内容来源：https://news.microsoft.com/source/emea/features/ai-that-drives-change-wayve-rewrites-self-driving-playbook-with-deep-learning-in-azure/

内容总结：

伦敦街头实测：Wayve以颠覆性AI方案重塑自动驾驶未来

伦敦市中心一个繁忙的十二月早晨，阴雨绵绵，交通拥堵。然而，一辆搭载Wayve自动驾驶技术的电动汽车正从容穿行其中。当一名行人突然从停靠车辆后闯入车道时，车辆平稳而果断地自动刹停，全程无需安全员干预。这看似寻常的一幕，背后是一家英国初创公司对自动驾驶技术路线的彻底重构。

与许多同行从复杂的规则编码和多传感器融合起步不同，成立于2017年的Wayve选择了一条“另类”路径：专注于打造一个纯粹的、基于端到端深度学习的AI驾驶员系统。该系统的核心是一个受人类大脑启发的神经网络模型，主要依靠摄像头感知环境，通过海量数据学习驾驶逻辑。

“我们真正将自动驾驶视为一个AI问题，并构建了一个数据驱动的、端到端的深度学习技术栈。”Wayve联合创始人兼首席执行官亚历克斯·肯德尔表示。这一策略旨在开发一个通用性强的AI驾驶方案，只需短暂适配，便能安装在不同品牌、型号的车辆上，并适应全球各个城市的道路环境。

为实现这一雄心，Wayve深度依托微软Azure云平台。利用Azure存储、Azure Databricks以及基于Azure Kubernetes服务构建的AI基础设施，Wayve连接了数千个图形处理器，打造出一台能够训练和验证其自动驾驶模型的灵活“超级计算机”。目前，搭载Wayve技术的车辆已在英国、美国、德国和日本的城市中开展测试。

Wayve的差异化战略获得了市场认可，公司已累计融资13亿美元。微软是其重要支持者之一，双方于2025年10月达成新协议，显著扩大Wayve对Azure服务的使用，并签署战略框架协议，将在技术推广、市场营销等多方面深化合作。此外，Wayve也与优步、日产等企业达成合作，计划今年在伦敦启动有限度的载客试运营，并于2027财年开始日产车型的量产搭载。

Wayve工程师将AI模型的训练过程类比为人类学习驾驶：首先需要长达“16年”的“预训练”，以建立基础的空间感知和协调能力；随后才是具体的道路规则学习。公司通过测试车队采集的视频、模拟环境数据等，在Azure的强大算力支持下，持续“教导”模型应对复杂动态环境。

在实际道路测试中，系统不仅能够处理常规交通场景，还能良好应对训练中未见的“长尾场景”，例如为横穿道路的动物减速。肯德尔展望，自动驾驶服务将深刻改变城市生活，通过提升车辆共享利用率，减少对停车位的需求。他认为，Wayve所代表的“具身AI”趋势，即将人工智能融入物理世界，将在未来十年于交通、物流、医疗、制造等多个领域带来巨大变革。

从伦敦街头到全球合作，Wayve正凭借其以深度学习为核心的颠覆性方案，试图为自动驾驶领域书写新的规则。

中文翻译：

AI驱动变革：Wayve借力Azure深度学习重塑自动驾驶规则

伦敦——在苏荷区一个异常繁忙的周四早晨，十二月灰暗的天空飘着冷雨。车流时走时停，但多半是停滞不前，连人行道也拥挤不堪。

终于，在宏伟的大英博物馆旁，车流开始恢复些许活力。一辆自动驾驶的四门电动轿车内，安全操作员双手掌心向上置于大腿上，虽未主动操控，却时刻保持警觉。车辆无需任何人工干预，自行平稳驶向特拉法加广场。

片刻之后，一位行色匆匆的男子从停靠的车辆后突然步入车道。这辆由AI操控的轿车果断刹车停稳，车内四位乘客随之微微前倾；而那位粗心的行人头也不回地穿过了马路。安全操作员全程未触碰踏板——车辆完全在自主行动。

如今，由AI驱动的自动驾驶汽车已驶入许多大城市的街道。但为我们提供此次行程的Wayve公司，自2017年在英国剑桥成立之初，便选择了一条与众不同的道路。

本质上，Wayve打造了一个AI驾驶员系统。无论车型品牌，也无论身处何国何市，该系统理论上都能安装于任何新车，仅需数周微调即可驾驭车辆。这一方案依托一种受人类大脑启发、名为“神经网络”的AI模型。Wayve的AI驾驶员主要依靠摄像头实现安全的点对点导航。

“我们真正将自动驾驶视为一个AI问题，并致力于构建一个数据驱动、端到端的深度学习技术栈。”Wayve联合创始人兼首席执行官亚历克斯·肯德尔在其位于伦敦国王十字区的工作室表示。

为实现目标，Wayve正借助微软Azure的强大能力。具体而言，它利用Azure存储、Azure Databricks以及结合了Azure Kubernetes服务的Azure AI基础设施，将数千个图形处理单元连接成一台灵活的超级计算机，用以训练和验证自动驾驶AI模型。

搭载Wayve技术的车辆在后备箱配备了一台强大的中央计算机，其中预装了Wayve的AI程序。通过车载摄像头，AI模型能够识别路标和交通信号灯，即使在伦敦这样繁忙的城市，也能感知环境并做出相应决策。目前，配备Wayve系统的车辆已在英国、美国、德国和日本的多个城市投入运营。

“起步时，我们采取了一种非常规路径，”Wayve联合创始人兼首席执行官亚历克斯·肯德尔说，“这一理念延续至今。我们真正将自动驾驶视为一个AI问题，并致力于构建一个数据驱动、端到端的深度学习技术栈。”

灵活可扩展的战略

这一战略与该领域的其他竞争者形成鲜明对比。后者最初采用基于规则的人工设计方法，将驾驶分解为不同问题集，并在车辆中集成复杂的传感器和计算机阵列。

Wayve则希望采用一种更通用、灵活的方法，能够快速扩展并被不同汽车制造商部署。它利用深度学习构建了一个神经网络——这是一种受人类大脑工作机制启发而设计的计算机算法。它由多层互连节点构成，能够从视频、其他形式的传感器数据乃至模拟环境（类似电子游戏）中学习模式。

“我们的目标不是构建完整的垂直技术栈，”肯德尔表示，“我们不造车，不构建自己的云基础设施，也不建立自己的出行网络。”

“我们的专长在于AI驾驶员系统。我们旨在与最强大、最优秀的伙伴合作，无论是汽车公司、像Uber这样的出行平台，当然还有微软以及支撑我们一切工作的Azure基础设施。”

“我深感感激的是，微软对Wayve下了赌注，”他说，“在我们对抗所有其他自动驾驶巨头时，微软很早便以合作伙伴身份给予了我们支持。”

自成立以来，Wayve已融资13亿美元。微软是其坚定的支持者之一。2025年10月，Wayve与微软就使用Azure服务达成新协议，此举显著扩大了Wayve对Azure服务的使用。两家机构还签署了战略框架协议，意味着它们将继续以多种方式合作，包括将开发中的技术推广至其他汽车制造商，并在市场营销和销售方面协作。其他公司也在与Wayve制定合作计划。

在6月宣布的与Uber的合作中，公司计划今年在伦敦使用配备Wayve系统的车辆启动小范围的载客服务试运营。Wayve还与日产汽车达成协议，后者将在2027财年开始量产搭载Wayve系统的汽车。

“我们能够拿到日产在日本的一款全新车型，那是一个我们从未进行过驾驶测试的国家，”肯德尔说，“仅仅四个月后，我们就能让这款新车证明，我们的系统可以在东京全城实现自动驾驶。”

亚历克斯·佩尔辛是Wayve的首席工程师，他领导公司的“预训练”团队，负责开发作为AI驾驶员核心的模型。

“我们喜欢用的类比是：人类学习驾驶前，已有十六七年的时间学习空间感知、手眼协调等能力，”他解释道，“然后他们可能接受约40小时的驾驶课程，学习交通规则和车辆操控。预训练就相当于那最初的十六年。”

与微软共创全新可能

利用从测试车队收集的视频和其他数据，以及模拟数据（可类比电子游戏）和其他类型数据，Wayve的工程师们正在教导AI模型如何在动态环境中安全导航。

“模型在学习物体如何在空间中移动、不同摄像头的视角如何关联、这些视角如何与操作动作相关联，以及速度等因素如何影响未来世界的呈现方式。”佩尔辛说。

他补充道，训练Wayve AI模型所需的数据密集型系统依赖于微软的大规模服务能力。他提到Azure Blob存储（blob是“二进制大对象”的缩写，此处指Wayve生成的数PB视频及其他类型数据）以及Azure Kubernetes服务系统，这些工具对于实现Wayve支持模型训练和运行所需计算能力的目标至关重要。

佩尔辛回顾了Wayve与微软如何协作开发工具，帮助Wayve创造出真正创新的成果。

“一个具体例子是，AKS过去只支持1000个节点，”佩尔辛说。一个节点通常指一台可能运行多个GPU的服务器，而集群则是一组节点。“我们需要比这更大的单一集群，现在该服务已支持5000个节点。这意味着我们无需自行搭建和管理Kubernetes服务……从而加速了我们自身的开发进程。”

机器学习工程师玛尔塔·沃林斯卡在Wayve的驾驶性能团队工作。她的任务是使模型适应配备不同摄像头设置以及雷达、激光雷达等其他类型传感器的各类车辆。

她指出，新车已具备许多AI功能，如车道检测和一定程度的辅助驾驶，但Wayve的技术将事情提升到了不同层次。

她说，令她和其他Wayve工程师及计算机科学家印象深刻的是，模型对于训练中可能未曾遇到的实际路况反应出色。

“比如为过路的鹅或松鼠减速之类的情况，”她说，“正是这些长尾场景，我们的模型能够很好地泛化处理。”

自动驾驶汽车的益处

在从Wayve伦敦总部（靠近国王十字区）前往特拉法加广场的行程中，我们充分体验了英国首都交通的复杂性。

起步平稳，频繁的停车亦然。虽未遇到鹅或松鼠，但车辆明确识别并停车避让了另一位粗心的行人——这次是在信号灯变红后横穿马路。它将四位乘客安全送达特拉法加广场并返回，全程无意外发生。安全操作员在其规划的路线中从未需要干预。

Wayve首席执行官肯德尔对Wayve及其竞争者可能在伦敦及其他地方产生的影响充满热情。

“我认为伦敦市民将会对自动驾驶汽车服务感到欣喜，因为它带来的益处是巨大的。”他说。

他表示，无人驾驶汽车还可能通过减少对停车位的需求来改变城市环境，因为自动驾驶汽车可以共享或租赁，从而提高使用效率，减少闲置停放时间。肯德尔指出，Wayve正在开发的技术最终是更大趋势——“具身AI”——的一部分，这一趋势尚未像Copilot这样的大型语言模型那样获得广泛关注。

“我认为在未来十年，我们将看到具身AI的兴起，将AI带入物理世界，”他说，“这为我们带来的机遇，当然涉及我们生活中与物理交互相关的绝大部分领域，无论是自动驾驶汽车、物流、医疗保健、机器人技术、制造业还是家用机器人。物理世界中的所有这些应用都能受益于AI。”

英文来源：

AI that drives change: Wayve rewrites self-driving playbook with deep learning in Azure
LONDON – On a more than typically busy Thursday morning in Soho, the gray December sky spat rain. Traffic was stop-and-go, but mainly stop. Even the sidewalks were congested.
Finally, alongside the imposing British Museum, the flow of cars and trucks regained some momentum. Inside a four-door EV sedan that was driving itself, a safety operator sat passively but alertly behind the wheel, hands resting palm up on his thighs. The car glided forward without any assistance from him, en route to Trafalgar Square.
A few moments later, a harried man stepped into our path from behind a parked car. The AI-guided sedan braked firmly to a halt, giving the four passengers inside a gentle shake; the heedless pedestrian crossed the street without looking back. The safety operator had not touched the pedals; this car was acting independently.
Self-driving cars, powered by AI, roll through a number of big city streets these days, but the company behind our ride, Wayve, took a different route when it was founded in Cambridge, U.K., in 2017.
In essence, Wayve has built an AI-powered driver who could potentially be installed in any new car, no matter the make or model, and drive it, in any country or city, with just a couple weeks of fine-tuning. This approach relies on a form of AI model inspired by the human brain known as a “neural net.” Wayve’s AI Driver mainly uses cameras to safely navigate from point to point.
“We’re really approaching autonomous driving as an AI problem and building a data-driven stack with end-to-end deep learning.”
Alex Kendall, co-founder and CEO of Wayve, at its workshop in the King’s Cross neighborhood of London. Photo by Chris Welsch for Microsoft.
To achieve its goals, Wayve is harnessing the power of Microsoft Azure. In particular, it is using Azure Storage, Azure Databricks and Azure AI infrastructure with Azure Kubernetes Service to connect thousands of graphics processing units into a flexible supercomputer that can train and validate the AI model for autonomous driving.
A car equipped with Wayve technology has a powerful central computer in the trunk that is pre-loaded with the Wayve AI programming. Through the car’s cameras, the AI model can read road signs and traffic lights, and even in a busy city like London, perceive its environment and act accordingly. There are now Wayve-equipped vehicles operating in cities in the U.K., the United States, Germany and Japan.
“When we started, we were building a very contrarian approach,” said Wayve co-founder and Chief Executive Alex Kendall. “It’s still with us today. We’re really approaching autonomous driving as an AI problem and building a data-driven stack with end-to-end deep learning.”
A flexible and scalable strategy
This strategy contrasts with that of other competitors in the field, who started by hand-engineering a rules-based approach that tackled driving as different sets of problems and integrated a complex array of sensors and computers in the vehicle.
Wayve wanted a more generalized and flexible approach that could be scaled up quickly and deployed by different carmakers. It used deep learning to construct a neural network – a computer algorithm inspired by our understanding of the way the human brain works. It is made up of layers of interconnected nodes and learns patterns from data like video, other forms of sensor data and even simulated environments (in a way like a video game).
“We’re not aiming to build the full vertical stack,” Kendall said. “We’re not building our own cars. We’re not building our own cloud infrastructure. We’re not building our own mobility network.
“Our expertise is in the AI Driver, and we aim to partner with the biggest and best, whether it’s a car company or a mobility platform like Uber, or of course with Microsoft and the Azure infrastructure that underpins everything we do.”
“What I’m so grateful for is that Microsoft made a bet on Wayve,” he said, “backed us as a partnership quite, quite early on when we were going up against all of the other self-driving giants.”
Wayve has raised $1.3 billion since its inception.
Microsoft is among its believers. In October of 2025, Wayve and Microsoft agreed to a new deal on Wayve’s use of Azure services, a commitment that significantly expands Wayve’s use of Azure services. The two organizations have also signed a Strategic Framework Agreement, which means they will continue to work together in a variety of ways, including expanding the use of the technology being developed to other car and vehicle makers and collaborating on marketing and sales. Other companies are also making plans with Wayve.
In a collaboration with Uber, announced in June, the company plans to start operating a limited trial of passenger service in London with Wayve-equipped cars this year. Wayve has also announced a deal with Nissan, which will begin mass production of Wayve-equipped cars in fiscal year 2027.
“We were able to take a new vehicle from Nissan in Japan, a country where we had never driven,” Kendall said. “And in just four months, we were able to take this new vehicle and show that our system could drive autonomously all throughout Tokyo.”
Alex Persin is Wayve’s principal engineer. He leads the company’s “pre-training” team, developing the model that is the AI Driver.
“The analogy we like to use is that when a human learns to drive, they have 16 or 17 years of learning spatial awareness and hand-to-eye coordination and things like that,” he said. “And then they have maybe 40 hours of driving lessons where they learn the rules of the road and how to handle a car. Pre-training is that first 16 years.”
Working with Microsoft on something new
Using video and other data gathered from its fleet of test cars, as well as simulated data (think video games) and other kinds of data, Wayve’s engineers are teaching the AI model how to navigate safely through dynamic environments.
“The model is learning how objects move in space, how the views from the different cameras relate to each other, how they relate to the actions and how things like speed affect what the world will look like in the future,” Persin said.
He added that the data-hungry system for training Wayve’s AI model relies on Microsoft’s large-scale capacities. He cited Azure Blob Storage (blob is short for binary large object – and in this case means petabytes of video and other sorts of data being created at Wayve) and the Azure Kubernetes Service (AKS) system as tools that were essential in reaching Wayve’s goals to support training and the computational demand required to run the model.
Persin reflected on how Wayve and Microsoft have collaborated on the tools that helped Wayve create something truly new.
“One concrete example is AKS used to only support 1,000 nodes,” said Persin. A node is usually one server that may have up to several GPUs operating in it, and a cluster is a group of nodes. “We wanted single clusters bigger than that, and now the service supports 5,000 nodes, which has meant that we didn’t have to go and run our own kubernetes service ourselves … so that has accelerated our own development.”
Marta Wolinska, a machine learning engineer, works on Wayve’s driving performance team. Her work is adapting the model to different types of vehicles with different camera setups and other kinds of sensors, like radar and lidar, which refers to light detection and ranging.
She points out that new cars already incorporate many AI features, like lane detection and some degree of assisted driving, but that Wayve’s technology takes things to a different level.
She said what has impressed her and other Wayve engineers and computer scientists is how well the model reacts to real-world situations it might not have encountered in training.
“Like slowing down for geese crossing the road or squirrels, that kind of thing,” she said. “It’s really those long-tail scenarios that we generalize to really well.”
The benefits of self-driving cars
During our Wayve-equipped car’s trajectory between Wayve’s London headquarters, near King’s Cross, to Trafalgar Square, we got a good sampling of the complexity of the British capital’s traffic.
Takeoffs were smooth; the frequent stops too. No geese or squirrels were encountered, but the car did clearly see and stop for another careless pedestrian — this one crossing after the light had turned. It delivered its four passengers to Trafalgar Square and back without incident. The safety operator never needed to intervene during the trajectory he had planned.
Wayve CEO Kendall is enthusiastic about the impact Wayve and its competitors might have in London and elsewhere.
“I think Londoners are going to be delighted by self-driving car services because the benefits it brings are enormous,” he said.
He said driverless cars could also change the urban environment by reducing the need for parking spaces because self-driving cars could be shared or hired, making them more productive and spending less time in parking spots. Kendall said that the technology Wayve is developing is ultimately part of a larger trend — “embodied AI” — that has not gotten as much attention as large language models like Copilot.
“I think over the next decade we’re going to see the rise of embodied AI bringing AI into the physical world,” he said. “What this gives us the opportunity to do is, of course, the enormous part of our lives that involves physical interactions, whether this is self-driving cars, logistics, health care, robotics, manufacturing, domestic robotics. All of these applications in the physical world can benefit from AI as well.”