Toyota, AV Makers See Global Driverless Sales Capped at 25% by 2041

丰田与自动驾驶厂商预计到2041年全球无人驾驶汽车销量上限为25% 地方监管碎片化、全球车道基础设施匮乏以及车辆准入认证壁垒，将在未来15年内把全自动驾驶车辆限制在新车销量的四分之一以下。 本场讨论由Automotive Ventures的Steve Greenfield主持，汇集了来自May Mobility、Tensor Auto和丰田研究院（TRI）的高管，共同探讨汽车AI的技术与商业化落地。对话深入探讨了端侧多模态感知、混合推理架构以及制造业中的扩散模型，同时涉及TRI的“Driving Sensei”等以人为本的系统。与会嘉宾还审视了运营层面的现实问题，包括L4级制造商责任界定、动态保险定价，以及经销商维修工位向自动驾驶维护中心的转型。最后，发言嘉宾给出了前瞻性预测，估计在15年内全球将有10%至25%的乘用车实现完全自动驾驶，并指出监管碎片化而非技术才是最主要的瓶颈。 公众对 AI 的热情与行业从业者存在 40 个百分点的信任鸿沟，丰田等车企正通过双轨投资与自研布局，并开发“Driving Sensei”等 AI 领航员以协同教练人类驾驶员而非彻底取缔人类。 (6:48) May Mobility 将离线学习模型与在线实时推理相结合，通过直接从感知到的周围车辆行为中学习，仅消耗约 2 万英里道路数据即实现了应对复杂交通场景能力的数倍提升。 (13:30) Tensor Auto 针对 L4 级私人自动驾驶汽车推出搭载超 100 个传感器的从零设计硬件平台，并将全部车内生物识别与多模态自然语言指令限定在车载边缘计算端处理，彻底隔绝云端隐私风险。 (17:13) 丰田研究院（TRI）通过仅 60 至 100 次人类示教训练出用于装配制造的扩散策略（Diffusion Policy），这一物理具身 AI 成果已剥离创立 Walden Robotics 并直接部署于丰田工厂。 (18:55) 针对 15 年后全球全自动驾驶汽车在新车销量的渗透率，嘉宾给出了 10% 至 25% 的预测区间，普遍认为阻碍普及的关键已非技术本身，而是碎片化的区域法规与车规准入壁垒。 (44:54)

Autonomous vehicle developers are transitioning from rigid automation toward multimodal, human-collaborative artificial intelligence, seeking to bridge a steep public trust gap while navigating commercial integration, dynamic liability models, and uneven global infrastructure. Hybrid Reasoning and Multimodal Edge Perception (13:30) Autonomy stacks are moving away from brittle, rules-based programming toward hybrid architectures that pair offline learned models with real-time online reasoning. Jacob Crossman explains that May Mobility expanded its driving capabilities using only 20,000 miles of training data by designing models that learn from observing surrounding vehicles, resolving split-second uncertainty through on-the-fly simulation 'tournaments' that default to cautious maneuvers. At the physical layer, Toyota Research Institute (TRI) adapted generative diffusion policies from large language models to robotics, mastering complex manipulation tasks with only 60 to 100 human demonstrations—a breakthrough published in Science Robotics and commercialized through spinout Walden Robotics. Perception hardware is similarly diversifying beyond standard computer vision. Tensor Auto engineered a purpose-built robocar featuring over 100 sensors—including 5 LiDARs, 11 radars, 22 cameras, and over 20 microphones—to capture environmental context like emergency sirens and acoustic tire friction. To protect consumer biometrics and voice interactions, Tensor executes all natural language and biometric processing locally on vehicle edge processors rather than the cloud. Avinash Balachandran notes that integrating auditory friction data directly into dynamic chassis control allows vehicles to maintain control at their physical handling limits, mirroring expert human driving. Bridging Public Mistrust and Augmenting Drivers (6:48) A critical barrier to autonomous adoption is public hesitation. Balachandran cites survey data revealing a 40-point trust deficit: while 57% of AI practitioners report excitement about AI, only 17% of the general public shares that view. To build trust, TRI prioritizes human-machine collaboration over outright driver replacement. Its 'Driving Sensei' initiative develops AI co-pilots that actively coach motorists, targeting vehicle crashes as the leading cause of death among teenagers globally by augmenting human competence rather than removing agency. Generational adoption is also reshaping passenger safety standards. Amy Luca notes that parents in cities like San Francisco increasingly favor autonomous ride-hailing over human-driven rideshare for unaccompanied minors to avoid placing children with strangers. Manufacturers are addressing this segment with edge-case controls, such as interior facial recognition lockouts that prevent children from exiting at intersections before reaching authorized drop-off points, while adapting to emerging municipal frameworks that restrict solo travel for children under eight. Reimagining Dealerships, Fleets, and Urban Land (21:56) Autonomy is poised to disrupt retail distribution networks and municipal infrastructure. Rather than bypassing franchise dealerships, Tensor Auto operates within dealer channels to automate service cycles: vehicles can independently drive themselves to dealership bays overnight for sensor calibration and maintenance. Dealerships can also dispatch autonomous vehicles to prospective buyers' driveways for self-guided sales demonstrations, removing geographic lot limitations. For shared passenger transport, May Mobility relies on specialized fleet management partnerships to maintain complex compute and sensor stacks. At the municipal scale, shared autonomous fleets could transform urban land allocation. Crossman notes that between 15% and 20% of urban real estate in cities like Ann Arbor is dedicated to parking; widespread on-demand transit could repurpose these spaces for green areas and housing. Concurrently, commercial freight is automating rapidly due to clear economic returns, with European freight operators deploying autonomous multi-truck platooning on highways to reduce aerodynamic drag and fuel consumption. Operational Liability and Dynamic Insurance Pricing (36:48) Commercial deployment requires establishing unambiguous legal liability when autonomous systems fail. Luca explains that Tensor Auto and insurance broker Marsh separate responsibility by operational mode: during Level 4 autonomous travel—where pedals and steering wheels retract or are absent—passengers carry zero operational control, and Tensor absorbs 100% of vehicle liability. Conversely, when a user engages manual or Level 2 driving controls, legal liability returns entirely to the human driver. This division facilitates dynamic, usage-based insurance underwriting. Because onboard vehicle computers log operational states mile by mile, insurers can split monthly premiums between manufacturer-backed autonomous miles and driver-operated miles. Luca anticipates that Level 4 miles will be substantially cheaper to insure due to lower software accident rates compared to human error, creating economic incentives for autonomous transit while industry players continue sharing safety data to guide underwriting standards. Looking ahead 15 years, panelists anticipate that fully autonomous vehicles will represent a modest share of the 90 million passenger vehicles sold annually worldwide—ranging from Amy Luca's conservative forecast of 10% to Avinash Balachandran's projection of under 20% and Jacob Crossman's 25%. The panel concludes that technological readiness is no longer the primary hurdle; instead, fragmented municipal regulations, disparate international homologation requirements, and regions lacking baseline roadway infrastructure will dictate the speed of global deployment.

来自May Mobility、Tensor Auto和丰田研究院（TRI）的高管指出，人工智能正在从感知决策、舱内交互到售后生态全方位重塑汽车工业，但全自动驾驶的落地进程仍面临技术范式分歧、商业模式差异、法律责任归属以及全球监管碎片化等多重制约。 算法演进：生成式模型与在线推理融合 (13:30) Jacob Crossman指出，May Mobility在过去18至24个月中实现了将类似大语言模型（LLM）的Transformer架构与实时决策系统深度融合的突破。该系统不仅依靠自车轨迹，更通过传感器学习环境中其他车辆的交互行为，仅需约2万英里的驾驶数据便实现了规划能力的跨越。针对信号灯突变等突发不确定性场景，系统通过在线模拟运行内部轨迹“锦标赛”，在歧义瞬间主动采取减速避让等保守策略，待状态明朗后即刻收敛回平顺的拟人驾驶节奏。 在基础模型与控制上，Amy Luca说明Tensor Auto通过真实道路视频与海量合成极端场景仿真训练Tensor World世界模型。Avinash Balachandran则介绍丰田研究院（TRI）发表于《Science Robotics》的扩散策略（Diffusion Policies）研究：机械臂仅凭60至100次人类远程示教即可掌握复杂动作，该成果已通过衍生企业Walden Robotics入驻丰田工厂。此外，TRI还在探索将轮胎摩擦等音频特征融入底盘动态控制，以拓展车辆极限姿态下的安全边界。 商业路径：私家车与共享车队的分化 (21:56) 在商业化推进路径上，行业呈现出显著分化。Amy Luca表示Tensor Auto主攻面向个人的L4级自动驾驶私家车（Tensor Robocar），正加速在迪拜交付，满足高净值家庭对未成年子女和宠物的接送需求。该车采取全新正向研发，彻底取消了踏板和方向盘，依托车内多智能体系统与用户日程联动，将载具转化为具备主动建议路线和优化能耗的个人具身伴侣。 与个人私家车路线相对，May Mobility聚焦微公交与共享出行，已落地亚特兰大、安娜堡等城市，并与Uber、Lyft及Grab合作。Jacob Crossman指出，共享车队能深刻释放城市空间，安娜堡目前约15%至20%的城市土地被停车场占据，无人车队普及可将其转变为公园与住宅。此外，干线物流重卡因工况相对标准化且具备削减人力与油耗的明确投资回报率（ROI），加之欧洲等地推行的自动跟车编队技术（Platooning），商用货运的无人化落地速度快于高密度城市客运。 舱内交互：端侧多模态感知与人机共驾 (17:13) Tensor Auto在车身内外布置了由5个激光雷达、11个毫米波雷达、22个摄像头和20多个麦克风组成的超100个传感器矩阵，支持手势指引与自然语言结合的多模态指令。针对车舱持续采集生物识别与音频信息的隐私风险，Amy Luca强调其系统采用类似iPhone的本地化边缘计算架构，数据在车端闭环处理而不强制上传云端，构筑了严格的数据安全边界。 丰田研究院则提出了以“放大人类能力”为核心的人机共驾范式。Avinash Balachandran指出，交通事故仍是全球及美国青少年的首要死因，TRI开发的“Driving Sensei”智能副驾旨在长期训练与辅导驾驶员习惯，而非单纯剥夺人类驾驶权。针对行业面临的信任鸿沟——调研显示57%的AI从业者对技术前景感到兴奋，而普通公众的兴奋度仅为17%——丰田认为通过人机协同提升驾驶安全是建立跨群体信任的关键。 产业重塑：经销维保变革与动态保险定价 (36:48) 自动驾驶正深刻重塑汽车售后运营与经销网络模式。Avinash Balachandran与Amy Luca指出，L4级车辆可实现夜间自主前往服务中心完成日常维保、传感器标定和清洁，大幅缓解白天车间的人力瓶颈；在销售端，车辆还能自主上门为潜在买家提供无接触展示与试驾，打破实体展厅的地域限制。在May Mobility的运营中，复杂的高阶算力与传感器维护则交由专业车队服务商协同打理。 在法律责任界定方面，Tensor Auto与保险机构Marsh达成合作机制：当车辆处于L4级自动驾驶且方向盘收起时，乘客免责，由主机厂承担全部事故责任；而在L2级人工驾驶模式下，责任依然由人类驾驶员承担。这种责任划分催生了动态里程保险，由于L4级事故率远低于人类驾驶，系统记录并区分两类驾驶里程，能够显著降低自动驾驶模式下的保险费用。 协同网络与十五年普及率的现实博弈 (44:54) 在车际与路网协同方面，丰田研究院指出传统专用硬件面临装车门槛导致的“网络效应”困境，行业正全面转向蜂窝车联网（C-V2X）。这一架构不仅能即刻向后方车辆广播匝道黑冰等局部险情、协调路口车速消除走停波动，还能与行人手机通信预测通行意图，构建多智能体动态交通协商生态。 对于未来15年全球每年约9000万辆新车中全自动驾驶渗透率的预测，三位专家均持审慎态度。Jacob Crossman预测为25%，认为货运与高频车队将主导渗透；Avinash Balachandran预测低于20%，指出丰田全球市场中大量地区缺乏基本的道路标线与车道规范；Amy Luca则将预期从最初的20%至25%下调至约10%，强调未来技术可行性已非瓶颈，全球碎片化的地方法规、国家车辆准入审批（Homologation）及行政阻力将是放缓落地的决定性因素。 总体而言，尽管自动驾驶技术在多模态大模型、端侧推理以及售后生态协同上取得了关键突破，但高阶系统的落地节奏深受商业落地场景、区域基础设施完备度及跨司法管辖区法规准入的严重制约。自动驾驶从技术验证迈向真正意义上的全球规模普及，仍有赖于法律责任规范的成熟以及公众信任赤字的有效弥合。

## Background 本内容由频道 Ai4 于 2026 年 8 月 22 日发布在 YouTube 平台上，记录了一场题为《AI 赋能的车辆：更智能、更安全与全面互联》的圆桌论坛。根据页面简介，来自顶尖汽车制造商（丰田研究院）和前沿初创企业（May Mobility、Tensor Auto）的行业专家与 Automotive Ventures 的主持人齐聚一堂，共同探讨 AI 如何重塑驾驶体验、车队安全与工程周期，并将多传感器数据转化为实时洞察以构筑真正的自动驾驶未来。 ## TL;DR 该研讨会深入探讨了物理具身智能（Embodied AI）与自动驾驶在乘用车、商用货运及工业制造场景中的全面落地。来自丰田研究院（TRI）、May Mobility 和 Tensor Auto 的专家拆解了混合决策架构、大模型驱动的车载智能体、超百颗传感器的端侧隐私设计以及机器人扩散策略等关键突破。对话进一步延伸至车联网落地瓶颈、传统汽车经销商体系的重塑、L4 级别的事故责任界定与动态保费机制。最后，嘉宾们一致指出未来 15 年制约 L4 级新车普及（预测渗透率在 10% 至 25% 之间）的最大障碍并非技术成熟度，而是碎片化的全球监管政策与整车安全认证门槛。 ## Atmosphere & style 现场氛围极其开放且富有极客思辨色彩，主持人打破常规、鼓励观众全程即兴插话，激发出多轮关于儿童乘车安全、极端警报场景与经销商转型的深度问答。三位嘉宾在代表传统整车巨头、前装量产新势力与车队算法方案商的不同视角下坦诚交锋，既有对“具身智能生活管家”的乐观畅想，也有对公众信任鸿沟与繁琐法规的务实审慎。整个探讨节奏紧凑且干货密集，在尾声轻松幽默的“100 美元渗透率赌局”中迎来高潮，给听众带来兼具技术深度与现实商业穿透力的思想碰撞体验。 ## Key takeaways - [05:03](https://www.youtube.com/watch?v=eLgm-nWKdMU&t=303s) 公众与从业者之间存在巨大的 AI 认知鸿沟（57% 专业人员对 AI 兴奋，但普通公众仅有 17%），车企与技术方案商正通过多元路径探索技术落地与社会信任构建。 - [06:43](https://www.youtube.com/watch?v=eLgm-nWKdMU&t=403s) 汽车正从单纯的代步工具演化为“具身智能体（Embodied AI）”，底层多智能体协同运行，实现主动日程管理、车辆能耗优化及自然语言多模态交互。 - [10:36](https://www.youtube.com/watch?v=eLgm-nWKdMU&t=636s) 丰田提出“放大人类能力”的双轨 AI 战略，通过“驾驶导师（Driving Sense）”等项目利用自动驾驶 AI 指导并纠正新手司机危险操作，在保留人类驾驶乐趣的同时提升行车安全。 - [12:51](https://www.youtube.com/watch?v=eLgm-nWKdMU&t=771s) May Mobility 研发出感知与决策双轮驱动的混合决策系统（Hybrid System），仅依托约 2 万英里的多智能体交互数据即实现了应对复杂长尾极端工况能力的量级跃升。 - [15:12](https://www.youtube.com/watch?v=eLgm-nWKdMU&t=912s) Tensor Robocar 采用围绕 100+ 传感器矩阵正向研发的原生无人车架构，并坚持所有语音与生物特征数据在车端边缘计算本地处理，构建严格的隐私隔离防护体系。 - [18:55](https://www.youtube.com/watch?v=eLgm-nWKdMU&t=1135s) 丰田研究院取得“扩散策略（Diffusion Policies）”具身智能突破并分拆 Walden Robotics，实现了仅需人类数十次演示即可自主接管工厂复杂制造流程的人机协同模式。 - [21:56](https://www.youtube.com/watch?v=eLgm-nWKdMU&t=1316s) 自动驾驶技术将深度重构城市空间与文旅出行，随叫随到的车队有望释放城市 15%~20% 的停车场用地用于绿地与住宅建设，并将枯燥通勤转化为高价值生产力与生活时间。 - [24:42](https://www.youtube.com/watch?v=eLgm-nWKdMU&t=1482s) 干线与中途自动驾驶重卡凭借立竿见影的降本增效商业闭环，落地速度在多维度上超越乘用车，并催生了卡车紧密编队行驶与智能化车队调度新生态。 - [26:17](https://www.youtube.com/watch?v=eLgm-nWKdMU&t=1577s) 无人驾驶在家庭用车场景中展现出更高安全性与接送需求，但面临单人承运低龄儿童等严苛政策限制，促使车企采用“可开可坐”折叠方向盘等双模方案过渡。 - [30:27](https://www.youtube.com/watch?v=eLgm-nWKdMU&t=1827s) 车联网正在从受限于网络效应瓶颈的传统专用硬件 V2X，转向基于泛在蜂窝网络的多智能体协同（车端 Agent ↔ 行人手机 Agent 动态协商意图）。 - [32:16](https://www.youtube.com/watch?v=eLgm-nWKdMU&t=1936s) AI 与自动驾驶深度重塑汽车经销商业务：车辆可于夜间自主前往门店进行传感器校准维保，甚至能在预约时间“自主开到买家门口”进行自我推销演示。 - [36:48](https://www.youtube.com/watch?v=eLgm-nWKdMU&t=2208s) L4 级自动驾驶厘清了责任归属（L4 模式下车厂承担全部法律与经济责任），并催生基于 AI 接管里程比例的动态车险定价体系（L4 里程保费显著更低）。 - [40:13](https://www.youtube.com/watch?v=eLgm-nWKdMU&t=2413s) 面对黄灯临界点等高度不确定性场景，自动驾驶系统通过内部微型模拟评估多种备选方案，在不确定性骤增时采取保守避让策略，保障行车决策鲁棒性。 - [41:40](https://www.youtube.com/watch?v=eLgm-nWKdMU&t=2500s) 多模态声音感知（外部麦克风识别警笛、轮胎声音监测路面附着力）结合 5G 联网与 LLM 上下文验证，成为解决突发警报与边缘极端工况的关键补充。 - [44:49](https://www.youtube.com/watch?v=eLgm-nWKdMU&t=2689s) 业内预测 15 年后全球 L4 级全自动驾驶新车渗透率约在 10%~25% 之间，制约普及的核心瓶颈已转向跨国法规碎片化、整车安全认证（Homologation）流程及固有新车更换周期。 ## Detailed walkthrough ### [00:04](https://www.youtube.com/watch?v=eLgm-nWKdMU&t=4s) 开场破冰与互动规则：让观众随时打断提问 主持人 Steve Greenfield（来自 Automotive Ventures）在座无虚席的小会议厅向观众致意。他幽默地调侃道，大家能够历经周折找到威尼斯人/帕拉佐（Palazzo）酒店这个偏远角落的会议室，并且一直坚持到大会最后一节，实在非常不容易。为此，他决定打破传统的圆桌流程——不把问答环节留到最后死板的几分钟，而是鼓励台下听众随时举手打断发问。 Steve 承诺自己会全程扫视全场，只要有人举手就立刻接入问题。在正式抛出技术议题之前，他先请三位背景截然不同的嘉宾依次介绍自己、背后的公司以及当前推进的核心业务。 --- ### [00:45](https://www.youtube.com/watch?v=eLgm-nWKdMU&t=45s) 嘉宾自我介绍：全栈自动驾驶、私人 Robotaxi 与汽车巨头的未来探索 圆桌嘉宾覆盖了主机厂、初创软硬件平台与学术级研究院三种视角。首先发言的是 May Mobility 的自动驾驶高级副总裁 Jacob Crossman。Steve 特别感谢了 Jacob 刚刚在上一个时段带来的精彩主题演讲，并极力推荐大家去看演讲中展示的车辆运行视频。 ```mermaid flowchart TD accTitle: 自动驾驶嘉宾背景分布与战略方向 subgraph OEM["全球最大车企研究院"] T["Toyota Research Institute (TRI)<br/>Avinash Balachandran"] T --> T1["人机协同物理具身智能"] T --> T2["TRI 风险投资 (如注资 May Mobility)"] T --> T3["Woven by Toyota 负责商业化落地"] end subgraph May["L4 级全栈方案与车队运营"] M["May Mobility<br/>Jacob Crossman"] M --> M1["线上推理 (Online Reasoning)"] M --> M2["离线学习 (Offline Learning)"] M --> M3["网约车伙伴 (Lyft, Uber, Grab, Choco-Choco)"] end subgraph Tensor["前装量产个人无人车"] A["Tensor Auto (原 AutoX)<br/>Amy Luca"] A --> A1["全球首款私人 Robotaxi (Tensor Robocar)"] A --> A2["前装量产：100+传感器 (5激光雷达/11毫米波/22相机)"] A --> A3["中东迪拜率先部署交付"] end ``` Jacob 介绍，May Mobility 总部位于密歇根州安娜堡（Ann Arbor），目前已在多座城市展开无人驾驶车队部署，包括亚特兰大（Atlanta）、阿灵顿（Arlington）、明尼阿波利斯（Minneapolis）以及安娜堡总部周边。公司与 Lyft、Uber 建立了深入合作，近期还宣布了与 Choco-Choco 的商业交易，并且东南亚出行巨头 Grab 也是其战略投资方。Jacob 统领公司的技术研发，日常除了处理技术攻坚和风险控制外，最让他兴奋的是深入挖掘 AI 技术的底层结合点——尤其是如何将 May Mobility 独有的“线上实时推理”（Online Reasoning）与“离线学习”（Offline Learning）融为一体，从而推动自动驾驶技术的大规模安全扩展。 随后发言的是 Tensor Auto 首席营销官 Amy Luca。Steve 调侃她可能是全场飞行里程最多的人（Amy 笑称自己是美联航的百万里程会员，家里还有只小狗在眼巴巴等她回去）。Amy 介绍道： - Tensor Auto 拥有约 10 年的技术沉淀，其前身正是行业熟知的 AutoX； - 大约一年前，公司推出了定位为“全球首款个人专属无人车”的 Tensor Robocar，目前正全力推进在阿联酋迪拜的交付落地。虽然当下受到中东地缘政治局势的一些客观挑战，但公司目标依然是争取在今年年底前，让首批属于私人车主的无人驾驶汽车真正停进个人车库； - 该车完全从零正向设计研发（Ground-up），彻底告别了改装车（Retrofit）路线，整车搭载了超过 100 个传感器，包括 5 颗激光雷达（LiDAR）、11 颗毫米波雷达（Radar）和 22 颗高清摄像头； - 车辆主打“AI Agentic（智能体化）”理念，内置自然语言处理系统，乘客可以像同真人交谈一样直接对车下达复杂指令。Amy 负责全盘的市场营销与商业化落地推进。 最后介绍的是丰田研究院（Toyota Research Institute, 简称 TRI）副总裁 Avinash Balachandran。作为全球按销量计最大的汽车制造商旗下的先锋研究机构，TRI 成立已有约 10 年时间。最早 TRI 聚焦于全自动驾驶（FSD）核心技术，后来该商业化业务剥离成立了 Woven by Toyota；而当前的 TRI 则迈向更广阔的未来前沿领域： - 深入探索先进驾驶（Advanced Driving）、机器人学（Robotics）以及能源与新型材料； - 核心愿景不再局限于机器单向替代人类，而是转向“人与智能物理 AI 系统（无论是车辆还是机器人）如何深度协作共存”； - 旗下设有风险投资基金，May Mobility 便是其重点投资的组合公司之一，双方在生态上协同探索。 --- ### [05:03](https://www.youtube.com/watch?v=eLgm-nWKdMU&t=303s) 17% 与 57% 的认知撕裂：丰田眼中的人机协作与全景赋能 Avinash 在现场做了一个即兴举手表决，询问台下有多少人对当下的 AI 应用感到“无比兴奋”，又有多少人感到“极度恐慌”。现场反馈呈现出鲜明的分化。 Avinash 随即列举了一组全美统计数据：在 AI 领域的专业从业者（Practitioners）中，有高达 57% 的人对技术前景感到极其兴奋；然而在普通公众中，这一比例竟骤降到了区区 17%（Steve 特别打断追问是 17 还是 70，Avinash 明确确认是 "17%"）。作为一家每年在全球销售超过 1,100 万辆汽车的巨头企业，丰田对这一巨大的社会认知鸿沟感到尤为审慎： > “根本没有人真正掌握标准答案。我们唯一能做的，就是通过尽可能多的路径和手段去反复探索。” 丰田选择的应对之道不仅是通过资本连接外部创新企业，在内部更聚焦于如何用 AI“放大人的能力”（Amplify the person），并将这种赋能渗透进汽车全生命周期——涵盖工程研发、外观造型设计、整车制造、终端销售乃至售后保修的每一个环节。 --- ### [06:43](https://www.youtube.com/watch?v=eLgm-nWKdMU&t=403s) 未来半年至一年的核心战役：从代码辅助到“物理具身智能体” Steve 将话题拉回现实，向台下嘉宾抛出了第一个硬核问题：在未来 6 到 12 个月内，各家公司最核心、排在第一顺位的 AI 战略举措到底是什么？ ```mermaid flowchart LR accTitle: 核心 AI 战略推进方向对比 subgraph MayMobility["May Mobility 核心聚焦"] M_A["路线拓展与车队规模化"] M_B["端到端大模型 + 实时在线推理"] M_C["AI 辅助全流程软件研发<br/>(让非技术人员直接生成功能)"] end subgraph TensorAuto["Tensor Auto 核心聚焦"] T_A["迪拜前装量产商业化交付"] T_B["多 Agent 协同架构<br/>(动力控制 / 个人行程助理)"] T_C["全场景日历、生活与即时通讯联动"] end ``` May Mobility 的 Jacob 指出，公司的头号目标始终是在更多真实场景和路线上拿掉安全员、扩大运营网络。实现这一目标的技术底座是双向的：一方面将 Transformer、大语言模型（LLM）等前沿架构同其独特的在线推理能力更深地融入行车技术栈；另一方面则是把生成式 AI 大规模引入日常的工程软件开发中。 Jacob 举了一个充满极客氛围的生动花絮：在安娜堡总部举办的为期数天的全员集会活动“May Days”黑客松比赛中，团队挑战为车载中控屏开发一款可玩的新车载游戏。最终在 Jacob 所在的战队中，亲自动手利用 AI 编写并跑通整个游戏代码的，竟然是他们的公关参谋长（Chief of Staff for Communications）。这彻底颠覆了他对生产力工具边界的认知： > “过去完全无法直接向工程工具库添砖加瓦的产品或运营职能同事，现在借助 AI 工具，竟然拥有了直接产出代码与工程能力的可能性。” Tensor Auto 的 Amy 则直言，他们当前的主要精力正全力扑在迪拜商业化交付冲刺上，而在技术底层，公司的绝对重心是全面强化车辆的“Agentic（智能体）属性”： - 现有的绝大多数 L4 级无人车（如 Waymo）在人机交互上依然高度依赖手机 App 或屏幕物理点击输入，乘客与汽车缺乏自然的双向语言沟通； - Tensor 的愿景是将无人车塑造为一个“长着轮子的专属生活管家”（Rolling Personal Assistant）。它可以深度读取用户的手机日历与生活习惯：早晨车辆会自动给用户发短信提醒：“你今天几点有重要会议，但当前路况比平时拥堵，必须提前出发；顺便问一下，路上需要为你预留停靠星巴克买咖啡的时间吗？” - Amy 强调，行业如果仅仅停留在讨论“自动驾驶汽车”（Autonomous Vehicles），格局就狭隘了。汽车本质上正在蜕变为**具身智能（Embodied AI）**。在车辆底层，多个子 AI Agent 正有条不紊地协同分工——有的在专注调控电池与动力系统，有的则负责行程动态管理，使车辆从单纯的“A点到B点代步机”跃升为生活中不可或缺的超级助理。 在 Amy 阐述完毕后，Avinash 正式准备拆解丰田研究院在未来的三层核心战略。 ### [10:36](https://www.youtube.com/watch?v=eLgm-nWKdMU&t=636s) 丰田的双轨 AI 战略：赋能人类与“驾驶导师”计划 丰田研究院（TRI）副总裁 Avinash Balachandran 阐述了丰田在业务中全面布局 AI 的双重路径。第一条路径是产品端的直接融合，即将智能体（Agents）、自动驾驶技术以及各类交互功能更无缝地植入车辆中；第二条路径则是重塑业务流程，例如在设计环节利用 AI 工具大幅压缩车辆工程研发周期，同时丰田北美汽车公司（Toyota Motor North America）的团队也在积极探索利用 AI 高效处理保修与召回信息，全方位提升运营效率。 然而，从更高维度审视，丰田核心关注的是 AI 时代下的人类主体性（Human Agency）、自主权与存在价值，并思考如何通过 AI 举措赋能人类。Avinash 指出，业内关于自动驾驶的大多数讨论都聚焦于“彻底取代驾驶员”，这固然有其适用场景；但作为一家年销量达 1100 万辆汽车的全球车企，丰田绝大多数用户所处的环境短期内根本无法覆盖全自动驾驶设施。丰田希望尽早将前沿技术红利带给大众，因此提出了一种“放大驾驶员能力”（Amplify the Driver）而非“替代人类”的新范式。 这一理念具体落地为“驾驶导师”（Driving Sense）项目。由于车祸是美国及全球多地青少年群体的头号死因，TRI 尝试将高性能自动驾驶 AI 系统转化为安全教练，在日常行驶中循序渐进地指导、纠正新手司机的危险操作，在不剥夺人类驾驶乐趣与主体性的前提下，全方位提升行车安全性。 - **AI 赋能业务两翼**：在产品端无缝集成智能体与自动化能力；在运营端利用 AI 加速工程设计周期，并智能化分析保修与召回数据。 - **“驾驶导师”（Driving Sense）项目**：针对青少年车祸高发痛点，利用顶尖自动驾驶算法作为实时辅助与教学系统，帮助驾驶员提升技能而非直接接管控制权。 --- ### [12:51](https://www.youtube.com/watch?v=eLgm-nWKdMU&t=771s) 过去两年的 AI 突破与 May Mobility 的混合决策架构 主持人 Steve Greenfield 结合自身经历幽默地提到，自己前几天在亚特兰大开车时忽然意识到，年轻时在路上冒过无数不必要的风险，“按理说我能活到现在真是个奇迹”，如果车辆能主动防止年轻人做蠢事将极具价值。随后，他向嘉宾提问：回顾过去 18 至 24 个月，各位所在团队最具代表性、最引以为傲的具体 AI 成果是什么？ May Mobility 自主研发高级副总裁 Jacob Crossman 回顾了团队将大语言模型等生成式 AI 技术迁移至自动驾驶决策的突破。在 2023 年，May Mobility 的系统主要依靠 AI 进行环境感知，而底层的路径决策与规划仍依赖传统规则引擎，且仅能在小范围区域内运行。过去两年间，工程团队以惊人的速度将其升级为感知与决策双轮驱动的“混合系统”（Hybrid System）。 ```mermaid flowchart TD accTitle: May Mobility 混合自动驾驶系统演进 subgraph 2023传统架构 ["2023 早期架构"] A1["传感器数据输入"] --> B1["AI 感知层"] B1 --> C1["传统规则决策引擎"] C1 --> D1["小范围有限运行"] end subgraph 混合决策架构 ["最新混合系统 (Hybrid System)"] A2["自车 2 万英里数据 + 周边车辆多智能体轨迹"] --> B2["基础感知与世界交互模型"] B2 --> C2["AI 策略生成 + 鲁棒决策引擎"] C2 --> D2["复杂道路场景能力指数级跨越"] end ``` 令团队惊喜的是，该系统仅凭借约 20,000 英里（约 3.2 万公里）的实际行驶数据就完成了高效训练。其突破性的关键在于全新的多智能体学习机制：系统不仅从自车的驾驶行为中汲取经验，还会通过高精度传感器实时捕获并学习道路上周围其他车辆的驾驶轨迹与博弈行为。在成熟决策引擎的配合下，该架构使车辆应对复杂极端工况的能力实现了“数个数量级”（orders of magnitude）的跃升，并为下一阶段的指数级迭代奠定了基础。 - **极低数据量驱动高鲁棒性**：仅依托约 2 万英里自车及周边环境多目标交互数据，即完成混合策略模型训练。 - **决策系统能力跨越**：将 AI 深度引入路径规划与策略生成，系统应对长尾复杂场景的能力获得量级提升。 --- ### [15:12](https://www.youtube.com/watch?v=eLgm-nWKdMU&t=912s) Tensor 原生无人车：百余传感器、全车无盲区与端侧隐私计算 Tensor Auto 首席营销官 Amy Luca 介绍了其团队在全无人驾驶领域的重大跨越。Tensor 曾在中国通过改装车队运营了约 1,000 辆 Robotaxi，积累了丰富经验；六年前团队转入隐身研发（Stealth Mode），创始人决定回到美国“从零起步”（Start from scratch）打造真正的原生自动驾驶汽车，并于去年正式发布了 Tensor Robocar。 该车在工业造型上完全打破常规，采用“围绕传感器套件进行整车设计”（Car designed around the sensor suite）的颠覆性理念，通过 AI 协同设计彻底消除了车身内外的感知盲区。在硬件配置上，Tensor Robocar 堪称传感器怪兽，全车搭载了超过 100 个传感器。面对 L4 级自动驾驶取消方向盘和油门刹车踏板所带来的用户信任挑战，Tensor 构建了世界基座模型（World Foundational Model），利用海量真实路况片段与高保真合成仿真数据互相补充，覆盖远超人类驾驶员一生所能经历的边缘极限工况（Edge Cases），并秉承开源精神在官方 YouTube 上公开了技术架构。 ```mermaid flowchart LR accTitle: Tensor Robocar 感知与隐私计算架构 subgraph 硬件感知矩阵 ["全车 100+ 传感器矩阵"] S1["5 颗激光雷达 (LiDAR)"] S2["11 颗毫米波雷达 (Radar)"] S3["22 个高清摄像头"] S4["20+ 车内外麦克风"] end subgraph 端侧算力与交互 ["本地化边缘计算 (Edge Processing)"] S1 & S2 & S3 & S4 --> P1["车端实时大模型处理"] P1 --> P2["多模态自然语言理解<br/>(指向交互 / 语音控车)"] P1 --> P3["本地生物识别与数据加密<br/>(零云端上传 / iPhone级隐私)"] end ``` 在交互体验与数据隐私层面，Tensor Robocar 展现了高度智能化与安全克制： - **多模态自然语言直觉交互**：乘客可以直接用手指着特定车窗发出语音指令“把这扇窗摇下来”，或者指示“在这里左转”、“去那栋大楼的正门放我下车，然后自己去停好”，系统均能精准理解执行。 - **全端侧隐私计算架构**：鉴于车内搭载了海量传感器并实时采集乘客语音与生物特征数据，Tensor 坚持所有数据处理完全在车端本地完成（Edge Processing），不向云端传输任何生物隐私信息，构建了类似 iPhone 的本地化数据隔离与权限防护体系。 --- ### [18:55](https://www.youtube.com/watch?v=eLgm-nWKdMU&t=1135s) 丰田研究院的扩散策略突破：从机器人学术前沿到工厂人机协作 Avinash Balachandran 重点回顾了丰田研究院约两年前在机器人领域取得的标志性科研突破——“扩散策略”（Diffusion Policies）。这项技术被业内公认为视觉-语言-动作（VLA）模型以及当前具身智能（Embodied AI）热潮的核心先驱之一，其本质是将驱动 ChatGPT 等生成式 AI 的前沿机制成功迁移至物理机械臂控制中。 这项研究最具开创性之处在于其超高的数据利用效率：人类操作员只需通过远程遥控进行 60 到 100 次动作演示，系统便可在无需人工编写底层控制代码的情况下，快速训练出让机器人自主完成高难度操作的模型。该成果发表于顶级期刊《Science Robotics》上并获得了学界的广泛引用。 丰田并未将该技术停留在论文阶段，而是迅速将其与汽车工厂的制造流水线深度结合。与前面提到的“AI 作为导师教人类新手驾驶”形成生动镜像，制造车间展现的是“人类向 AI 演示传授技能”的全新协作模式： - **人机协同减负**：一线制造工人向装配机器人演示操作难度高、重复性强或对人体负担重的工作流程。 - **任务自主接管**：机器人学会后即可独立接管这些复杂制造环节，从而让人类员工从高负荷劳动中解脱出来，专注于更具创造性与高附加值的制造工序。目前该系统已在丰田多家制造工厂展开试点部署。 ### [20:35](https://www.youtube.com/watch?v=eLgm-nWKdMU&t=1235s) 丰田工厂的人机协作与机器人前沿探索 丰田研究院（Toyota Research Institute, TRI）在具身智能领域取得了实质性突破，他们不仅发表了开创性的学术论文，还在真实工厂环境中完成了落地部署。该技术团队刚刚完成了独立拆分，新公司“Walden Robotics”已于约三周前上市并筹集了大笔资金，目前正在丰田工厂以空前的大规模推进系统整合。TRI 的核心优势不在于单纯的数据量，而在于沉淀在制造流程中的深厚专业知识（Expertise），以及丰田独有的“人机融合”理念。 ```mermaid flowchart TD accTitle: TRI具身智能演进路径 subgraph Core["核心能力与商业化"] A["丰田制造专业经验"] --> B["TRI 开创性算法研究"] B --> C["分拆成立 Walden Robotics"] C --> D["丰田工厂大规模部署"] end subgraph NextFrontier["下一代前沿：人机直觉协作"] E["现状：孤立演示"] -.-> F["突破：无需言语的默契协作"] F --> G["典型场景：人机合力搬运大桌子"] end B --> NextFrontier ``` 团队目前正将精力投向下一个前沿领域：**解决机器人与人类交互缺失的问题**。当前行业内绝大多数机器人演示都是在孤立无人的环境下进行的，如果想真正为人类释放价值，就必须建立更直觉化的人机协作机制。例如两个人搬一张大桌子时往往无需语言交流就能默契配合，但如果协作伙伴换成机器人，系统就需要全新的感知与交互技术来实现丝滑的共同操作。 - **Walden Robotics 分拆上市**：约三周前完成分拆并成功公开上市，专注于在丰田生产线等复杂工业场景中落地具身智能。 - **丰田核心壁垒**：强调丰田真正的核心优势是工艺与人的经验专长，而非单纯的海量数据。 --- ### [21:56](https://www.youtube.com/watch?v=eLgm-nWKdMU&t=1316s) 自动驾驶对智慧出行与城市空间的重塑 来自美国酒店住宿协会（American Hotel and Lodging Association）的 Jesse Brady 向专家组提问：自动驾驶技术将如何彻底改变长途出行与未来旅游业？ ```mermaid flowchart LR accTitle: 自动驾驶对出行与城市的重塑 subgraph City["城市与基础设施变革"] A["V2X 协同"] --> B["动态交通管理与信号交互"] C["夜间无人化运维"] --> D["自主前往校准与充电"] E["车队呼叫服务普及"] --> F["15%~20% 停车场释放为绿地/住宅"] end subgraph Experience["个体体验与产业价值"] G["解放驾驶注意力"] --> H["通勤时间转化为生产力/陪伴"] I["旅途即目的地"] --> J["酒店与长途移动体验深度绑定"] end ``` 专家们指出，自动驾驶正在全方位重构出行生态： 1. **智能城市与车联网（V2X）**：自动驾驶汽车能够直接与交通信号灯、智能安全闸机及停车场通信，带来全新的动态交通管理模式。 2. **夜间自主维护与校准**：车辆在车主睡眠时可自主前往维护中心完成传感器重新校准，次日清晨完好归位，彻底省去用户的维护精力。 3. **重获时间与消除通勤损耗**：自动驾驶最大的价值在于“把时间还给人类”。过去一小时专注于路况的枯燥通勤，可以转化为高效工作时间，或是与家人、孩子沟通交流的温馨时刻。 4. **城市土地利用的根本性释放**：May Mobility 强调以安娜堡（Ann Arbor）为例，城市中约 15% 至 20% 的土地被停车场占据。当随叫随到的自动驾驶车队普及后，大量停车场将转型为城市公园或住宅区。 5. **旅途体验与文旅产业结合**：对于酒店与旅游业，行车过程本身将成为旅行体验的核心部分，行程与目的地的界限变得模糊，为消费者创造出全新的价值闭环。 --- ### [24:42](https://www.youtube.com/watch?v=eLgm-nWKdMU&t=1482s) 干线与中短途自动驾驶卡车的并行发展 来自内布拉斯加州一家小型货运公司的 Taylor 提问：干线（Long-haul）与中途（Mid-haul）重卡货运的自动化进程是与乘用车并行推进，还是作为独立轨道发展？ 专家组对此给出了清晰的商业逻辑分析： - **商业闭环更早成型**：自动驾驶卡车的发展速度在许多方面甚至快于乘用车，核心在于其商业模式（Business Case）更加明确——干线货运通过替代或补充长途人力、优化固定线路调度，能够直接为物流企业带来立竿见影的降本增效价值。 - **国际趋势与卡车编队技术（Platooning）**：尽管丰田并未在美国销售重型卡车，但全球范围内（尤其是欧洲）正在大规模开展示范运营。欧洲不仅测试单车自动驾驶，还广泛应用“自动卡车编队行驶”技术，通过紧密跟车降低空气阻力以实现大幅节油。 - **车队调度管理（Fleet Management）新生态**：重卡自动化的推进催生了大量专注于智能车队分发、动态调度及干线资产管理的新兴企业与技术方案。 --- ### [26:17](https://www.youtube.com/watch?v=eLgm-nWKdMU&t=1577s) 驾驶资格、家庭用车与政策监管的边缘挑战 现场观众 Daniel 提出了一个现实问题：他在女儿 7 岁时曾与朋友打赌，赌女儿 17 岁时是否还需要考驾照，如今女儿已经 12 岁，自动驾驶面对的监管与落地进展究竟如何？ ```mermaid flowchart TD accTitle: 自动驾驶家庭场景与监管考量 subgraph Scenarios["新兴家庭用车场景"] A["旧金山日常"] --> B["Waymo 代替 Uber 接送青少年 (无陌生司机更安全)"] C["高端市场 (如迪拜)"] --> D["无成人陪同的儿童接送 / 宠物托管日托转运"] E["车载人脸识别"] --> F["精准锁门防误下车 (红灯安全防护)"] end subgraph Regulation["政策与车企应对"] G["Zoox 监管规范 (如禁止单人承运 8 岁以下儿童)"] H["Tensor 策略: '可开可坐' (折叠方向盘/踏板设计)"] I["呼吁: 针对边缘场景迭代软硬件，避免过度监管"] end B --> G D --> H F --> I ``` 针对这一话题，嘉宾们展开了多维度的讨论： - **出行信任度的代际转变**：在旧金山等已常态化运营的城市，许多家长更愿意让孩子乘坐 Waymo 等无人车而非 Uber，因为车内没有陌生人类司机，安全性反而更高。在加州卡梅尔的 The Quail 经典车展上，有孩童看到 Tensor 满屏且无方向盘的概念设计后直呼希望将其作为自己的第一辆车。 - **“可开可坐（Drive or be driven）”的双模理念**：Tensor 认为驾驶本身的乐趣依然存在，因此在其豪华车型中保留了可折叠的方向盘与脚踏板，用户想自己掌控时依然可以无缝切换为传统驾驶模式。 - **政策与年龄法规的制约**：类似 Zoox 等自动驾驶车队正受到严格政策限制，例如明确规定不能单独搭载 8 岁以下儿童。人类青少年在获得驾照后可以独立承载 8 岁儿童，但自动驾驶汽车面临的法律门槛显然更为苛刻。 - **车内安全机制与新兴场景**：在面向迪拜等高端市场的家庭转运场景中，无人车甚至被用于接送儿童或将宠物送往日托所（Doggy daycare）。针对儿童在红灯期间可能误开门下车等极端边缘情况（Edge cases），车辆可以通过车内人脸识别技术智能锁定车门，家长也可全程远程视频监控。专家们呼吁政策制定者从具体边缘案例中吸取经验，在保障安全的同时避免“过度监管（Overregulate）”，以推动软硬件系统的快速迭代。 ### [30:27](https://www.youtube.com/watch?v=eLgm-nWKdMU&t=1827s) 车联网通信（V2V/V2I）落地瓶颈与多智能体协同愿景 探讨车联网技术时，丰田研究院（Toyota Research Institute）的 Avinash Balachandran 指出，丰田以及全行业多年来一直在深入研究车对车（V2V）与车对基础设施（V2I）的互联通信。这项技术主要聚焦两大应用场景：其一是通过聚合全车队采集的环境信息提供先验预警，例如前车若在某处匝道遭遇隐蔽的“黑冰”（Black Ice），后车就能在驶入前提前收到减速警报；其二是优化城市路网的交通流控，通过车速引导与信号灯配时协同，让车流平滑穿过交叉路口，最大限度消除拥堵回流。 然而，车联网在落地上面临着经典的“双边网络效应”难题。V2V 与专用硬件的红利只有在全路网普及率达到相当高的临界规模后才会真正显现；这意味着早期尝鲜的购车者在相当长一段时间内几乎无法获得实际收益，市场缺乏足够的装机动力。面对高昂的专用通信频段与硬件门槛，行业近年来已逐渐将注意力转向成熟的蜂窝网络通信。这种思路不仅能覆盖车辆，还能将行人的智能手机纳入交互生态——在无保护路口，行人的手机与靠近的车辆互相传递意图，系统会及时提示行人“此时是否适合通行”。展望未来，各端甚至会运行专属的 AI 智能体（Agent），汽车与手机的 Agent 自主协商、推演全局最优解，在瞬息万变的路况中协同保障行人与车辆的绝对安全。 ```mermaid flowchart TD accTitle: V2X 通信技术演进路径 subgraph S1["传统专用网路 (V2V / V2I)"] A["车对车 / 车对路专用硬件"] --> B["车队数据聚合 (黑冰/危险预警)"] A --> C["信号灯配时协同 (优化交通流)"] B -.->|"痛点: 早期装机门槛高<br/>缺乏网络效应"| D["推广受阻"] C -.-> D end subgraph S2["泛在蜂窝互联与智能体协同"] E["蜂窝网络 (Cellular / C-V2X)"] --> F["跨终端互联 (车载系统 + 行人手机)"] F --> G["多智能体协商 (Vehicle Agent ↔ Phone Agent)"] G --> H["复杂路口意图预判与主动避障"] end ``` --- ### [32:16](https://www.youtube.com/watch?v=eLgm-nWKdMU&t=1936s) AI 对传统汽车经销商体系的重塑：从售后维保到“主动上门”的销售新范式 在现场问答环节，一位从事汽车经销商售后软件开发（Fixed Ops）的观众提问：当前 AI 浪潮在短、中、长期将对拥有成百上千家门店的传统汽车经销商网络产生何种冲击？ Avinash Balachandran 表示，丰田与旗下经销商始终保持紧密沟通，经销商对拥抱 AI 展现出极高热情，不断主动询问落地场景。在短期，经销商最迫切的痛点集中在车辆售后维保管理与复杂的质保信息追踪上，AI 能在此提供立竿见影的降本增效；长期来看，物理实体 AI（Physical AI）与自动化机器人有望直接承担部分维修工序。丰田的核心理念在于“赋能并放大人类的能力”（Amplifying the Person），重点不在于零敲碎打地拼凑插件，而是基于 AI 彻底重构整个经销商的工作流。此外，AI 还能通过高度个性化的配置方案，显著拉动汽车后市场（Aftermarket）的选装与定制化收入，为主机厂（OEM）和经销商共同开辟全新的价值增量。 Tensor Auto 的 CMO Amy Luca 随即从初创车企的商业化视角展开补充。她透露，Tensor 并非走特斯拉式的直营模式（DTC），而是依然选择与经销商结盟。在传统维保之外，自动驾驶时代的维修需求已扩展至精密传感器清洁与标定。更颠覆性的改变在于时间的解耦： - **无人化夜间维保**：车辆可以在车主熟睡的深夜自行前往经销商店铺，向系统自主汇报所需维修项目，维修技师处理完毕后车辆再自行开回车库，车主甚至无需亲自踏进门店一步。 - **走出展厅的“自我推销”**：未来的销售链路将被彻底改写。一台搭载 L4 级自动驾驶的私人车辆甚至可以“自己把自已卖出去”——它能在预约时间自主开到潜在买家的家门口，主动演示各项黑科技，并引导买家完成体验。 Amy Luca 强调，尽管营销环节始终需要人类的情感联结，但实体展厅的物理空间限制已被打破，汽车不再坐等顾客光临，而是主动走向用户。 May Mobility 的 Jacob Crossman 则指出，May Mobility 采用车队运营模式而非面向个人消费者，但同样极度依赖生态伙伴关系。伴随着电气、机械架构与复杂的自动驾驶 AI 深度耦合，车队的日常运维与系统标定极具挑战，May Mobility 的策略是专注自身核心技术，将车辆物理维护深度交托给专业合作伙伴。 --- ### [36:48](https://www.youtube.com/watch?v=eLgm-nWKdMU&t=2208s) L4 级自动驾驶的法律责任归属与动态保险定价体系 关于法律责任与事故定责的现场提问，引发了全场对自动驾驶商业落地的深层讨论。一位观众提问：“如果我坐在一辆处于 L4 级自动驾驶状态的车里发生了碰撞，且被判定负有事故责任，究竟是由乘客买单还是主机厂承担？” Amy Luca 明确解答了这一难题，并介绍了 Tensor Auto 与全球知名保险经纪及风险管理公司达信（Marsh）的深度合作： - **L4 模式（责任归厂商）**：当车辆处于全自主的 L4 模式时，车内方向盘收起，乘客完全无法介入车辆的物理操作，行车控制权与软件决策全由 Tensor 掌控。此时一旦发生事故，乘员不承担任何责任，法律与经济责任全部由 Tensor 承揽。 - **L2 模式（责任归驾驶员）**：若处于辅助驾驶或人工驾驶状态（此时踏板和方向盘处于可用状态），驾驶责任则严格归属于持有驾照的实际操作者。 ```mermaid flowchart LR accTitle: L4 vs L2 责任归属与动态保费机制 subgraph DrivingModes["驾驶控制模式"] M1["L4 级完全自动驾驶<br/>(方向盘收起 / 软件接管)"] M2["L2 级 / 人工驾驶<br/>(踏板方向盘启用 / 驾驶员操控)"] end subgraph LiabilityMap["责任归属与保费模型"] L1["车厂承担全责 (Tensor)<br/>对自身算法安全性背书"] L2["驾驶员自负全责<br/>(车主个人保单履约)"] P1["L4 里程保费显著更低<br/>(AI 出险率远低于人类)"] P2["传统人类驾驶保费基准"] end M1 --> L1 --> P1 M2 --> L2 --> P2 ``` 基于这种控制权的分界，车险行业迎来了“动态保险定价”（Dynamic Insurance Pricing）的根本性变革。车载计算平台会精准统计并向保险公司同步数据周期内（按月或按季度）L4 托管里程与人类驾驶里程的具体比例。由于经过严密验证的自动驾驶系统事故率远低于人类驾驶员，主机厂有充分信心为算法背书，“每行驶一英里的 L4 里程，其保险费率将显著低于人类驾驶里程”。 针对 May Mobility 的运营现状，主持人 Steve Greenfield 追问其车队是否采取自保模式。Jacob Crossman 回应称，May Mobility 同样与专业保险机构保持深度合作，双方共享行车数据以协同推进承保模型的迭代，不过当前 May Mobility 尚未像 Waymo 那样公开发布其相较于人类驾驶的安全对比数据。随着问答深入，现场观众转向了更底层的技术逻辑，开始探讨自动驾驶算法如何应对 AI 本身所具有的非确定性（Non-deterministic）特征。 以下是根据视频转录第 5 部分整理的详细内容记录与分析： ### [40:13](https://www.youtube.com/watch?v=eLgm-nWKdMU&t=2413s) 应对驾驶中的高不确定性与动态决策 在驾驶过程中，车辆常会面临高度不确定性的瞬间（例如黄灯变红灯的临界抉择）。Jacob Crossman 解释道，人类司机在面对突发或模糊状况时，会在极短的时间内快速在脑海中“预演”各种可能发生的情景与后果。基于这种快速的直觉模拟，人类通常会采取相对谨慎的策略。 为了让自动驾驶表现得更像人类，May Mobility 在其推理系统中引入了类似的决策机制。在决策过程中，系统会在微型的内部模拟（类似一场“小型锦标赛”）中实时运行并评估多种备选方案。当环境不确定性增高时，生成的不同预测结果也会增多，此时系统会倾向于选择减速或微调方向等保守策略。这种动态调整仅持续几分之一秒或数秒，一旦状况明朗、环境重回确定性，车辆便会恢复正常的平稳行驶。 ```mermaid accTitle: 自动驾驶不确定性决策流程 flowchart TD subgraph S1 ["1. 感知与评估"] A["高不确定性场景<br/>(如黄灯临界点/突发状况)"] --> B["生成多种潜在路线与行为选项"] End subgraph S2 ["2. 内部模拟与筛选"] B --> C["内部微型模拟/锦标赛评估<br/>(运行多种 outcome 预测)"] C --> D{"环境是否明朗？"} End subgraph S3 ["3. 决策执行"] D -- "否 (不确定性高)" --> E["采取谨慎策略<br/>(减速/微偏避让)"] D -- "是 (确定性强)" --> F["恢复正常平稳驾驶"] E --> B End ``` ### [41:40](https://www.youtube.com/watch?v=eLgm-nWKdMU&t=2500s) 声音感知在无人驾驶安全中的关键作用 观众席的 Matt（来自旧金山）提出了一个有关车辆外部声音感知能力的疑问。他提到旧金山遍布数百个警报喇叭，每周二中午都会进行测试；如果遭遇海啸或地震警报，即便街上空无一人且没有视觉迹象，车辆也应当能够识别声音警报并采取避险行动（如靠边停车并提醒乘客前往高处）。 针对这一问题，小组嘉宾探讨了声音感知的必要性与应用： * **外部麦克风与声音识别：** Amy Luca 指出，Tensor Auto 的车辆配备了外部麦克风系统，使 AI 能够听取并处理车外声音（例如远比人类乘客更早地侦测到紧急救护车或警车的声音）。 * **上下文验证与网络互联：** 面对“海啸警报”这类极端边缘案例（Edge Case），车辆搭载的大语言模型（LLM）通过 5G 网络实时连接互联网。AI 结合环境上下文与网络数据进行双重验证，避免因误听而盲目做出让乘客跳车的错误判断。 * **极限控制与多模态感知：** Avinash Balachandran 补充指出，优秀的人类司机十分依赖听觉（例如通过轮胎摩擦声判断车辆是否失滑）。丰田研究院（TRI）为此专门研究了如何利用轮胎声音反馈来改善车辆在冰雪路面或剧烈变道时的极限操控能力。 - 声音感知在无人驾驶中的典型应用：提前听知远处紧急车辆警笛、利用轮胎声音监控车辆极限附着力与打滑状态。 - 边缘案例的处理机制：结合多模态声音感知、5G 联网数据校验以及 LLM 环境上下文理解，防止非规则决策引发误判。 ### [44:49](https://www.youtube.com/watch?v=eLgm-nWKdMU&t=2689s) 15 年后全球全自动驾驶新车渗透率预测 在论坛的尾声，主持人 Steve Greenfield 向三位嘉宾发起了一场“100 美元赌局”，要求他们抛开当前的官方表态，对 15 年后（全球每年约 9000 万辆新车销售中）全自动驾驶（L4 级）新车的渗透率给出预测： | 嘉宾 | 预测比例 | 主要思考维度与理由 | |:--- |:--- |:--- | | **Amy Luca** | **20% – 25%** | 保持乐观预期，但考虑到消费者更换新车的周期较长（通常 10–15 年），普及速度受到固有更新换代节奏的制约。 | | **Jacob Crossman** | **~25%** | 物流与货运交付领域会率先实现大规模落地（去司机化经济效益明显）；但由于各地区法规政策极其复杂且繁琐，整体普及受到牵制。 | | **Avinash Balachandran** | **小于 20%（可能在 10% 左右）** | 技术本身并非最大障碍，真正的瓶颈在于全球各国的法规、政策认证及基础设施差异（如无车道线地区）；小型创业公司需要跨越漫长的合规审查。 | 嘉宾们强调，技术可行性在 15 年内不会是主要瓶颈（各大主流 OEM 届时都会具备 L4 级技术实力），真正的挑战在于**法规政策与全球认证（Homologation）**： - **政策与区域监管碎片化：** 尽管 Waymo 在政策制定与 L4 验证方面开辟了道路，但不同国家乃至城市（如加州、欧洲各城市、迪拜）的测试与验证标准截然不同，合规成本高昂。 - **硬件与整车安全认证：** 对于推出全新车型的企业而言，新传感器配置与车辆本身的整车安全认证（不仅是软件安全）耗时极长。 - 论坛在主持人对现场观众积极提问的感谢与掌声中圆满结束。 ## Notable quotes > [05:03](https://www.youtube.com/watch?v=eLgm-nWKdMU&t=303s) 根本没有人真正掌握标准答案。我们唯一能做的，就是通过尽可能多的路径和手段去反复探索。 — Avinash Balachandran > [06:43](https://www.youtube.com/watch?v=eLgm-nWKdMU&t=403s) 过去完全无法直接向工程工具库添砖加瓦的产品或运营职能同事，现在借助 AI 工具，竟然拥有了直接产出代码与工程能力的可能性。 — Jacob Crossman > [12:51](https://www.youtube.com/watch?v=eLgm-nWKdMU&t=771s) 按理说我能活到现在真是个奇迹 — Steve Greenfield > [15:12](https://www.youtube.com/watch?v=eLgm-nWKdMU&t=912s) 把这扇窗摇下来 — Amy Luca > [15:12](https://www.youtube.com/watch?v=eLgm-nWKdMU&t=912s) 在这里左转 — Amy Luca > [15:12](https://www.youtube.com/watch?v=eLgm-nWKdMU&t=912s) 去那栋大楼的正门放我下车，然后自己去停好 — Amy Luca > [30:27](https://www.youtube.com/watch?v=eLgm-nWKdMU&t=1827s) 此时是否适合通行 — Avinash Balachandran > [36:48](https://www.youtube.com/watch?v=eLgm-nWKdMU&t=2208s) 每行驶一英里的 L4 里程，其保险费率将显著低于人类驾驶里程 — Amy Luca ## Mentioned Steve Greenfield, Automotive Ventures, Avinash Balachandran, Toyota Research Institute (TRI), Toyota Motor North America, Woven by Toyota, Walden Robotics, Amy Luca, Tensor Auto, AutoX, Tensor Robocar, Jacob Crossman, May Mobility, Lyft, Uber, Grab, Choco-Choco, Waymo, Zoox, Marsh, American Hotel and Lodging Association, Jesse Brady, Science Robotics ## Video description Steve Greenfield - General Partner, Automotive Ventures Avinash Balachandran - Vice President, Toyota Research Institute Amy Luca - CMO, Tensor Auto Jacob Crossman - Sr. VP of Autonomy, May Mobility Vehicles are rapidly becoming intelligent edge devices. In this panel, experts from top automotive OEMs and suppliers discuss how AI is redefining everything from driver experience to fleet safety. We'll look at how data from cameras, sensors, and ECUs is being turned into real-time insights and how AI is shortening design cycles, improving safety, and laying the groundwork for true autonomy. Join this panel to see the future of the AI powered automotive industry! [Source](https://www.youtube.com/watch?v=eLgm-nWKdMU)

[5:03] Avinash Balachandran: 根本没有人真正掌握标准答案。我们唯一能做的，就是通过尽可能多的路径和手段去反复探索。 [6:43] Jacob Crossman: 过去完全无法直接向工程工具库添砖加瓦的产品或运营职能同事，现在借助 AI 工具，竟然拥有了直接产出代码与工程能力的可能性。 [12:51] Steve Greenfield: 按理说我能活到现在真是个奇迹 [15:12] Amy Luca: 把这扇窗摇下来 [15:12] Amy Luca: 在这里左转 [15:12] Amy Luca: 去那栋大楼的正门放我下车，然后自己去停好 [30:27] Avinash Balachandran: 此时是否适合通行 [36:48] Amy Luca: 每行驶一英里的 L4 里程，其保险费率将显著低于人类驾驶里程

[17:13] Amy Luca: Roll that window down [17:13] Amy Luca: turn left here [17:13] Amy Luca: go to the front door of that building in the front, let me out, and park. [6:48] Avinash Balachandran: Driving Sensei [40:06] Jacob Crossman: tournament [6:48] Jacob Crossman: drive-out [6:48] Jacob Crossman: May Days [0:48] Amy Luca: Tensor Robocar [30:09] Avinash Balachandran: chicken-and-egg

[0:04] We're going to do things a little different. Since  you found us in the far corner of the palazzo here [0:10] and you stuck around to the very end, we'll take  questions throughout. Right? Don't start right [0:15] now. I got a little bit of a preamble to do. But  if you if you have questions, there's a reason why [0:20] you're in the room today. If you have questions,  raise your hand. I'll I'll watch the audience here [0:24] and uh we'll take questions throughout. So,  um, let's do that and we'll get started. So, [0:30] um, spend a little time introducing yourself, your  company, and your role a little bit. You know, [0:35] I had a chance to talk to all of you ahead of  time and like we've got good good diversity up [0:39] here in terms of what they're actually working  on, different philosophies around um um autonomy [0:45] and technology uh in general. But maybe Jacob,  start with you. For those that were here on the [0:49] last session, you were the keynote, so thank you  for that. He had some cool videos. If you missed [0:54] his videos, he had some great videos. talk a  little bit about your role May and uh what you [0:59] work on day-to-day. Uh sure. My role I'm a senior  vice president of autonomy. I lead the technology [1:04] development at May. May is an autonomous vehicle  company based out of Ann Arbor and we're deployed [1:10] in several cities. Um most notably uh uh uh  sorry Atlanta, Arlington, a lot of A's. We we [1:18] also drive around in Ann Arbor around our our uh  headquarters in Minneapolis. So we're partnered [1:23] with uh a lot of uh of the people that you would  know Lyft Uber uh recently we announced a deal [1:29] with Chowochow uh as well as Grab who is a  uh investor in May. So terms of day-to-day [1:36] um I lead all the technological development uh at  my role as you can probably imagine uh mostly I [1:41] deal with problems coming up uh risks etc. But  what I really like to do is to dig into the AI [1:47] technologies. uh we have a unique technology  I just talked about in the previous uh session [1:51] uh where we combine online reasoning with offline  uh learning uh and that's fascinating and kind of [1:57] diving into how that can help us scale our uh our  autonomy. Very cool. Thank you. Um Amy may have [2:04] more air miles than me which is pretty awesome or  not depending on who you are but I'm a United dog [2:11] waiting for you at home. I believe I'm a million  mileer. Yes. Yeah. Yeah. Yeah. U Amy Luca. I'm the [2:16] chief marketing officer for Tensor Auto. Uh you  may have heard of us. Uh we're about 10-year-old [2:21] company. You may have heard of our previous name,  which was Auto X. Um and we launched uh last year, [2:27] almost uh a year ago exactly, the world's first  personal robo car. Um the Tensor Robocar. And [2:34] we are in the middle right now of deploying in  Dubai. uh little challenging at the moment for [2:40] for geopolitical reasons but uh hopefully uh  you'll be hearing about uh the first personal [2:47] robocars in someone individual garage by the end  of this year. Uh we are focused around the idea of [2:56] uh a fully autonomous and AI agentic vehicle.  So we have natural language processing. You can [3:02] talk to the car as if it's a human being to give  it commands. Um, and it is built from the ground [3:07] up. So there's no retrofit. Um, it actually has  a lot of AI design in it. Um, over 100 sensors [3:14] including five LAR, 11 radar, um, 22 cameras, um,  lots of other technology on board as well. And, [3:22] uh, yeah, hopefully you'll have your own roboc  soon. And your role and your responsibilities. [3:27] I am the chief marketing officer. So I  handle a lot of obviously the marketing, [3:31] but also commercialization. Uh we are a small and  mighty startup-minded company still so we have [3:36] a lot of hands- on deck uh when it comes to the  commercialization aspect of the brand. Uh so yeah [3:42] that's my my role. Avan you work for the largest  automaker on earth you know that that's hard to [3:51] follow. Uh but yes Toyota is the largest automaker  in the world by by volume. My name is Avvin uh and [3:56] I work at the Toyota Research Institute. So we  are the part of Toyota that's really looking at [4:01] the at the future and trying to figure out what's  next for Toyota. So we've been around for about 10 [4:05] years and so 10 years ago we were actually working  on full self-driving kind of technologies. That [4:10] was the main focus of our research efforts. Since  then you know we've evolved that that part of of [4:15] TRRI has spun off into woven by Toyota. That's  commercializing that efforts. Uh we've started [4:21] doing a lot more work in robotics in energy and  materials in all kinds of different areas trying [4:26] to figure out what's next. So that's TRI's role  and I'm a vice president in there. I I basically [4:32] uh continue our advanced driving research. So  these are things beyond sort of full self-driving [4:37] and I'll be excited to talk more about that in as  this panel progresses but really around ideas of [4:42] how do we have people and these intelligent uh  physical uh AI systems whether they be cars or [4:48] robots collaborate and work more closely together.  So I have teams working in the automotive space as [4:53] well as in sort of robotics uh and and doing that  kind of stuff. And so kind of just to maybe zoom [4:58] out and give you a little bit of context about  Toyota uh and how we think about this uh I guess [5:03] you know maybe just as a general raise of hands  here how many people are super excited about AI [5:08] and and its applications in the world today and  how many people are are very terrified of this [5:15] and so you see there's a mix here and actually  if you look at statistics in the US about 57% of [5:21] AI practitioners are super excited and about 17%  of the actual public is super excited. 17 or 7D [5:28] 17 17. So there's this gap, right? And Toyota is  a company that sells roughly, you know, over 11 [5:33] million cars a year all over the world. And we're  really concerned about this gap. How do we build [5:37] products that work for everyone? And so in order  to do that, uh, we actually believe that there's [5:42] nobody really knows the answer. And the thing that  we can do is try as many different pathways as [5:47] possible through as many different means. And as  a specific example, within TRRI, we have a venture [5:52] fund. that fund invests in companies including May  mobility which is one of our portfolio companies [5:57] and helps us explore with them this opportunity  that they they're doing internally we're looking [6:01] at developments around using AI to really amplify  the person and using that both in the vehicle as [6:07] well as more holistically across the automotive  process from engineering design manufacturing [6:13] sales warranty etc. So, we've got um a good  diverse uh perspective up here. You've got [6:19] the largest automaker in the world and their  perspective on autonomy. We've got Tensor who [6:24] wants to democratize ownership. We're all going  to own fully autonomous vehicles that will take [6:28] us around, but then at night they're going to go  into the Uber fleet and generate revenue for us [6:32] overall. And then we've got May over here who  in the short term is testing AVs on the roads. [6:37] But long term, I'm assuming if I ask Jacob,  probably more of a software technology play. [6:43] um getting their technology within other  automakers maybe Toyota auto or not. So um we're [6:49] at an AI conference. So my qu first question I  guess is um you know what's top of mind for you in [6:55] terms of initiative? So over the next like 6 to 12  months what's the number one AI based initiative [7:01] that your company has? And Jacob maybe I'll start  with you. Sure. Um our our goal really is is drive [7:08] route in as many places as we can. So we we focus  our attention on the things that make that happen. [7:14] AI is a critical piece of this, not only in our  stack. AI is central to the stack, both the things [7:19] that most people think of as AI, uh the learned  models, uh from, you know, transformer-based [7:24] things, things based off of LLMs, as well as our  online reasoning capabilities. Um we're also, [7:29] uh beginning to use AI processes a lot more in  terms of actually developing the software. Um, [7:35] one really cool thing, slightly tangential, but  um, we had our May days, which is when we all get [7:40] together for a few days and, uh, at May in Ann  Arbor, and we had a little competition where, [7:46] um, we we were going to build a new game that you  could play in the center console for our cars. And [7:54] in the team that I was on, it was actually our  chief of communications or our chief of staff [7:59] for communications who actually built that  model using AI. she built the actual game and [8:06] I thought that was really really cool. So this  is really opening up new possibilities to have [8:09] maybe people in product and other roles that maybe  haven't traditionally been able to add to our our [8:15] uh our uh tools and to our capabilities the  ability to start directly contributing. So [8:21] uh that's been interesting as well. So AI is at  the center of how the car drives and it always has [8:26] been and now we're beginning to bring uh AI into  how we develop the software for the car. Cool. [8:31] Amy, same question for you. I think right now  we're we're headlong into the commercialization [8:38] part of our business. Um, but I think the focus  from an AI perspective is really around the [8:42] agentic features of the vehicle and building out  more of those so they're intuitive. I think when [8:48] you earn a Whimo or you're any in any sort of a  level four car right now, a lot of it is based [8:53] on an app or you know it's it's input. it's not  natural language talking to the car or having that [8:59] car really understand from an agentic perspective  your life um and how it can help you and be more [9:06] of a personal assistance. So we think of that  technology of how do we build AITIC systems [9:11] within the car that are actually more than just a  point A to point B but actually helping you with [9:15] your daily life. Um a lot of our technology  right now is focused on how does it integrate [9:20] with other applications uh say on your phone.  So for instance um in the morning your tensor [9:28] could text you and say you have a meeting at this  time traffic conditions are this you know a little [9:34] heavier wake up wake up we got to go earlier we  got to go earlier and hey by the way do you want [9:39] to stop at Starbucks on the way there we'll plan  that into your day so if you think about that as [9:43] like your personal assistant and that's really  what we're trying to to build um a robo car to [9:48] be more than just a point A to point B. So a lot  of our technology innovation is really around [9:54] um how do we make this rolling personal assistant  uh to be indispensable to your daily life more [9:59] than a vehicle and actually we were talking  about it earlier in this conference. I almost [10:03] think like talking about autonomous vehicles is  almost not enough. It's really embodied AI and [10:10] these AI agents that are going to be helping to um  do more than just manage your life but also like [10:17] in our car for instance it manages power systems.  you know those different there's different agents [10:22] that are playing different roles within the  vehicle at any given time. Um but yeah that's [10:26] what we're working on. Average TRRI's priority.  Yeah. So I think um there's kind of three parts to [10:32] this this answer that I'll give. One is you know  like many of the panelists here I think there's [10:36] various ways we're thinking about how we can use  AI in the business and that's kind of twofold. [10:40] One is the most obvious. How do we integrate  it into our products that we have integrate [10:44] you know agents autonomy other features more  directly into the vehicle make it more seamless. [10:50] uh and and these are things that we're working on  right now. The second one is how do we incorporate [10:55] it into our business processes uh and you  know examples of that you know we see in our [10:59] design process we're looking at using AI tools to  accelerate and condense the vehicle engineering [11:04] cycle so that that can be done a lot faster. We  actually have a team right here from Toyota Moans [11:09] North America Muhammad and team that are looking  at using AI for warranty and recall information [11:14] and so just everywhere across the business we're  asking ourel hey can we make use AI in a way that [11:19] makes this this business more efficient uh but  sort of sort of zooming out you know we're what [11:24] we're really thinking about here is in the age  of AI as you think more deeply about it the real [11:30] kind of challenge is in human agency autonomy  and purpose and how do we build AI initiatives [11:37] that empower those things. And so to give you a  specific example, as we think about autonomy in [11:42] cars, you know, a lot of the discussion is around  pulling the driver out and there are lots of good [11:46] reasons and use cases for where that's uh useful.  Uh but as a company that sells 11 million cars [11:51] all over the world, uh most of our consumers are  in places where this access to this technology [11:56] doesn't exist. But we do want to provide them  the benefits of this technology as soon as we [12:00] can. So cars can be safer, uh more enjoyable.  And so one of the metrics we're looking at, [12:05] one of the paradigms we're looking at is instead  of thinking about just AI sort of replacing [12:09] the driver, is there a way that we can use AI to  amplify the driver? And we're looking specifically [12:14] at this idea of teaching the driver. So imagine  that we are building really performant AI systems [12:19] that can drive super well. Is there a way that we  can use that uh and we call this project driving [12:23] sense to teach, for example, teenage drivers  to get better with driving over time. So, I [12:28] don't know how many of you have teenage kids, but  certainly I'm sure you worry about them getting [12:32] their driver's licenses. And unfortunately,  in the US, as in many parts of the world, [12:36] car accidents are the leading cause of death of  teenagers. And so, can we use AI and these these [12:41] AI systems not just purely from a way of removing  uh agency from the human and and taking over the [12:47] driving task? Are there ways we can augment it in  places where that that kind of capability doesn't [12:51] exist? It's interesting. In the last like 3 days,  I was driving somewhere in Atlanta. I was like, [12:57] I shouldn't really be alive. The risks  that I've taken driving over the years, [13:02] especially when I was younger, I mean, just take  undo risks, right? So, it would be great if the [13:06] cars could protect the drivers from doing  stupid things because as males specifically, [13:11] we do a lot of stupid things when we're young. So,  I think that's that's pretty exciting. So, let's [13:15] transition to the third question I would have  here, which is um I guess what are you proudest [13:21] of or what stands out? If you look back over the  last 18 months, two years, some AI initiative [13:26] within the company that's really stood out and  the more granular you can be for the audience, [13:30] the better. But the example, the granularity,  what did it achieve and why are you proud of that? [13:38] So um for us the last two years uh have been about  how to take the newest learnings uh uh that we see [13:46] in AI the large language models etc and how to  adapt them to cars and um we had a drive route [13:54] capability uh in 2023 worked in smaller domains  uh it was pretty exciting uh it was primarily [14:01] it did it did have some AI capabilities in the  perception system but not necessarily in the the [14:06] driving components, the policy generation,  etc. And we were able to very very rapidly [14:12] uh convert that to a hybrid system that we use  today, uh which uses both capabilities. And we [14:17] were really surprised at how fast we were able to  get that system to work. Uh and really the secret [14:23] of that was that we we learned on about 20,000  mi of driving. Uh we we did some really clever [14:29] things. We didn't just learn by what our car did,  we also learned by the other cars that our sensors [14:33] picked up. So, we were able to learn from that. Uh  but when with that very little amount of driving [14:38] because we already had this robust decision-making  engine, we were able to get literally uh orders [14:45] of magnitude improvements on the the complexity we  could drive in and the situations we could handle. [14:49] Um and it all happened very very fast. And uh as  as the leader of all the engineers, I was quite [14:54] proud of the team for uh what they were able to  accomplish doing that. Um, and it's just setting [14:59] the stage for uh the next year, which where we've  got many more advances uh in in uh in line for [15:04] that. And I think we're going to we're going  to have another couple of orders of magnitude [15:07] improvement as well. Nice. Thanks, Jacob. Maybe I  mean, I think there are a couple things. I mean, [15:12] for us, it was a huge achievement to launch our  car last year. We went stealth six years ago [15:18] um after doing robo taxis in China. We had about a  thousand robo taxis, retrofitted cars um in China. [15:24] and our founder came back to the United States  and said, "Okay, let's start from scratch." Um, [15:29] so we used AI in a lot of the design process for  what the Tensor Robocar looks like. And I think [15:34] it's one of the things that people don't realize  when they look at the car that the car is actually [15:38] designed around the sensor suite, not the sensor  suite put onto a car. Um, and people say, why does [15:43] the car look that way? And so I think one of the  things that our engineers are really proud of is [15:46] how they've used AI in that design process to make  a car that really um has no blind spots. I think [15:55] the other thing too is we our world foundational  model similar to u what you're talking about with [16:00] your model was taking lots of and by the way  you can see all about our our founder believes [16:05] in sharing everything with the world. So our  tensor foundational model is actually on our [16:10] YouTube channel explains what it is. Um but it  was really around the idea that we can take these [16:16] um clips of um real road activity, real road  cases and augment that with a lot of synthetic [16:24] um scenarios so that we can actually have every  single edge case, not every but like you know lots [16:29] and lots of edge cases way more than any human  driver will ever inter uh interact with. um and [16:36] kind of continually building that model and just  similar to what you said like learning from others [16:41] um being able to make these cars as safe as  possible because the biggest factor in level [16:47] four autonomy for consumers and I think you  talked about you know are we fearful of AI [16:52] well people are really fearful of giving their  life to a car and our car in level four has no [16:57] steering wheel so when there's no steering wheel  and there's no pedals in level four you're really [17:02] trusting a vehicle uh and so all of those um AI  simulation models. Um sharing that with the world, [17:10] helping other companies and learning from  other companies, I think is going to just [17:13] um you know, we're really proud of of that sort  of open-source mindset that we have in our company [17:19] uh to hopefully make cars um safer and more  trusted by the people that will eventually own [17:24] them. If you haven't Googled tensor, the tensor  car, Google the tensor car and take a look at [17:30] it. How many sensors on the car total? It's over  100. um five LAR, 11 radar, 22 cameras um of think [17:38] over 20 microphones. So the car literally um you  know is looking seeing everywhere all at once and [17:46] inside. So for it to be an AIC vehicle that has  uh can listen to your voice for instance you can [17:53] point to the window and say roll that window down  and it will roll that window down or turn left [17:58] here or go to the front door of that building you  know in the front and and let me out and park. So, [18:04] it does understand natural language. And by the  way, I know it sounds crazy, but as the marketing [18:08] person, they actually let me play with it and it's  pretty amazing. And I every time I have to drive [18:14] long distances, I'm like, why can't we have this  now? Um, but yeah, it's it's it's very rich from a [18:19] sensor suite perspective. And I think the other  thing that our founder really believed in was [18:23] privacy because one of the other factors when you  have a level four vehicle and all of these sensors [18:29] and taking all this biometrics and listening to  everything you have to say is is it private? And [18:33] so actually all of our processing happens um in  the car. So we're not connected to a cloud for any [18:39] of your biometrics and nothing is shared. So think  about it like an iPhone. Um, lots of data goes in [18:44] your iPhone, but Apple doesn't necessarily get it  unless you get good permission. And so our safety [18:48] and security or privacy is based on that as well.  Very cool. Arash. Yeah, great question. And I [18:55] think there's lots to be be proud of at Toyota and  TRRI. I think the example I'll pick is, you know, [19:00] roughly about 24 months ago. Uh, TRRI was the  pioneer in a in a technology called diffusion [19:05] policies. It was really sort of the precursor  for vision language action models and all the [19:10] growth you see in robotics. recently uh and so  the the team there did really good work to kind [19:15] of build that out. Basically, can we use the same  technology that powers chat GPT and use that for [19:20] robotics and um you know was based on collecting  demonstration data you know collecting relatively [19:26] small quantities of it 60 to 100 demonstrations  of a teley operator doing something with a robot [19:32] and then very quickly without any code training  a model to allow the robot to do that. TRI did [19:37] that pioneering work. We published uh you know a  seminal paper in this in science robotics that's [19:41] highly cited and so from a research perspective  it was awesome you know we did we did this good [19:46] thing but the team didn't stop there they really  focused then on how do we apply this to Toyota [19:50] and in a way that's consistent with our values  and so there you know we work very closely with [19:55] manufacturing and once again looking at you know  uh the the points that I made earlier around human [20:01] agency you know what is the most useful way we  can combine this technology with the people in [20:05] Toyota and we looked at things like having our  manufacturing [clears throat] team demonstrate [20:10] to robot tasks that were either difficult or uh  uh stuff that was harder for them to do so that [20:16] they could focus on higher value tasks. And so  earlier I mentioned this paradigm of having an [20:21] AI teach a human. In this way we are actually  looking at the opposite. Can a human teach an [20:25] AI have an assistant that you can show tasks to  and have the robot go off and do that. And so [20:30] they actually worked with several uh parts of the  factory do that. there there are trial deployments [20:35] that are going going down with that as well. Uh  but also what we did was we actually ended up [20:40] uh spinning that out. So that just got spun out uh  and went public about 3 weeks ago. Walden Robotics [20:45] uh and uh so they've raised a large amount of  money and now can pursue this at a really massive [20:49] scale and they're starting with integrations in  Toyota factories which I think is really really [20:54] cool. And the last thing I'll add is so this  team has kind of published some seminal work [20:57] actually done deployments in the factory and tried  to incorporate it what's unique about Toyota which [21:02] is our specific way of thinking about how we  can incorporate humans people in this process [21:07] uh but also in trying to capture expertise which  is what is our real advantage at Toyota it's not [21:12] data it's sort of the expertise uh but now the  team is moving on to the next frontier and what do [21:17] you see missing in a lot of these robotics kind of  demonstrations is any interaction with a person so [21:22] these are all happening in isolation and so We're  saying well you know if you really want robotics [21:26] to unlock value for people right now how do we  think about ways in which people in robots can [21:31] can interact in a more intuitive way. So imagine  you know you need to move a large table me and [21:35] Steve could probably do that without talking to  each other but if I had a robot instead of Steve [21:39] as my partner there would need to be some sort of  technology that allows us to really intuitively be [21:44] able to move a table uh you know 10 feet or  whatever it is and so those are kind of the [21:47] problems that we're looking at. Very cool. Very  cool. Um, questions for the audience. Got one. [21:56] Mic's coming. There you go. Thank you. Fire away.  Hi there. Uh, my name is Jesse Brady. I'm with [22:01] the American Hotel and Lodging Association.  Uh, thought this was really insightful. I [22:06] noted a comment about, you know, what about when  we'll be able to use this to drive long distances [22:11] and just interested in how you think these  technologies will change the future of travel. [22:19] I mean I think that it's going to change travel  fundamentally in every single way. Uh we were [22:24] actually talking earlier about smart cities.  So when you have agentic or autonomous vehicles [22:31] um they could traffic management could be very  very different because the cars can actually [22:36] you know talk to a stoplight or you know how does  it talk to other you know smart safety gates and [22:42] parking all those sorts of things. We would talked  about how it will change um just maintenance in [22:48] general. Like my if I have my car and it needs  to sensor recalibration, it can go do that in [22:54] the middle of the night while I'm sleeping um  and not then come home and be ready. But I think [22:59] um travel efficiency, us being able to get our  time back. I mean, I think the biggest and most [23:05] impactful thing that autonomous vehicles are going  to do for us is give us our time back. I mean, [23:10] if you think about just if you commute  to work and you have an hour commute, [23:14] think about lost productivity because you actually  have to focus on the road. Um, getting connection [23:18] back with your family, talking to your kids, um,  you know, all those sorts of things. I think from [23:23] a transport perspective, just, you know, I don't  even know if we can possibly understand today all [23:29] the impact that's going to happen, but, um,  it's pretty exciting to think about. At May, [23:34] uh, we've always had this vision that autonomous  vehicles are going to transform cities. So just [23:39] imagine for a minute all the parking lots become  parks instead of parking lots or perhaps you get [23:45] more residential space. You think of how much of  cities I think I want to say like Ann Arbor where [23:50] where we're located like I don't know like 15 20%  of the land mass is used for parking lots. So when [23:56] you can start calling vehicles they can always  take you where you want to go. That's a huge [24:00] change. So, not only will it change the individual  uh life who's who's using the autonomous vehicles, [24:05] it will change the communities in which uh the  vehicles are being used. To add something really [24:12] quick, I think these are excellent points.  I think one of the interesting things if you [24:15] think about autonomy in vehicles as it relates  to travel is that the journey itself now becomes [24:20] a bigger part of the experience. And so, you  can think about that in various ways. You know, [24:25] getting you more time in the vehicle, etc.  But you know if you're coming from sort of [24:28] the hotel industry you can imagine that the now  the destination is just as important as the ride [24:33] right there ways in which you can connect both  of those things. So where you're going and the [24:37] experience in the car linked in some way and  you're able to capture more value for the for [24:42] the consumer d in that interaction. Great question  up here. Oh he sorry. My name is Taylor. I'm with [24:50] a small trucking company out of Nebraska. Uh, I  guess I'm curious if you guys have a perspective [24:54] on the process of like these like actual trucks  that transportate like that transport materials [24:59] like are those being autonomized at the same  time as cars like in tandem? Is it a separate [25:05] thing? Like what's your guys' perspective on  that? Long haul or mid-h hall trucking? So, [25:11] while May in specifically is not uh automating  those right now, uh they are being automated. [25:16] I I'm going to guess that you're going to have  uh some ideas as well there. So uh so yeah I'll [25:21] let my colleagues or the others here talk. Yeah I  mean so certainly you know Toyota doesn't sell any [25:27] any heavy trucks in the US uh but you see a lot  of companies that are involved in in automated [25:32] trucking and I think in in many cases you know  the one of the reasons that that was an earlier [25:37] uh one of the reasons that that thing developed as  fast as it did was because the business case was a [25:42] lot clearer in terms of you know the the value you  could capture if you were able to automate some [25:46] of these routes. And so I think we're seeing both  of those and we're seeing that internationally as [25:50] well in Europe there huge um demonstrations of  of not just autonomous trucking but platooning [25:56] for the purposes of saving fuel efficiency. Uh  and so um along with that I think so there's a [26:01] whole bunch of work in automated truck trucking.  I think we're also seeing a lot of interest in [26:05] then sort of the fleet management distribution of  these as well. And so there's a lot of interesting [26:09] companies and and work that's happening in  that area. Great. Hi, my name is Daniel. Um, [26:17] when when my daughter was seven, I placed a little  bet with a friend uh whether she'll have to get a [26:23] driver driver license when she'll get to be 17 and  uh now she's 12 already and my time is running and [26:31] I was wondering how are you tackling this all the  regulations? We saw a tensor. [laughter] I mean, [26:38] it was interesting last year the brand at  the Quail Motorsport um event which is in [26:44] Carmel and there was a kid that came up and got  in the car and his dad was there and he's like, [26:48] "This is so cool. There's screens all over and  there's no steering wheel and this is great." And [26:52] he told it to his dad, he's like, "I want this car  to be my first car when I'm 16." And I was like, [26:57] "You don't have to be 16 to have this as your  first car." Um, and if you actually see right now [27:01] like in San Francisco Whimos, parents are more apt  to put their kid in a Whimo than an Uber because [27:06] there is no driver because it's safer than having  a stranger drive them. Um, so they put their [27:11] teenagers in an Uber or in a Whimo than over an  Uber. So I think it's interesting. Uh, we believe [27:18] though that people still want to and will continue  to want to drive when they want to drive. So we [27:24] say drive or be driven. Our car has a collapsible  steering wheel and pedals, but if you want to [27:30] drive, you can drive it just like a regular car  because I think there's still some enjoyment we [27:33] have in driving. So, will your your uh daughter  have a driver's license? I hope so. Um I because [27:40] it's it's fun to drive, but I think most people um  we make it to a world where people fewer and fewer [27:46] people will actually have a driver's license.  So, go ahead. I I'll just add one note which [27:50] I think a lot of this is also driven by policy  and regulation. So you may have seen this week I [27:55] think there's some regulations around zuks and its  deployment and one of the stipulations for example [28:00] is I think there's some limitation on age limit  so they can't take uh kids under the age of eight [28:05] alone and things like that and so um you know  to to your question I think some of this is also [28:10] driven by sort of regulation over time and we're  seeing you know the regulations that are required [28:14] for AVs are not the same as a teen driver right  a teen driver can take an 80ole without without [28:19] needing any any permission so I think uh you know  we'll kind of see have to see how that evolves [28:23] there's like certain safety things. Actually,  we were talking about uh issues with some Whimos [28:27] sometimes an edge case where they locked. But I  mean, when parents can actually observe their kids [28:32] in the car while they're being taken to grandma's  house, for instance, or somewhere, and they're not [28:37] allowed to get out because they know the car knows  through facial recognition that it's the child [28:42] in the car. They can't just stop at a stoplight  and accidentally get out. Those are the sorts of [28:47] things that we have to think through when you're  thinking of autonomous vehicle technology that [28:51] you know we're not thinking about now. Um and so  that's the exciting part is we have to all these [28:56] things to figure out. Begs an interesting question  that I hadn't thought about before. Right. So at [28:59] least in America what are you like 16 17 before  you can get a driver's license and that before [29:04] that age even if your parents bought you a car  you couldn't drive the car but are we going to see [29:09] especially for like more affluent families that  kids that are four years old own a car? I don't [29:14] know if a four-year-old is gonna I don't know if  it's funny actually we are launching in Dubai and [29:19] we are we are a luxury vehicle right and so a lot  of the customers we've talked to and things like [29:24] that are for family transport um and you know you  don't need a parent to transport your kids or your [29:31] pets by the way um so you know can you know take  your doggy a dog to doggy daycare um but I think [29:38] those are it really does come down to regulation  it depends on the city the municipality the [29:44] country how they're going to regulate um autonomy  and always there's going to be that one case [29:49] that's going to pop up um that kind of like oh wow  you know maybe we should regulate it more but I [29:54] think um you know hopefully we don't overregulate  um too much and we we learn from uh the edge cases [30:02] and and make the software and the hardware  better interesting I had a question regarding [30:09] uh what one of you said in the introduction I  think uh Abanach you mentioned mentioned that [30:14] um cars can interact with each other auto like and  what I'm interested in is like reducing traffic [30:22] especially like within Toyota cars but also like  within cars from different companies. Have you [30:27] worked or is there work being done on that right  now? Yeah, there's there's a lot of work being [30:32] done on that uh not just by Toyota but by by  lots of lots of other folks. Uh and so yeah, [30:37] you're you're mentioning this notion of vehicle to  vehicle or vehicle infrastructure communication. [30:41] That's been an area of investigation quite some  time and I think there's a couple of different [30:45] use cases. One is can we use that information to  basically aggregate information from the fleet to [30:50] give you more useful data. So for example, you're  driving on a road and you know the off-ramp has [30:54] like a a bunch of black ice at least you can  be warned ahead of that because another car [30:59] experienced that. The second one is can we use  that to improve traffic traffic flow. So can we [31:03] for example uh you know when you're when you're  going through an urban area you know can we time [31:08] your speed etc. so that we are able to actually  go through these lights you know with the least [31:12] amount of of backflow and stuff and people are  working on that. I think the challenge to that is [31:16] uh actually being able to get vehicle-to- vehicle  uh infrastructure and and hardware hardware into [31:21] the cars. And the big problem here is the benefits  really acrew only when a significant amount of [31:26] the fleet has it. And so for the early adopters,  they're not really getting any benefit for quite [31:30] some time. Uh and then you know how do you create  benefit that is valuable enough that people want [31:35] to do it? More recently, people are looking uh  rather than a specialized hardware and specialized [31:39] bands can we use uh cellular networks and things  like that. So there's a lot of work looking at [31:43] that and looking as at that as a way of you know  creating safer intersections for example so you [31:50] can connect to people's phones and and they know  cars are coming and they can indicate to you you [31:54] know maybe you should cross or you shouldn't  cross or convey your intent. You can imag even [31:58] imagine a crazy world with agents where you  know car has agent your phone has agent and [32:02] they all know what you're trying to do and these  agents interact and can kind of figure out what [32:06] the optimal thing to do is in any specific you  know environment to to maintain maintain safety. [32:12] uh and so um yeah I think there's there's lots  of ways in in which people are investigating this [32:16] uh though there's still some challenges with the  implementation other questions yeah uh so I work [32:23] in the uh dealership side of automobiles uh so I  was curious uh based on what you're seeing with [32:29] AI we've been building software to help them with  fixed stops and all their dealership kind of stuff [32:35] a lot of what you're mentioning obviously is going  to have an impact on dealers so could you comment [32:40] on maybe some short, medium, and long-term  things that uh all of us you have maybe you [32:45] can start uh cuz you have you know obviously  hundreds of dealers to deal with uh um but [32:50] then everybody else if you have some comments on  what does what does any of this mean good or bad [32:56] uh for automotive dealerships? Yeah, so we we  keep in close contact with our dealers. We have [33:02] uh you know significant engages with them and  one of the the main themes we've heard is this [33:07] huge desire to pull AI. They're like how can  we use it? tell us you know what are the ways [33:11] in which it's useful. I think the big pain points  we've heard from dealers are things around vehicle [33:16] service and maintenance uh warranty information  and and keeping track of that. And so there's a [33:21] bunch of I think in the short term use cases  that dealers themselves have identified that [33:25] I think you know AI could uh produce some real  tangible results and I think longer term there's [33:31] opportunities for us to think about you know  physical AI systems uh robots etc that can help [33:36] with some of these service tasks and and other  things within the dealerships as well and so uh [33:40] I don't have all the answers here but I think it's  a continual conversation with the dealers and like [33:44] I mentioned earlier our philosophy in case I uh  you know you missed it is really around amplifying [33:49] the person and so the dealers, you know, what  would be the most useful in their context and [33:53] how do we rethink the dealership workflow in the  context of AI rather than just sort of adding AI, [33:59] you know, peace meal into existing processes.  And so this is something that we're trying to [34:02] work most closely with our dealers to figure out.  We're selling through dealers as well. We're not [34:07] we're not a direct consumer brand. Um, and the  thing the conversations we're having aside from [34:12] the maintenance side of it, right? meaning um  sensor maintenance and sensor cleaning and all [34:17] the things that have to happen is the time, right?  So, if my if my car doesn't have to go in with me, [34:23] a dealer may never interact with the owner of  the car. Uh the car may bring itself in for the [34:30] service and tell the technicians or whoever has to  work on it all the things that need to be done can [34:35] be done at night while the the owner is sleeping.  The other thing we talk a lot about is the sales [34:39] process changes fundamentally. When you're talking  about a personal vehicle, that vehicle can go sell [34:43] itself. It can drive itself to your house and  take you through all the features and actually [34:49] go through that sales process. Now, hey, I'm a  marketing person. So, do I believe humans are [34:54] going to be cut out completely? No, because we  still want to have that human connection. But, [34:59] do we have to be limited by a dealership location  footprint? it changes things a lot because if the [35:06] cars can go out to customers as opposed the  customers have to come to the cars um it's a [35:12] different model. So we're playing around with a  lot of the different commercialization aspects [35:17] for dealers. Um as a matter of fact I'm speaking  at the California New Car dealers association. [35:21] And there have an event um in October and we're  going to talk about those sorts of things is how [35:26] do dealers have to think about autonomous vehicles  differently and how does the customer experience [35:32] the owner experience um change and evolve uh in  the world of autonomy. May mobility is a little [35:39] different in that we operate with fleets. Uh so we  don't deal directly with uh dealerships. However, [35:45] May is a company that really focuses on  partnerships. So we'll partner with fleet [35:50] managers. we partner with other uh entities that  help us maintain the vehicles etc. That's going [35:55] to be a big deal too right and now your vehicles  uh not only do they have the mechanical electrical [35:59] but they got these really complex AI systems that  have to be maintained uh calibrated etc. So uh May [36:06] has always been a company that we we don't try  to do everything ourselves we have the core uh [36:10] and that we partner with uh with others to help  us and part of that is uh people who can help us [36:15] uh maintain the vehicles and keep them uh on the  road. And I'll just say one more thing. Sorry, [36:21] I should have mentioned it earlier as well. I  think we see a lot of opportunities for dealers to [36:24] capture additional value because of AI. So value  that you're not capturing today, but that AI would [36:30] enable true personalization and customization uh  options that allow you to then maybe, you know, [36:34] one of the things we're looking at is can we use  AI to basically uh increase aftermarket sales by [36:40] allowing people to customize things. And so  there additional value streams that we think [36:44] AI will unlock that will be beneficial uh not  just for the OEM but for the dealers as well. [36:48] Other questions? Yeah. Uh, thank you for taking my  question. How do you see the liability environment [36:57] evolving? For example, if I'm in a level four  autonomous vehicle, it's in a crash, it's found [37:02] liable. Am I liable, the manufacturer? What how  do you see that playing out? Thank you. We I [37:08] can actually speak to that. We actually partner  with Marsh, uh, which is our insurance provider. [37:12] This is actually a really really interesting  area because for instance with a tensor robbo [37:17] car when it's in level four mode the owner has no  liability. It's it's a sorry liability. The owner [37:22] can do nothing. You can't operate the car. You  can tell it where to go but the operational um [37:29] you know function of the car is is ours. It's our  software. It's it's it's uh we're controlling the [37:34] car. So liability is with tensor in level four.  anything under level four because we also you can [37:40] drive the car. So level two for instance then it  is the liability of the owner or the driver uh the [37:47] person with the license in the car who's operating  the car and that's one of the reasons why we do [37:51] not have a steering wheel when we're in level  four but all other modes obviously have to have [37:54] a steering wheel and pedals. What's interesting  about insurance though is Marsh and we've talked [37:59] about it. We had to see uh at CES we did a whole  session on it and it's actually on our our YouTube [38:04] um uh channel as well and you can so you can hear  about it. But dynamic insurance pricing becomes a [38:10] real thing because the car's computer can actually  inform the insurance company how much of the [38:16] travel that month or that period that quarter was  done by the car liability on tensor and how much [38:24] of it was done by the owner. And so theoretically  you could have very dynamic insurance pricing [38:30] based on how you're operating the vehicle. And  I think those are some interesting things and [38:34] obviously because there's an onboard computer and  you're can pull that information. Even now I think [38:40] some insurance companies you get you know safe  driver discounts for you know being able to share [38:44] some of the data but dynamic insurance pricing is  a real thing and will need to be uh accommodated [38:49] for level four. you or is the implication from  what you said that a mile on level four will be [38:55] cheaper to ensure than a mile driven by a human.  A mile in level four would be cheaper because [39:00] there is no driver liability at all. So less  likely to incur an accident and less likely to [39:07] incur an accident which is we're willing to take a  liability, right? Because we trust our technology. [39:11] So any sort of level four vehicle or car in a  level four mode, you would hope that the company [39:18] that uh you're buying that vehicle from uh backs  their technology and so we do um and that's the so [39:24] we we take the uh insurance liability uh for uh  any sort of accident if it's level four. Jacob, [39:32] can you can you uh talk publicly about you guys  self-insure? Uh we do work with insurance partners [39:38] uh that work with us and and that includes both  insuring us as and our vehicles as well as we have [39:44] partnerships where we share data with them etc.  So it's an evolving landscape. I'm not an expert [39:48] on it but we do have partnerships. Do you guys  publish your your safety numbers like compared [39:54] to like regular miles driven by humans? Um I  don't think we publish them publicly right now [39:59] but yeah I know Whimo does but um other questions.  Hi uh thank you for all the great information very [40:06] stimulating thought question I had was how do you  take into account the non-deterministic nature of [40:13] AI into your decision-m say for example you're  going through a yellow or red light that may [40:18] change vary based on the the situation and in your  simulations curious how you handle that so sure um [40:25] I I talked a little bit about this earlier in my  talk there's a lot of uh situations uh where the [40:31] uncertainty is high there's bursts of uncertainty  you while you're driving. And in those times, [40:36] um, it's important to, you know, as a human, what  we do is we we'll like visualize the different [40:42] things that could happen like on the fly. And  we'll say, okay, uh, if I do this, if I do that, [40:46] you're it's very very fast and you'll make a  decision that's generally a little bit more [40:50] cautious than say if you were certain about what  was going on. And that's the kind of technologies [40:55] that you need. Everybody's kind of moving more  and more toward especially with these AI models [41:00] uh giving uh the inference system options and we  take that to its extreme at May where we actually [41:05] play out the options in a little simulation or a  little tournament sometimes we call it you know [41:10] who's going to win in a situation. So if there's  a lot of uncertainty that means there'll be a lot [41:14] of outcomes that are different from each other and  we're going to tend to pick options that slow down [41:19] maybe veer a little bit or etc. and and that will  happen for fractions of a second or a few seconds [41:26] and then everything clears up and and things  become much more um deterministic and then you [41:32] can kind of proceed with the your normal driving.  It's very humanlike in a way. Question. Uh hi, [41:40] my name is Matt and I'm from a different I work  for a different industry but I'm interested in [41:45] autonomous vehicle. Uh thank you for the  great session. Uh [sighs] actually when it [41:51] comes to this make the AI powered vehicle safer  it raised a big concern for me which is lack of [41:57] microphone in the you know autonomous cars. In the  previous session, Jacob pretty well explained the [42:05] uh use of the evidence from lighter, radar, and  cameras to train the AI model. And I'm glad that [42:12] uh Amy mentioned that uh they use 20 microphone,  but I assume that they are uh inside the car and [42:20] for you know the passenger. Uh I'm coming from San  Francisco and in San Francisco you know there are [42:27] hundreds of speakers around the city and every  Tuesday at noon they uh announce something and [42:34] they are for um uh alerting people about the  tsunami earthquake and something like that. [42:41] Imagine an slow day and no one in the street,  the passenger in the robo car and uh there's [42:48] no sign of tsunami but the speakers started  announcing the tsunami and to me the car should [42:56] uh pull over and shout the passenger just run  you save your life or go to the elevated area [43:05] uh to you know to save the passenger's life. So  uh I would like to know how the robo car handled [43:12] this situation. Who wants that one? I mean we we  have speakers inside the car and out external. So [43:20] the the AI system or our technology is listening  externally. So, for instance, emergency vehicles [43:29] are an autonomous vehicle can hear an emergency  vehicle way before a human passenger would ever [43:35] hear it inside and process those sounds. I can't  speak to a tsunami that's not a tsunami. Um, [43:41] but I would imagine that um, you know, in those  sorts of scenarios, it's not just the car. Um, [43:48] and we're not a rules-based AI system. So, it's  got to not only understand what it was saying, but [43:54] understand the context of what's going on in the  environment. So, um you know, our LLM is connected [44:02] and our car is connected to to the internet to  through 5G. So, I would hope that in those very [44:09] strange edge cases that the car did validate that  there was indeed a tsunami uh warning and not tell [44:15] its passenger to jump out. I I'll just add a quick  note. I think the meta point you're making is that [44:21] good drivers use sound. So I don't know how many  of you have been in a situation where your car [44:25] has started skidding uh and uh what do you hear  you know usually you know the sound from your [44:30] tires. So we've actually done work uh because  I uh on on how we can utilize that to kind of [44:35] improve vehicle control at the limits uh you  know especially when you're driving on snow or [44:39] you're doing aggressive lane change or something  like that. But I think your point is well taken, [44:43] right? The best human drivers utilize all senses  at their at their disposal and I think we want [44:49] the AI to do the same. Okay. So time has come. I  have one last question for each of you. I'm going [44:54] to bet each of you individually 100 bucks and you  have to give me a number. So 15 years from now, [45:01] we're going to abstract away from current. Uh  there are 90 million new cars more or less sold on [45:06] the face of the earth every year. What percentage  of new car sales 15 years from now are fully [45:12] autonomous globally? I will see what they really  believe. I mean, I would love to be optimistic, [45:21] but I think 15 years from now, I would hope we're  approaching 20 25%. Okay, good number. I think [45:30] it's hard because um it's it's pe the the drivers  with their individual cars that they switching [45:37] cars takes a long time. Maybe 10 years, 15 years.  You never know. I keep cars for 10 years. Give [45:42] us a number. Give us a number. I'm thinking I  justifying I'm giving him my chain of thought. [45:47] I I'm thinking that it's going to be the places  where delivery logistics where things will move [45:52] the fastest because there's a lot of advantages  to just taking the drivers out. Um you want his [45:57] number? Do you want his number? I want the number.  Come on, baby. I'll say uh you're in Vegas from [46:02] now 25%. Okay. 25 and yeah. Um so I think at you  know uh at Toyota we sell all over the world and [46:11] certainly in certain markets there is appetite for  this stuff. Uh but we see you know obviously huge [46:16] challenges you know imagine places that don't have  lane discipline. There aren't any autonomy systems [46:20] for for those places right. My feeling is you guys  don't like giving me numbers. Yeah. Yeah. Well you [46:25] know I think the real answer is I don't know. And  so I'm going to tell you what I I can speculate, [46:30] but my guess is less than 20%. Less than less  than 20%. I don't I don't think it's going to [46:35] be a case of in 15 years whether the technology  is viable, right? It's going to be a case. So I [46:41] think especially like Toyota and bigger OEMs will  be they will have level four systems in their car. [46:47] Um and that's not going to be the problem. The  problem is regulations, not even just consumer [46:53] adoption, but all the different policies. Even now  we're we're launching in Dubai. Our headquarters [46:58] is in San Jose. So we have California where you  know we have a driverless permit in California. [47:04] We're trying to launch in Europe. Europe  every single country and every city decides [47:08] they want to have a different way of testing and  validation. So you know for companies launching [47:14] Toyota probably is going to be one of the leaders  because they do have the scale and the budget. But [47:19] for us we're a small relatively small startup. We  have to go through all that process. Um and it's [47:25] Whimo is paving the way. I mean they've been  really good with working with policy makers [47:29] and trying to figure out how do they validate  level four technology and especially as we have [47:35] like for our in our case it's a brand new car.  So homologation was a big issue related to all [47:41] these new sensors and making sure the car is safe  not just the software is safe. Um all of those [47:46] process take time. So yeah I mean I would love  to be optimistic and say 25% but I think you're [47:51] right. It's probably going to be closer to minus.  Yeah. 10 plus or minus. Yeah. Not not a whole lot. [47:57] Yeah. Well, you've been a great audience. Great  engaged questions. Thank you. Round of applause [48:01] for our panelists, please. And if you if you got  questions, I'm sure you can grab them as they come [48:06] off stage here. But thank you very much for having  us and uh have a good rest of the conference.