Autopilots for aircraft have made the most difference. Even stable aircraft require the continuous attention of a pilot in order to fly. In the early days of transport aircraft, this created very high demands on crew attention and high fatigue.
The first autopilot was developed by Sperry Corporation. It slaved a gyroscopic artificial horizon and magnetic compass to hydraulically operated rudder, elevator and ailerons. It permitted the aircraft to fly straight and level on a compass course without a pilot's attention, thus covering more than 80% of the pilot's total work load on a typical flight.
The straight-and-level autopilot is still the most common, least expensive and most trusted type of autopilot. It also has the lowest pilot error, because it has the simplest controls.
Modern autopilots generally divide a flight into taxi, take-off, ascent, level, descent, approach, landing, and taxi phases. Autopilots exist that automate all of these flight phases except the taxiing, and some incorporate automated collision-avoidance, as well.
Modern autopilots use computer software to control the aircraft. The software reads the aircraft's current position, and controls a fly-by-wire system to guide the aircraft. In such a system, besides classic flight controls, many autopilots will control throttles to optimize the air-speed, and move fuel to different tanks to balance the aircraft in an optimal attitude in the air.
Although autopilots handle new or dangerous situations inflexibly, they generally fly an aircraft with a lower fuel-consumption per mile than all but a few of the best pilots.
The autopilot reads its position and the aircraft's attitude from an inertial guidance system. Inertial guidance systems accumulate errors over time. These errors are corrected by using satellite navigation systems and altimeters. The disagreements between the two are resolved with digital signal processing, most often a six dimensional Kalman filter. The six dimensions are usually roll, pitch, yaw, altitude, lattitude and longitude.
The hardware of a typical autopilot is a set of five 80386 CPUs, each on its own printed circuit board. The 80386 is an inexpensive, well-tested design that can implement a true virtual computer. New versions are being implemented that are radiation-resistant, and hardened for aerospace use. The very old computer design is intentionally favored, because it's inexpensive, and its reliability and software behavior are well-characterized.
The custom operating system provides a virtual machine for each process. This means that the autopilot software never controls the computer's electronics directly. Instead it acts on a software simulation of the electronics. Most invalid software operations on the electronics occur during gross failures. They tend to be obviously incorrect, detected and discarded. In operation, the process is stopped, and restarted from a fresh copy of the software. In testing, such extreme failures are logged by the virtualization, and the engineers use them to correct the software.
Usually, one of the processes on each computer is a low priority process that continually tests the computer.
Generally, every process of the autopilot runs more than two copies, distributed across different computers. The system then votes on the results of those proceses. Extreme values are discarded before they can be used to control the aircraft.
Some autopilots also use design diversity. In this safety feature, critical software processes will not only have different computers, and several copies, but each copy will also be by a different engineering team. It is unlikely that different engineering teams will make the same mistakes. As the software becomes more expensive and difficult, design diversity is becoming less common because fewer engineering companies can afford it.