Greatest aggregate distance driven by an autonomous vehicle fleet

The greatest distance covered by a fleet of AI drivers (on public roads without a safety driver) is around 5 million miles (8 million km). The two biggest players in autonomous vehicle research – Cruise (a subsidiary of General Motors) and Waymo (a subsidiary of Alphabet Inc) – both reportedly achieved this milestone in early September 2023.

The concept of the self-driving car has been a familiar feature in science fiction for around one hundred years, but the advent of deep-learning neural networks in late 1980s led many researchers to believe it was finally close to becoming reality.

Although some early projects, such as Carnegie-Mellon University's 1989 ALVINN prototype, showed promise, it soon became apparent that the challenge was far greater than initially assumed and progress in the field slowed dramatically.

The current wave of self-driving car research, backed by tech giants and automakers, started in the early 2010s with a similar belief that recent advances in AI and sensor technology meant fully autonomous vehicles would soon be a reality. The current generation of autonomous vehicles use banks of high resolution LiDAR (light direction and ranging) scanners as well as conventional cameras to gather data about their environment, and deep-learning neural networks to interpret the inputs.

While the self-driving cars from companies such as Waymo and Cruise perform exceptionally well compared the first generation vehicles, an AI capable of driving the car in all situations (what is known as Level 5 Autonomy) remains – at least for now – tantalizingly out of reach.

The systems that are used in these trials on public roads have reached what is called Level 4 Autonomy; they have no on-board safety driver, and they are capable of autonomous operation (in the environment they're designed for) almost all the time.

When they encounter a novel or confusing situation, however – such as a policeman directing traffic with hand signals or a construction-related diversion – the vehicles still often need to call on remote human assistance. They're also optimized for urban environments with predictable grid-based road layouts and clear road markings.

It is hoped, however, that these millions of miles of real-world training data will eventually equip the AIs with enough experience to handle all the many and various edge-cases that autonomous vehicles can encounter on the roads.