CMBbRVE&t=331s) 多模态大模型在自动驾驶演进中扮演核心角色，行业正加速将激光雷达与雷达等非视觉模态全面融入编码器-解码器架构。 - [05:31](https://www.youtube.com/watch?v=y40-CMBbRVE&t=331s) 遵循 AI 领域的“苦涩的教训”（The Bitter Lesson），Waymo 致力于以极简的模型架构结合海量训练数据作为技术演进的长期北极星。 - [08:04](https://www.youtube.com/watch?v=y40-CMBbRVE&t=484s) 模型规模与丰富数据的扩展赋予了系统强大的跨地域泛化能力，雨雪等恶劣气候在工程技术上属于完全可解的问题。 - [08:04](https://www.youtube.com/watch?v=y40-CMBbRVE&t=484s) 第 6 代 Waymo Driver 在提升感知性能的同时显著精简传感器数量，实现整车硬件平台制造成本的大幅削减。 - [11:16](https://www.youtube.com/watch?v=y40-CMBbRVE&t=676s) 优化全链条单位经济效益涵盖硬件制造成本与场站运维开销，最终将直接转化为乘客端更亲民、更具竞争力的乘车资费。 - [11:16](https://www.youtube.com/watch?v=y40-CMBbRVE&t=676s) 借助 Avis 租车等成熟产业伙伴的基础设施网络负责场站流转与车辆维保，使 Waymo 避免了重复建设重资产运营网络。 - [13:43](https://www.youtube.com/watch?v=y40-CMBbRVE&t=823s) 面对全球数万亿英里的出行体量，自动驾驶仍处萌芽初期，短期内不会淘汰人类司机，社会与劳动力市场拥有充分的调整适应期。 - [15:22](https://www.youtube.com/watch?v=y40-CMBbRVE&t=922s) 乘客对无人驾驶的信任建立存在“数秒定律”：尽管初次乘车前伴随焦虑，但车辆起步平稳行驶数秒内即可建立起安全感并迅速常态化。 - [17:50](https://www.youtube.com/watch?v=y40-CMBbRVE&t=1070s) 海外拓展已确立为 Waymo 的核心战略方向，团队正全力推进此前已开展前期道路测试的海外城市作为首批国际落地站点。 ## Detailed walkthrough ### [00:04](https://www.youtube.com/watch?v=y40-CMBbRVE&t=4s) 自动驾驶赛道现状与 Waymo 的传感器路线哲学 在拉斯维加斯举行的 Ai4 会议现场，主持人 Jessica Mathews（《财富》杂志资深撰稿人）开场提到，亚马逊旗下的 Zoox 正在拉斯维加斯大道进行测试，而特斯拉也在会场楼下展示全自动驾驶（FSD）演示。当被问及本次行程是否去体验了竞争对手的技术时，Waymo 工程副总裁 Srikanth Thirumalai 幽默地回应称，自己昨天刚飞抵、几小时后就要飞走，行程匆忙，虽然从外面看到了特斯拉的展位，但并没有亲自试乘。 随后对话切入核心技术路线的争论：特斯拉和 Wayve 等公司主张纯视觉方案（Camera-only），声称不需要激光雷达（LiDAR）和毫米波雷达（Radar），而 Waymo 却始终坚持多传感器融合方案。Thirumalai 阐述了 Waymo 的根本原则：**“对安全和乘客体验极度坚持，对实现手段极度灵活”**（"Stubborn about the experience, flexible on the details"）。他表示，在自动驾驶早期阶段，向公众证明技术的安全性是至高无上的目标。底层软硬件技术随时间飞速发展——模型更强大、车载算力更强、传感器更精准且更便宜——Waymo 会根据技术成熟度持续调整技术栈，但绝不以牺牲安全性为代价。 激光雷达与毫米波雷达之所以不可或缺，在于它们能够形成关键的“互补与安全网”。当某一种传感器模态在特定恶劣环境（如极端天气）下受限时，其他模态能够立即补位，确保系统获得做出安全驾驶决策所需的全量信息。针对纯视觉方案是否足够安全的质疑，Thirumalai 强调，评价一个自动驾驶系统不能脱离整套构建、测试与验证流程。如果有人声称其系统更优秀，就必须拿出客观可衡量的统计数据——例如在数亿英里（100 million miles）真实行驶规模下经过验证的安全记录。只要有其他架构能以相同规模证明等同的安全性，Waymo 并不抱有任何教条主义（"We're not religious"）。 ```mermaid flowchart TD accTitle: Waymo 核心安全与工程哲学架构图 subgraph CoreValues["核心指导原则"] N["北极星目标：乘客安全与极致体验<br/>（顽固坚持，绝不妥协）"] end subgraph SensorSuite["多模态冗余感知"] C["摄像头<br/>（高分辨率语义视觉）"] L["激光雷达 LiDAR<br/>（精确 3D 几何与深度）"] R["毫米波雷达 Radar<br/>（测速与极端恶劣天气穿透）"] end subgraph Validation["规模化客观验证"] M["数亿英里实际道路行驶数据"] V["端到端严格测试与安全验证流程"] end CoreValues --> SensorSuite C & L & R -->|"模态互补与安全兜底"| V V --> M ``` - **核心指导思想**：对安全体验坚持原则，对底层算法与硬件选型保持灵活演进。 - **传感器互补性**：极端天气或复杂光照下，LiDAR、Radar 与摄像头互为冗余安全网。 - **衡量标准**：拒绝空谈概念，强调以数亿英里级的大规模公开安全数据作为客观评价依据。 --- ### [05:31](https://www.youtube.com/watch?v=y40-CMBbRVE&t=331s) 生成式 AI、多模态模型与“苦涩的教训” 讨论转向当前 AI 领域的热点——生成式 AI 与多模态大模型在自动驾驶中的应用。针对 Waymo 之前发表的关于 EMMA（基于多模态大语言模型的端到端自动驾驶研究）等技术成果，Mathews 询问生成式 AI 是否代表了自动驾驶的未来。Thirumalai 明确表示，多模态模型在自动驾驶中必然占有一席之地，但目前行业仍有大量工作需要推进。 现阶段多数视觉-语言模型（VLM）主要处理视频、图像和文本，而自动驾驶行业正致力于将激光雷达、雷达等更多传感器模态融入其中。从架构本质来看，强大的“编码器-解码器”（Encoder-Decoder）框架将长期延续，VLM 只是该范式的一种具体体现。行业的发展目标是让端到端架构变得尽可能简洁高效，尽管距离完全理想的极简端到端系统可能还有数年时间，但这已是当下的重点攻关方向。Thirumalai 引用了 AI 领域的著名论断**“苦涩的教训”（The Bitter Lesson）**，指出“极其简单的模型架构配合海量训练数据”正是指引技术演进的北极星愿景。当主持人问及 Waymo 是否在 Gemini 之外测试其他多模态大模型时，Thirumalai 笑着回答没有这个必要，“Gemini 已经足够了”。 --- ### [08:04](https://www.youtube.com/watch?v=y40-CMBbRVE&t=484s) 恶劣天气泛化能力与第 6 代系统的成本削减 Waymo 业务在过去一年迅速扩张，已进入 5 个城市并正在向纽约等更多城市展开测试。面对纽约降雪等恶劣气候带来的挑战，Thirumalai 对系统的泛化能力（Generalization）充满信心。内部研发和测试表明，随着模型规模的扩大和训练数据的丰富，模型展现出极强的跨地域泛化特性。他认为恶劣天气在技术上是完全可解的问题（Tractable），只要依托优秀的模型和充分的数据收集，剩下的只是时间问题。 随后，主持人提及自动驾驶的市场需求与运营成本：Uber CEO 近期透露，在奥斯汀，Waymo 车辆的日均订单量已经超过了 99% 的人类专车司机；但乘客端发现旧金山的 Waymo 乘车价格与 Uber 专车相差无几。Thirumalai 对此解释称，Waymo 正在通过硬件换代与运营优化持续压缩成本： - **第 6 代 Waymo Driver**：相较于第 5 代系统实现了显著的成本削减。 - **传感器优化配置**：在提升传感器性能的同时，精简了传感器总数量，做到了在不牺牲安全性和感知质量的前提下大幅降低整车硬件平台造价。 - **运营效率提升**：持续削减车队日常运维和调度成本，以推动自动驾驶服务在经济效益层面的长期规模化普及。 ### [11:16](https://www.youtube.com/watch?v=y40-CMBbRVE&t=676s) 降低单位经济成本与深化外部运营合作 Srikanth Thirumalai 探讨了自动驾驶技术的成本走势。任何新技术的起步阶段，单位经济效益（unit economics）和单次出行成本往往居高不下。然而，随着软硬件平台的迭代优化、传感器套件成本的下降，以及车队运营流程的成熟，Waymo 的全链条成本必将大幅压缩，并最终直接转化为乘客端更亲民、更具竞争力的乘车资费。 面对主持人 Jessica Mathews 关于“运营成本是否包含车队维护调度与控制室”的追问，Thirumalai 明确表示这涵盖了停车场站运维（depot operations）、车队实时调度以及车辆物理维护等全部环节。 ```mermaid flowchart TD accTitle: 降本路径与生态合作结构图 subgraph Cost_Reduction["单位成本优化路径"] A["硬件与传感器成本迭代下降"] --> C["综合运营单价下降"] B["车队运维流程精细化"] --> C C --> D["乘客端资费更便宜"] end subgraph External_Partners["外部生态伙伴协同"] E["Avis 租车: 场站运营与车辆维保"] --> G["加速新城市落地推广"] F["Uber 等平台: 用户生态与调度协同"] --> G end ``` 为了以最高速度将自动驾驶服务拓展至更多乘客，Waymo 选择重度依托成熟的产业合作伙伴。例如在进军达拉斯（Dallas）等市场时与大型租车巨头 Avis 达成合作，借助其在全国范围内的车辆维保、场站管理（depot operations）与资产运营底蕴，形成优势互补。谈及与 Uber 在奥斯汀（Austin）的合作反馈时，Thirumalai 坦言目前该区域投放的无人车数量相较于整个交通出行业盘仍处于极早期，尚无法透露详细业务指标，但多项先期信号均展现出非常积极和向好的趋势。 - **核心降本路径**：迭代降低底层计算平台、传感器套件以及运维中枢的日常开销，以达成正向商业闭环。 - **Avis 合作定位**：利用租车巨头成熟的基础设施网络负责场站流转与车辆维护，避免 Waymo 重复建设繁重的重资产网络。 - **奥斯汀落地进展**：车辆体量虽小，但与 Uber 合作的早期运营与用户引流指标整体符合预期。 --- ### [13:43](https://www.youtube.com/watch?v=y40-CMBbRVE&t=823s) AI 替代人类司机争议与出行市场的长期适应路径 随着通用人工智能与自动驾驶的爆发，行业内外对于“AI 是否会剥夺网约车（Uber / Lyft）和传统出租车司机生计”的忧虑日渐升温。Thirumalai 严肃回应称，Waymo 极其重视这一问题，绝不将人类从业人员的生计视作微不足道的小事。 但从产业体量与客观物理规律来看，颠覆不会在一夕之间发生。尽管 Waymo 已经累积了惊人的 1 亿英里（100 million miles）实际自动驾驶里程，但与全球每年数以万亿计（trillions upon trillions of miles）的总出行里程相比，无人驾驶依然处于萌芽初期。在可以预见的数年之内，自动驾驶车辆无法直接全面接管主流交通体系。Thirumalai 直言人类司机在短期内绝不可能被淘汰（"I don't see human drivers being obsoleted anytime soon"），未来的演进将是一条漫长的演进之路，社会各方将拥有充裕的调整与适应窗口，以一种更加可持续、渐进的方式将新型移动形态融入日常生活。 - **出行里程对比**：Waymo 累计行驶里程达 1 亿英里，但全球交通总需求达数万亿英里，渗透率目前极微小。 - **劳动力替代周期**：技术扩散与规模化上量需要经历较长周期，社会和劳动力市场有足够空间逐步调整与适应。 --- ### [15:22](https://www.youtube.com/watch?v=y40-CMBbRVE&t=922s) 用户接受度拐点与从“恐惧”到“常态”的心理转变 针对公众对自动驾驶车辆的信任构建过程，Thirumalai 分享了旧金山等地的真实反馈。自大约两年前 Waymo 正式在旧金山取得商业收费运营许可（permits to charge for trips）以来，整个城市的公众舆论与乘客态度呈现出持续向好的显著变化。 这种心理转变的模式几乎发生在每一位初次体验者身上： 1. **上车前**：伴随着期待、兴奋，以及由于未知技术带来的明显紧张与焦虑情绪。 2. **启动后数秒内**：一旦车辆平稳起步并自然应对复杂路况，乘客便会毫无例外地迅速把“驾驶座空无一人”抛诸脑后。 3. **日常化**：整趟行程迅速演化成与普通出租车无异的平凡日常乘车体验，这种踏实与安全感在第 10 次、第 50 次乘坐后被不断强化，技术也由此彻底褪去神秘感，变成可靠的基础设施。 - **旧金山商业化历程**：获准正式面向乘客收费已运行约两年，公众对无人驾驶出租车的接受度稳步抬升。 - **信任建立的“数秒定律”**：上车初期虽有心理防线，但车辆启动短短数秒内乘客便会建立心理安全感。 --- ### [17:50](https://www.youtube.com/watch?v=y40-CMBbRVE&t=1070s) 国际化扩张测试进展与尾声 在对话尾声，Mathews 问及 Waymo 何时将无人驾驶业务推向国际市场。Thirumalai 证实团队正在全力以赴推进该目标（"working feverishly on it"）。 Waymo 此前已官方披露在部分海外城市开展先期道路测试，这些经过验证的测试城市将大概率成为其首批启动正式运营的国际站点。尽管现阶段尚无法给出一张明确的具体落地时间表，但海外拓展已经成为公司核心战略推进的关键方向。对话在轻松感谢的气氛中正式圆满结束。 ## Notable quotes > [00:04](https://www.youtube.com/watch?v=y40-CMBbRVE&t=4s) Stubborn about the experience, flexible on the details — Srikanth Thirumalai > [00:04](https://www.youtube.com/watch?v=y40-CMBbRVE&t=4s) We're not religious — Srikanth Thirumalai > [05:31](https://www.youtube.com/watch?v=y40-CMBbRVE&t=331s) The Bitter Lesson — Srikanth Thirumalai > [05:31](https://www.youtube.com/watch?v=y40-CMBbRVE&t=331s) Gemini 已经足够了 — Srikanth Thirumalai > [13:43](https://www.youtube.com/watch?v=y40-CMBbRVE&t=823s) I don't see human drivers being obsoleted anytime soon — Srikanth Thirumalai > [17:50](https://www.youtube.com/watch?v=y40-CMBbRVE&t=1070s) working feverishly on it — Srikanth Thirumalai ## Mentioned Srikanth Thirumalai, Jessica Mathews, Waymo, Fortune, Ai4, Zoox, Amazon, Tesla, Wayve, EMMA, Gemini, Uber, Lyft, Avis, Waymo Driver ## Video description Srikanth Thirumalai, VP of Engineering, Waymo Jessica Mathews, Senior Writer, Fortune [Source](https://www.youtube.com/watch?v=y40-CMBbRVE)

[0:04] Srikanth Thirumalai: Stubborn about the experience, flexible on the details [0:04] Srikanth Thirumalai: We're not religious [5:31] Srikanth Thirumalai: The Bitter Lesson [5:31] Srikanth Thirumalai: Gemini 已经足够了 [13:43] Srikanth Thirumalai: I don't see human drivers being obsoleted anytime soon [17:50] Srikanth Thirumalai: working feverishly on it

[0:04] Srikanth Thirumalai: see the Tesla boom from the outside [0:35] Srikanth Thirumalai: stubborn about the safety of our passengers [0:35] Srikanth Thirumalai: flexible in the approach [0:35] Srikanth Thirumalai: not religious [8:51] Srikanth Thirumalai: no need [8:51] Srikanth Thirumalai: bitter lesson [13:43] Srikanth Thirumalai: within a few seconds to a tea, every person... instantly forgets that there's no person in the driver's seat

[0:04] So, we are in Las Vegas right now and we [0:07] have Amazonowned Zuks that are doing [0:10] test driving down the Las Vegas strip. [0:13] And then here at the conference, we also [0:14] have Tesla downstairs that's doing some [0:17] full self-driving demos. Have you been [0:19] surveying the competition and taking any [0:21] rides yourself this trip? [0:23] >> No, I have not. Okay. I'm just here I I [0:25] drove in. I I flew in yesterday. I fly [0:27] out in a few hours today, but I did see [0:30] the Tesla boom from the outside. [0:31] >> Okay, so next time [0:33] >> maybe. [0:35] >> So hopefully some of you got a chance to [0:37] listen to talk earlier, but I want to go [0:41] ahead and dive into a couple more [0:43] details about um Whimo's AI stack and in [0:47] particular, one thing um the LAR and [0:52] radar. You talked a lot about the [0:53] importance of that. Um, [0:57] that's particularly interesting right [0:59] now because you have players like Tesla, [1:02] you have players like Wave who are [1:04] talking about a camera only system and [1:06] they're saying we don't need LAR, we [1:07] don't need radar. So can you tell us a [1:10] bit why you have decided to go the [1:12] approach of incorporating LAR and radar [1:14] and why you think that's important? [1:16] >> Yeah, I can try. Right. So, [1:20] so you know at the very core there is [1:22] this one principle that we hold dear um [1:26] at Whimo and that is we're very stubborn [1:29] about the safety of our passengers and [1:33] the experience we want to create for our [1:37] riders in the way especially in the [1:39] early days as people are trying to [1:42] understand and assess the safety of this [1:45] new piece of technology and for us that [1:48] is paramount. Safety is paramount. The [1:50] writer experience is paramount. We're [1:51] very stubborn about that. And we're very [1:54] flexible in the approach we take to get [1:57] us there. And over the years, what we [2:00] found is that the stack that we have, [2:03] the approach we're taking, which is not [2:05] just about sensors or not just about [2:07] state-of-the-art AI, it's that whole [2:09] process, right, that I just talked about [2:11] in my talk, that's really important to [2:14] scale this piece of technology and and [2:16] and sort of really reassure people that, [2:20] you know, safety is really important and [2:22] these vehicles are safe. For us, that [2:25] was the northstar. That was the guiding [2:27] principle and you know over time [2:30] technology changes you know AI software [2:33] changes uh more powerful models become [2:35] available you know it's possible to run [2:37] a lot more um you know sophisticated [2:40] models on the car uh sensor technology [2:42] is going to evolve cameras are going to [2:44] get much better going to get much [2:46] cheaper right every every year every few [2:50] months you know stuff changes and so [2:52] what we want to do is to analyze and and [2:56] see what technology exists and where [2:58] this technology is at in order to [3:01] generate the or or create this [3:02] experience that our customers want. So [3:04] stubborn about the experience, flexible [3:06] on the details. That's how I would [3:07] characterize our approach. [3:09] >> What about the LAR and radar [3:11] specifically are critical to the safety [3:12] case. [3:13] >> I think they just build I I think they [3:15] sort of complement each other. It's just [3:17] it's just that simple. They complement [3:18] each other and when one modality doesn't [3:21] work quite well the other chips in and [3:23] and provides that additional safety net [3:25] right for us to make sure that under all [3:27] conditions you know extreme weather [3:30] conditions as well that that the [3:33] complement the full complement of the [3:34] sensors that we have provides us with [3:37] the information needed to make safe [3:38] driving decisions. It's just that [3:40] simple. Now, um, if someone actually [3:43] shows us that they can achieve a 100 [3:46] million miles of driving, think about [3:49] it, a 100red million miles of driving, [3:52] you know, with these models. Sure. You [3:56] know, we've done a lot of research, [3:58] right? Uh, we're aware, right, of of [4:02] what works, what doesn't work at our [4:04] scale [4:05] >> and what we need to do to deliver the [4:07] experience. And this is what we built. [4:09] Do you think that systems that are out [4:11] on the road today without radar and [4:13] without LAR are safe for customers? It's [4:16] not the system like I said you know it's [4:18] the whole process of how you actually [4:21] build test validate right the car right [4:25] that has to be looked at you know I [4:27] think for us to actually a lot of people [4:29] say well you know oh you know u company [4:34] X or company Y is better than Whimo [4:37] because they're able to do it with [4:41] these sensors or that AI. I'm like, [4:45] let's take an objective definition of [4:47] what better means. Okay? Like when [4:50] someone says, "Oh, this stuff is more [4:51] evolved. This stuff is better." I'm [4:53] like, "Yeah, okay. By what measure? By [4:57] what objective measure?" Right? I think [4:59] when we when if if if [5:02] we actually are talking about objective [5:04] measures, then we have to look at the [5:06] statistics, right? Um our safety record [5:10] at the scale that we're driving. And [5:12] when someone actually says, "Yes, I can [5:14] match your safety and your your scale [5:18] with a different system," that's great. [5:21] You know, we'll take that. [5:22] >> We're not we're not we're not we're not [5:24] religious. [5:24] >> So other companies do not have the same [5:26] statistics you do yet. So are you saying [5:28] you can't make the comparison quite? [5:29] >> I don't know. If they have it, they [5:30] haven't shared it. [5:31] >> Okay. So one of the things that we're [5:34] spending a lot of time talking about at [5:36] this conference is generative AI. That's [5:38] on a lot of people's minds right now. [5:39] Whimo has done a lot of research so far [5:42] on multimodal models and how to [5:44] incorporate that into driving and [5:47] self-driving in particular. You've [5:49] published some research on that um on [5:51] the Emma system in particular. Yeah. [5:54] >> And um [5:56] >> the findings so far have said that [5:58] there's a lot of opportunity there, but [6:01] then there's also some things that have [6:04] to be sorted out at this point. And I'm [6:06] I'm curious to hear your perspective on [6:10] if generative AI models, multimodal [6:12] models are the future of self-driving or [6:15] if it's too early to tell. [6:16] >> I mean, they certainly have a place [6:18] right in them. There's a lot more work [6:20] that needs to be done. Um, you know, [6:23] right now, um, you know, these models, [6:26] uh, use videos, cameras, and text for [6:30] the most part, right? there's a lot of [6:32] work going on in the autonomous vehicle [6:34] industry to you know add other [6:37] modalities to them. Um yeah I I would [6:41] say the core principle of you know [6:45] having this at a very high level an [6:47] encoder decoder model and the more [6:49] powerful you can make the encoder uh you [6:52] know and the decoder uh the better I [6:54] think that will persist and you know uh [6:58] now a VLM is just one manifestation of [7:00] that right and so uh I think I think [7:03] this thing is going to persist uh but [7:05] but there's a lot more work that's going [7:07] to be needed right to uh to kind of just [7:10] make the system uh as simple as possible [7:13] and a lot of people are working to make [7:16] that that that end to end architecture [7:20] um you know as simple as possible right [7:23] we're still maybe a I don't know um a [7:26] few years away from from that but it's [7:29] the it's it's an area of intense [7:31] research [7:32] >> right and and uh I do believe right that [7:35] in the next few years um you [7:39] what what that bitter lesson says you [7:41] you all you need are simple models and [7:43] lots of training data you know that is [7:45] the northstar vision that guides all of [7:47] us [7:48] >> so you've been testing with Gemini right [7:50] have you tested with any other [7:52] multimodal models [7:54] >> um [7:56] you know I we not really there's no need [8:00] to [8:02] >> Gemini's enough [8:04] >> so Whimo's been scaling really quickly [8:07] over the last year you're now in five [8:08] cities. I believe you're testing in many [8:11] more. Um, as you start incorporating and [8:15] as you start testing in cities like New [8:17] York, which you're doing, that have more [8:21] difficult weather conditions like snow, [8:24] how is that influencing the way you [8:27] think about your safety case and your [8:29] ability to scale? [8:30] >> Yeah. So you know one of the things that [8:34] we're observing um in our in our you [8:38] know internal research uh our [8:41] development our testing um is that you [8:44] know our models are generalizing you [8:47] know quite well and and it's sort of the [8:51] reason why we're very excited um about [8:56] you know expanding to all of the cities [8:59] that we've been sort of regularly [9:01] announcing. that we're going to expand [9:02] to. Um, and I do believe that as these [9:05] models get bigger, as they as they're [9:07] trained on more data, we will start to [9:10] see more and more generalizability sort [9:13] of, you know, working in our favor. Um, [9:15] and you know, I don't know where your [9:18] question was going, but the question was [9:20] like, is is technology, right, the [9:23] biggest hurdle for some of these cities [9:25] uh with with with bad weather? [9:27] >> Yeah, [9:28] >> I don't think so. I think all of these [9:30] problems are are tractable, you know, um [9:33] with uh you know, with with with good [9:36] models and uh and and [9:39] you know, the data that we were able to [9:40] collect. So, I'm not I'm not too [9:42] concerned. It's just we just need time. [9:44] >> So, bad weather is already solvable. [9:46] >> Yeah, it is. I think it is. [9:48] >> Okay. [9:49] >> Yeah. [9:49] >> So, clearly there's a lot of demand for [9:52] driverless vehicles. the CEO of Uber a [9:55] few days ago said that um in terms of [9:58] their daily rides they're doing Whimo is [10:01] driving more than 99% of their mandriven [10:06] fleet. Um which is pretty significant [10:10] and that's specifically in Austin. [10:12] >> Oh yeah. [10:12] >> Um but one thing that I was surprised [10:14] about when I took my first Whimo ride in [10:16] San Francisco was that the cost is [10:19] actually about the same. It was about [10:21] the same price for me to take a Whimo as [10:23] an Uber in San Francisco. And I wanted [10:25] to see if you could explain that to us. [10:27] >> Well, so we're always working on [10:29] reducing the cost, right? This is, you [10:31] know, with every generation of our of [10:33] our driver. Uh we're we're trying to [10:37] shave the costs of of uh of the vehicle [10:40] and the and the compute and the sensors. [10:43] Our sixth generation driver has [10:46] significant cost savings, right, over [10:49] our fifth generation system. We have uh [10:53] not just more sophisticated sensors, but [10:55] we have fewer of them, right? And we're [10:58] able to keep reducing the cost of the [11:00] platform and the sensing stack while not [11:02] compromising on the safety and quality [11:04] that we need. And you know and we're [11:07] also working to um you know uh keep [11:11] shaving costs off our operations right [11:13] increasing efficiency of our operations [11:16] which again directly sort of you know um [11:19] u helps us lower costs with any new [11:22] technology right the unit economics the [11:25] unit unit costs are going to be high in [11:27] the early days but over time you know um [11:30] I'm very uh you know optimistic that all [11:34] of our efforts to iteratively lower [11:36] costs will lower the cost of our [11:39] operations, lower the cost of our [11:40] platform, lower the cost of sensors and [11:42] that'll sort of, you know, get [11:43] translated to to uh to better affairs. [11:46] >> So, you think it'll eventually be [11:48] cheaper? [11:48] >> Yeah, I think it'll be cheaper. [11:49] >> Yeah. When you talk about the [11:50] operational costs, are you talking about [11:52] the fleet management, the operation [11:54] rooms, all of those things? [11:55] >> All that stuff. [11:56] >> Okay. And that will be streamlined over [11:58] >> depot operations, fleet management, [12:00] fleet vehicle maintenance, that sort of [12:02] stuff. Yeah. [12:03] How do partnerships that you're doing [12:05] with companies like AVIS play into that? [12:07] I know when you're launching into [12:09] Dallas, you're planning to work with [12:10] Avis on the fleet management side. [12:12] >> Yeah, that's super important to us, [12:13] right? Because like as I said, what we [12:15] want to do is we want to bring this [12:17] service to as many customers as possible [12:18] as soon as possible. and uh and and [12:22] frankly our partners play are going to [12:25] play a big part in that because they [12:27] have some core skills right that we want [12:29] to leverage um to to expand as fast as [12:32] we can. Um we talked about fleet [12:34] management, you know, depot operations, [12:37] uh vehicle maintenance. These are all [12:39] things that some of these partners are [12:40] very very good at. you know, AA [12:42] certainly, you know, is a um is is a is [12:46] a big name in in in things like fleet [12:49] management and uh and operations. And so [12:52] just leveraging their core strengths uh [12:54] which, you know, quite frankly, you [12:56] know, complements ours and allows us to [12:58] grow sooner um feels like the right [13:01] thing to do. So, we're super excited [13:03] about these partnerships uh because [13:05] it'll allow us to bring our service to [13:07] as many um more people sooner. What are [13:11] some of the things that you've learned [13:12] so far from those partnerships? So, with [13:14] Uber, for example, in Austin, [13:17] >> um, you know, it's it's too early to say [13:19] and I don't have the I don't have the [13:21] things off the top of my head, but, you [13:23] know, things seem to be pointing in the [13:25] right direction, right? I think I think [13:26] there I think there, you know, uh, it's [13:30] just too early to say. The number of the [13:32] number of vehicles that we have, right, [13:34] in in places like Austin, etc., relative [13:36] to the market is just so small, right? [13:38] It's very early days but uh very [13:40] promising um signals. [13:43] >> Something that a lot of people are [13:45] thinking about as we talk more and more [13:48] about AI is which jobs will AI replace [13:51] and I think specifically with robo taxis [13:55] there's a pretty clear answer that it [13:57] would be taxi drivers Uber and lift [14:00] drivers [14:02] being someone who's working on that [14:04] technology in the weeds in it every day. [14:07] How are you thinking about that? [14:09] >> Well, we take all that very seriously, [14:11] right? Um, I mean, jobs and livelihoods [14:14] are not something you can sort of like, [14:15] you know, brush off to the side. Um, but [14:19] as I said again, it's very early days [14:21] even though, you know, we have driven [14:23] 100 million miles so far. You know, like [14:26] the mobility market is like trillions [14:29] upon trillions upon trillions of miles, [14:31] right? And it's not as if this thing is [14:34] going to, you know, magically sort of [14:37] scale from where we are right now to to [14:42] sort of cover large swads, if not all of [14:45] u of transportation in uh in the next [14:49] few years. This is going to take a [14:50] while, right? So I don't see I don't [14:53] see, you know, human drivers being [14:56] obsoleted anytime soon, right? we have [14:59] this is going to be a long path for us [15:01] and along the way I'm sure right we'll [15:04] find uh you know ways to um incorporate [15:08] this technology and incorporate this [15:10] this new form of mobility into our daily [15:13] lives in a more sustainable way you know [15:16] >> say it'll be it'll be it'll be a long [15:18] path [15:19] >> a long time to adapt [15:20] >> yeah I think it'll be plenty of time to [15:21] adapt [15:22] >> so on on that thread I'm curious what [15:25] you've learned in ter in terms of [15:28] customer appetite for self-driving [15:30] vehicles. Obviously, the first time [15:32] anyone goes in one, it's a very new [15:34] experience for them. Um, but then they [15:37] get used to seeing self-driving vehicles [15:38] around. If they're living in cities like [15:40] San Francisco, if they're living in [15:42] cities like Austin, they start to get [15:44] used to it. I'm wondering in what's been [15:46] your experience because Whimo's in more [15:48] cities than any other company right now. [15:50] How long does it take for customers to [15:52] start to feel comfortable or the cities [15:55] themselves to start to feel comfortable [15:56] with [15:57] >> new technology? [15:58] >> Well, you know, I've heard secondhand, [16:02] right? So, you know, I can't quote uh [16:04] numbers or stats here, but I've heard [16:07] secondhand that, you know, customer [16:10] sentiment is improving, right? Um um [16:14] since the early days, right? What was [16:16] it? It was about two years ago when we [16:19] got the permits to to charge for trips [16:23] in San Francisco, right? And from then [16:25] to now, you know, I've heard that, you [16:28] know, people are view more people are [16:30] viewing Whimos positively [16:33] >> than at the start. Uh, you know, I don't [16:35] have the specific numbers with me, but [16:37] it doesn't surprise me, right? Because [16:40] when when you take a ride in one of [16:42] these, uh, initially, you know, you may [16:45] be anxious, right? Um you may be excited [16:48] right you've seen videos of people you [16:51] know um in these whimos you know getting [16:54] rides and so there is some anticipation [16:57] uh perhaps nervousness but within a few [17:00] seconds to a tea every person that I [17:04] have given a right to in a whimo [17:07] instantly forgets that there's no person [17:10] in the driver's seat and then they just [17:12] become it's just it just becomes another [17:15] regular you a cab ride, right? Um, and [17:19] that feeling of assuredness within a few [17:22] seconds sort of, you know, uh, I think [17:24] is very telling and it's not just the [17:26] first ride, it's, you know, the 10th [17:29] ride, the 50th ride. So, it doesn't [17:31] surprise me, right? That initially [17:33] there's apprehension because this is new [17:35] technology and you don't know what's [17:37] going to happen, but as you take more [17:39] and more rides, it becomes normal. And [17:42] so it doesn't surprise me that customer [17:44] sentiment over the over the years, the [17:46] two years that we've operated in SF, you [17:48] know, is going up. [17:50] >> All right, I'm getting the signal that [17:51] we're almost out of time, but one more [17:53] question. When is Whimo going to go [17:55] international? [17:57] >> Um, we're working on it. [18:00] >> Anything more specific? [18:02] >> We've announced, right, that we've [18:03] started testing in a few international [18:04] cities, right? And so those are likely [18:06] going to be the first cities that we [18:07] operate in, but uh we're working [18:09] feverishly on it. and um we're super [18:12] excited about about it. [18:14] >> So, no timeline just yet. [18:15] >> No timeline just yet. Sorry, [18:17] >> Con. Thank you so much. Really [18:18] appreciate it. [18:19] >> Thank you.