What Does 'Self-Driving' Really Mean? A Bird's-Eye View of Autonomy

The Long Road to Autonomy: A Brief History of Self-Driving Cars

A car rolling down a highway makes hundreds of decisions every minute. Some are conscious — should I change lanes? — and many are nearly invisible, like the tiny steering corrections your hands apply to keep you centered in the lane. Teaching a machine to make all of those decisions, fast enough and reliably enough that you would trust it with your life, is one of the most ambitious engineering projects of our time. Before we write a single line of code, we need a clear mental model of what an autonomous vehicle is, where the technology came from, and what problems such a system must solve.

The dream of a car that drives itself is older than most people realize. It did not begin with Tesla, with Google, or with any of the well-funded startups of the 2010s. It began in 1980s Germany, in a small robotics lab at the Bundeswehr University in Munich, where a professor named Ernst Dickmann was experimenting with computer vision long before "computer vision" was a phrase anyone used outside of research.

The 1980s: When "Vision-Guided" Meant Two Black-and-White Cameras

Dickmann and his team built a Mercedes-Benz van outfitted with two cameras and a rack of computing equipment. The cameras fed images to algorithms that detected lane markings, recognized the painted lines on the road, and produced a steering command. In 1987, that van drove itself along a stretch of German autobahn at speeds of up to 90 km/h — a remarkable achievement given that the processing power of the era was a fraction of what a modern kitchen blender uses.

This is a useful place to pause and underline a pattern that will repeat throughout the history of self-driving cars: progress has always been gated by hardware, not just ideas. Dickmann's algorithms were clever, but they could only run as fast as the silicon of the day allowed. Every decade that followed saw the same dynamic — a clever algorithm waiting, sometimes for years, for the hardware to catch up.

The 1990s: Neural Networks in the Driver's Seat

By the early 1990s, American universities picked up the baton. Researchers at Carnegie Mellon University built a series of test vehicles called NavLab. NavLab 1, completed in 1986, was a Chevrolet panel van with a single camera and a sun workstation in the back. NavLab 5, finished in 1995, became famous for the "No Hands Across America" trip: the vehicle drove itself 2,797 miles from Pittsburgh to San Diego, with autonomous control engaged for 98.2 percent of the distance.

What made NavLab 5 special was a small neural network called ALVINN (Autonomous Land Vehicle In a Neural Network). ALVINN took a low-resolution camera image as input and produced a steering angle as output. It was trained on a few hours of human driving, and it learned the basic geometry of "look at the road, follow the lane." Modern deep-learning-based driving systems are direct intellectual descendants of ALVINN. The core idea — show a network many examples of good driving, and let it figure out the rules — has not changed in thirty years. What has changed is the size of the network, the size of the dataset, and the size of the computer running the network.

The 2000s: DARPA and the Coming-Out Party

The first truly public inflection point came in 2004, when the U.S. Defense Advanced Research Projects Agency (DARPA) hosted the first Grand Challenge in the Mojave Desert. The goal was straightforward and brutal: build a vehicle that could drive itself 142 miles through the desert along a predefined route. Fifteen teams competed. The prize was one million dollars. The result: none of the vehicles finished. The best performer, Carnegie Mellon's Red Team, traveled just 7.4 miles before getting stuck on a rock.

The failure was instructive. It taught the field two hard lessons. First, perception in unstructured environments is brutally hard — a desert full of rocks, brush, and dry lake beds is in some ways more difficult to navigate than a city, because there is less prior information to lean on. Second, the engineering around the perception — the mechanical reliability, the software stability, the way the system handled its own failures — mattered as much as the algorithm itself.

In 2005, DARPA ran the Grand Challenge again. This time, five vehicles finished, and a converted Volkswagen Touareg called "Stanley" from Stanford University won the two-million-dollar prize. Stanley was built largely by a team led by Sebastian Thrun, a German-born computer scientist whose next move would shape the entire industry.

In 2007, DARPA hosted the Urban Challenge, this time on a closed urban mock-up at a former Air Force base. Vehicles had to merge into traffic, navigate intersections, obey traffic laws, and handle other robotic cars as if they were other drivers. Six teams finished. Carnegie Mellon's "Boss" won first prize, and the field of urban autonomous driving went from "vaguely plausible" to "concrete engineering problem."

The 2010s: From Research Project to Real Product

In 2009, Sebastian Thrun, by then at Google, launched the Google Self-Driving Car Project. The team was small — about fifteen engineers — and the technology was based directly on the lessons of the Urban Challenge. For years, the project operated in near-total secrecy, with the cars occasionally spotted on California highways covered in sensors and decals. In 2016, the project spun out as a separate company called Waymo.

Meanwhile, a different approach was emerging from a company best known for electric cars. In 2014, Tesla began shipping vehicles equipped with a forward radar, a forward camera, and a forward-looking computer called the Autopilot Hardware 1. In 2015, Tesla released a software feature called Autopilot, which could combine adaptive cruise control with lane keeping on well-marked highways. The system was the first consumer-facing product that resembled "self-driving," and it sparked a wave of public attention — and public controversy — that has not subsided since.

Throughout the 2010s, a wave of new companies entered the field. Cruise, founded in 2013, attracted a multi-billion-dollar investment from General Motors. Argo AI, founded by alumni of Google and Uber, was funded jointly by Ford and Volkswagen. Aurora Innovation was founded in 2017 by Chris Urmson (formerly of Google's self-driving car project), Drew Bagnell (from Uber), and Sterling Anderson (from Tesla). Mobileye, an Israeli company that had built many of the cameras and chips used in advanced driver assistance systems, was acquired by Intel in 2017 for $15.3 billion. By the end of the decade, the field had consolidated into a recognizable industry with hundreds of billions of dollars in combined market capitalization.

The 2020s: Robotaxis, Cameras-Only, and a New Public Debate

The early 2020s saw two divergent philosophies emerge. Waymo, Cruise, and most of the traditional robotics companies followed a "Lidar-and-HD-Map" approach: install multiple expensive sensors, build centimeter-accurate maps of the operating area in advance, and use the redundancy to achieve the safety levels required for true driverless operation. In 2020, Waymo launched Waymo One, a robotaxi service open to the public in parts of Phoenix, Arizona. By 2023, the service had expanded to San Francisco, and Waymo was operating thousands of driverless trips per week.

Tesla pursued a different path. Instead of lidar, Tesla relied on cameras and a custom-designed inference chip called the Full Self-Driving (FSD) computer. The argument, articulated forcefully by CEO Elon Musk, was that humans drive with two eyes and a brain, and a system designed well enough could do the same with cameras and silicon. By 2024, Tesla's FSD system was available in "supervised" form — meaning the driver must remain attentive at all times — and the company was making bold claims about future "unsupervised" capability.

The two approaches