Self-driving cars are among the most anticipated technologies. They are supposed to make mobility more accessible to people with disabilities (1), decrease traffic related fatalities by up to 90% (2) and give the world economy a seven trillion dollar boost (3). In this article, we explain the key technologies that have accelerated the development of self-driving vehicles and how they are integrated to create autonomous vehicles.

The first machine that is considered a self-driving vehicle was the Stanford Cart, which was built and further developed in the 1970s as a prototype for rover vehicles that could be used in space missions on other planets. It was able to detect obstacles in its vicinity and plan its route accordingly. This level of autonomy was required, since remote controlling a vehicle on other planets is not feasible due to the long time it takes to send commands back and forth – three minutes one way in the case of Mars. However, the cart was very slow and moved only 1 m every 10-15 min, because it needed the time to calculate its next movements (4).

Technology has advanced significantly since then, but the core design principles of the Stanford Cart can still be found in modern self-driving cars. These principles consist of three components: a sensory component, a computational component and an actuator component. The sensory component gives access to the vehicle’s surroundings via cameras, radar, lidar (a sort of radar using light) and satellite positioning. The computational component interprets the sensory input, makes predictions about how a given situation will develop and takes decisions accordingly. These decisions are then executed by the actuator component which, for example, triggers the brakes or steers in a certain direction. These components are configured in a closed loop, meaning that while a decision is being executed, it is re-evaluated and adjusted using the newly acquired sensory input. This constant re-adjustment allows self-driving cars to be flexible enough for the use on public roads.

While cameras, other sensors and machine controlled actuators have been established technologies for decades, the computational component that links them was never powerful enough. This becomes clear, when looking at its main tasks. Firstly, it has to integrate different sensory inputs to create a map of the vehicle’s surroundings. Secondly, it needs to reliably detect and classify objects – like other cars, traffic signs or pedestrians. Then, it must take into account prior information to extrapolate the current situation (e.g. whether a pedestrian will continue crossing the street). Finally, it has to decide on a reaction that will move the vehicle closer to its destination while complying with the law and avoiding accidents. To make it even more challenging, all these tasks must be executed in close to real time (5). All this only became feasible recently through an increase in processing power and the rise of machine learning. Especially the latter opened the door for object classification and decision making, since it removed the need to manually program individual objects and scenarios. Instead, engineers use data from simulations and real traffic situations to train an artificial intelligence to drive a car reliably (6). 

However, there are two limitations to this approach. One is that an artificial intelligence will only be able to handle situations that are similar to the data it was trained on. If it encounters an unfamiliar situation, its response will be unpredictable and potentially dangerous. This was illustrated in a tragic way, when a Tesla on autopilot misclassified a white truck as sky and crashed into it, killing its passenger (7). The other limitation is the time that the artificial intelligence needs to make its decisions. In slowly developing and predictable situations, like on a highway with low traffic, self driving cars perform very well, since they can both anticipate the future robustly and have a long time to plan their behaviour. In suddenly changing and unpredictable situations, for example an animal entering the road, self-driving cars have to react quickly and might not have the time to find and execute the ideal response (8).

Individual self-driving cars cannot harness the full potential of this technology. This can only be achieved through the connection and coordination of multiple cars. Connected self-driving vehicles can coordinate their braking and acceleration behaviour to avoid creating traffic jams or accidents (2, 9). Moreover, groups of vehicles – so called cohorts – can communicate with traffic lights to coordinate a green wave on demand (10), thus making traffic flow more efficient.

However, self-driving car technology also has its drawbacks. Some studies suggest that self-driving vehicles will lead to an increase in traffic volume, since more people – like children, elderly or people with disabilities – will have access to them via rental services like Uber (11,12). This may even lead to the disappearance of environmentally friendly public transportation systems like trains or busses, due to the high level of flexibility and convenience that self-driving cars promise. This would lead to a further increase in traffic volume, because more people would travel in individual cars, which have a low passenger capacity relative to their size (13). Moreover, the interconnected nature of self-driving vehicle infrastructure opens the potential for surveillance and tracking of individual’s movement (14). Furthermore, it is unclear at the moment whether the car’s owner or its manufacturer is liable in case of an accident (15).

Despite these issues companies like Waymo and Uber recently started to roll out early versions of self-driving car rental schemes in selected locations (16, 17). Waymo is even currently testing their technology on self-driving trucks for the logistics sector (18). With technology companies already exploring the market for self-driving car technology, it is crucial that the legal and structural issues that come with it will be addressed soon to pave the way for self-driving cars.