← 返回 Avalaches

在经过十几年以规则驱动的硬编码阶段后,Waymo 的模式由逐条道路规则教学转向端到端预测。文章称,Waymo 始于 2009 年,历经十余年与四代技术后,于 2020 年在亚利桑那州 Chandler 启动全面商业化。早期系统通过雷达、激光雷达(lidar)和多相机阵列采集道路数据,再借助仿真环境让车辆接触罕见边缘情境。第五代 Driver 于 2020 年推出,将此前 AI 与机型模型的混合方式改为以 AI 为主干。通过将在路上运行车辆的遥测数据与基础模型体系(driver、simulator、critic,以及训练于 Google Gemini 的视觉语言模型)融合,车辆开始对未见场景实现泛化,并能推断被遮挡行人的动作意图。

竞争正在扩大,新进入者尝试以 AI-first 技术路线。Wayve 与 Waabi 希望“从零起步”的设计胜过 Waymo 的先发优势,通过更少的摄像头和雷达,甚至部分取消 lidar,来压低成本。Waymo 已累计行驶 2 亿英里(约 3.22×10^8 公里)并载客运行,显示其已具备规模化运营,但 Waymo 仍依赖高端硬件且多为地理围栏限制下运行。Wayve 的 Alex Kendall 认为其端到端深度学习能形成可迁移的通用模型,从而广泛部署,尽管 Wayve 还未在公共道路上实现大规模验证。Tesla、Uber 背书的 Wayve 等也计划在更多区域快速扩张。

扩张路径显示速度正在加快:Waymo 目前覆盖 10 个美国城市,并表示年底前将在 17 个城市提供每周 100 万付费乘客行程,包括刚启动的伦敦。Waymo 预计到 2030 年占美国网约车市场超 7%。Uber 目标在全球 15 个城市配合合作运营商竞争,并布局 Lucid 与 Nuro 车队;Zoox 和 Tesla 也公布了激进路线图。中国运营方正将车队扩展至多座亚洲城市,WeRide 已于三月进入新加坡,Waymo 则在东京测试。行业资本和科技公司在 robotaxi 上投入数十亿美元:Uber 单独承诺了超过 100 亿美元的自动驾驶投资与车辆交付承诺,美国市场被估计为 1 万亿美元。

4578271e39c9.png



After decades of hard-coded planning, Waymo moved from rule-by-rule teaching to end-to-end prediction. The article says Waymo began in 2009, took over a decade and four generations, and launched full commercial operations in 2020 in Chandler, Arizona. Early systems gathered road data with radar, lidar, and multi-camera arrays, then used simulation to expose vehicles to rare edge cases. The fifth-generation Driver launched in 2020 shifted from mixed AI and model logic to AI as the backbone. By combining telematics from road vehicles with a foundation model stack—the driver, simulator, critic, and a Gemini-trained visual language model—cars began to generalize to unseen scenes and infer the intent of pedestrians hidden from view.

Competition is now widening as new entrants test AI-first approaches. Wayve and Waabi aim to use clean-slate design to move faster than Waymo’s first-mover advantage, using fewer cameras and radar and in some cases dropping lidar to cut cost. Waymo has already driven 200 million miles with passengers (about 3.22×10^8 km), indicating operational scale, but it still depends on high-end hardware and is often geo-fenced. Alex Kendall says Wayve’s end-to-end deep learning creates a general-purpose model that can deploy broadly, though Wayve has not yet proved this at public-road scale. Tesla and Uber-backed Wayve and other operators are also preparing global expansion.

Rollout trends indicate increasing speed. Waymo now serves 10 U.S. cities and says it expects 1 million paid rides per week across 17 cities by year end, including London. It is forecast to exceed 7% of the U.S. rideshare market by 2030. Uber targets 15 cities worldwide with partner operators and is deploying Lucid and Nuro fleets; Zoox and Tesla also publish aggressive roadmaps. Chinese operators are extending fleets across Asian cities, with WeRide expanding to Singapore and Waymo testing in Tokyo. Investors and technology firms are spending tens of billions on robotaxis: Uber alone has committed more than US$10 billion in AV commitments, while executives estimate the U.S. market at US$1 trillion.
2026-04-30 (Thursday) · 35820934e2b4611e2e5d0ce2fff0b32d05f14864