How we built the first real self-driving car (really)

Kyle Vogt
Published in
7 min readSep 11, 2017


Electric self-driving cars will save millions of lives and significantly accelerate the world’s transition to sustainable energy, but only when they’re deployed in large numbers. So, naturally, our singular focus at Cruise is to rapidly deploy self-driving cars at scale. Then we’ll drive down their cost and improve their capabilities, so we can deploy at an even larger scale.

While the videos we’ve previously released show the most advanced self-driving software ever demonstrated, the most critical requirement for deployment at scale is actually the ability to manufacture the cars that run that software. So, today, we’re unveiling the world’s first mass-producible car designed to operate without a driver. This isn’t just a concept design — it has airbags, crumple zones, and comfortable seats. It’s assembled in a high-volume assembly plant capable of producing 100,000’s of vehicles per year, and we’d like to keep that plant busy.

The car we’re unveiling today is actually our 3rd generation self-driving car, but it’s the first that meets the redundancy and safety requirements we believe are necessary to operate without a driver. There’s no other car like this in existence. In a few weeks, these cars will be a part of the fleet that carries Cruise employees anywhere in San Francisco using our app. For now, there will still be a human behind the wheel.

World’s first mass-producible driverless vehicle

In the near future you’ll see these cars on the road, and they’ll look like regular cars, but they’re actually some of the most technically advanced robots on the planet. They don’t need drivers, and there might not be anyone inside at all. They’ve been designed to emulate human driving behavior but with the human mistakes omitted. They don’t drink and drive, they don’t text while driving, and they don’t get tired. It has taken the collective effort of over two thousand people to create this product, and we believe that together we’ve managed to create something that will one day drive significantly better than any individual.

These robots didn’t appear overnight. We knew we’d discover new things along the way, so we took an iterative approach to development and built several generations of vehicles. In fact, by the time General Motors completed its acquisition of Cruise in mid 2016, we had already retrofitted our self-driving systems onto the Chevrolet Bolt EV platform to create our 1st generation test vehicles. We’ve already put hundreds of thousands of complex urban miles on these vehicles, and exposure to the many challenging situations we’ve encountered along the way has rapidly improved our software.

Cruise 1st generation test vehicle, retrofitted by Cruise in San Francisco

Hardware, however, is a different beast. No production car has been designed to operate without any help from a human. It’s a big jump to go from some assistance to none at all, perhaps even 100x as difficult. The number of critical systems that must be heavily modified, duplicated for redundancy, or newly designed means retrofitting an existing vehicle is out of the question. So behind the scenes and in parallel, a small army of our colleagues at GM, led by Doug Parks, began working with Cruise to develop a new car.

That new car was our 2nd generation test vehicle. It was designed by the same engineers who design regular cars, so it went through the same rigorous testing processes as most production vehicles. Most importantly, it was designed to be built on one of GM’s high volume assembly lines so that we could learn how to build at scale. We started rebuilding our systems using automotive grade components and automotive suppliers when possible, and started modifying and replacing code on the dozens of systems and controllers inside the vehicle so that they cleanly interfaced to the self-driving technology from Cruise. We even built some sensors and controllers from scratch because nobody else was building them. These 2nd generation cars have all the key elements for autonomous driving, but they don’t contain the redundancy and safety systems we believe are necessary for full driverless operation.

Cruise 2nd generation test vehicles, assembled at GM’s Lake Orion plant in Michigan

Shortly after we kicked off the design of our 2nd generation vehicles, we had yet another team start the design of our 3rd generation vehicles in parallel. As we found ways to improve the design or were able to make use of next generation sensing and compute technologies, we dropped those design changes into the 3rd generation vehicle. Unlike the previous generations, which were similar to Chevrolet Bolt EV design, the vehicles we’re unveiling today have almost completely new and fault-tolerant electrical, communication, and actuation systems that are unique to a driverless vehicle.

All of these new systems are important, because a driverless vehicle can’t rely on a human as a backup system. If something on a vehicle fails while there is an attentive human in the driver’s seat, they can yank the wheel or stomp on the brake pedal to avoid an incident. This isn’t the case for a car with no driver, so we built backup systems. And in some cases we built backups for the backups — and backups for those systems, too. Safety and validation teams have carefully considered plausible failure modes for all critical systems and fed changes back into the design. Our newest self-driving car might look like a regular car on the outside, but the vehicle’s core system architecture more closely resembles that of a commercial airplane or spacecraft. It’s a complex and time consuming process to design cars this way, but it’s the responsible thing to do.

Cruise 3rd generation self-driving car. Selected new parts in orange, modified in purple.

Hand-building a few hundred complex cars is tortuous and expensive, but it’s technically possible. People have done it. But things start to fall apart beyond that. The wiring harness in our new car has 4,085 wires and 1,066 connectors, meaning that without rigor and process, vehicles will spend most of their time in the shop while technicians chase “ghosts” (chafed wires, loose connectors, or anything else that causes things to suddenly stop working). Add rain, sea air, and humid weather and you’ve got a recipe for disaster unless you really know what you’re doing. As you can imagine, achieving massive scale with a low defect rate and high reliability is ridiculously hard. Cars are big, heavy, and have tens of thousands of parts. So you really need a well-run assembly plant to build something that works, such as the billion dollar plant we’re using in Lake Orion, MI.

GM assembly plants are state of the art facilities, and to observe the speed and precision at which vehicles are rapidly assembled from a pile of parts is a humbling experience (especially after building dozens by hand). The plant used to build our self-driving cars is massive, requiring the cooperation of over 1,000 people and spanning the area of 75 football fields. There are hundreds of robots, vehicles on tracks that weave, rotate, and climb through the facility, and processes that have been honed to near perfection over decades. Parts from thousands of suppliers feed into every station in the right order, at the right time. It’s a logistical symphony, and truly a sight to behold.

​Cruise 3rd generation self-driving car assembled in Lake Orion, Michigan. ​

While quality high-volume assembly ensures vehicles are produced as designed, the design still has to be right. Initial designs are never perfect, so our development methodology is optimized around speed of iteration. In some cases we’ve gone from a problem observed in the field to a change implemented at the plant within a single day. Today, we’re actually on our 3rd vehicle generation in 14 months (with more in the works). If we followed typical OEM development cycles this would have taken 6 years.

This kind of development requires full alignment, teamwork, rapid communication, and compromise between dozens of teams. Since Cruise is a fully owned subsidiary of GM, we’ve avoided the seemingly impenetrable wall of politics, competitive fears, NDAs, liability concerns, data ownership, asset financing, long-term alignment, and a host of other issues that plague other relationships and partnership models. We have none of that, which means more time is spent focusing on what matters most. We’re very proud to be a part of GM.

The launch of the world’s first mass-producible driverless car is a major accomplishment, but we’re not across the finish line yet. Building a few of these self-driving vehicles, or even a few hundred, won’t accomplish what we set out to do. And a self-driving business that depends on humans sitting behind the wheel is fundamentally unsustainable, so that won’t cut it either. We will achieve success by integrating the best software and hardware to deploy truly driverless vehicles at scale.