How unmanned vehicles work

In 2016, the volume of investments in unmanned vehicles amounted to 80 billion dollars, of which 40 were invested in microchips, 16 in sensors and 4 in artificial intelligence. It is speculated that this industry could provide a big boost for AI and machine learning algorithms. In addition, a large number of vacancies in this niche are already open today. But while the bright minds of programmers are working on solving the assigned tasks, we will tell you about how self-driving cars work and whom such progress can directly affect.

The idea of ​​unmanned vehicles originated at the beginning of the 20th century and was then implemented in the form of a car with an antenna, which was controlled from the passenger seat of the car following it. Nowadays, antennas are also widely used for monitoring unmanned vehicles, but not for all types.

Today, the market offers a variety of unmanned gadgets, from remote control cars to full-fledged unmanned vehicles. There are drones, drones and submarines, but what do we mean by unmanned aerial vehicles? Remote control by the operator? With the operator on board, but without his active participation? Let’s figure it out!

What’s in the think tank of a self-driving car?
There are two approaches to car learning: classical and neural network. The first consists of four modules: localization, recognition, planning and management. When working with the second approach, the car receives data from the cameras, sends them to the neural network, and it already determines where to turn and at what speed to go.

The implementation of the neural network approach requires a huge amount of input data for each of the possible situations in order to teach a machine to behave like a human. This is quite difficult to do, which is why many self-driving cars use the classic approach.

Path planning is almost as important as the car’s perception of the world around us. Therefore, let’s look at the path planning methods, there are only three of them, if we talk about the main ones.

1. Algorithms for finding paths on graphs

Qualitative algorithms for finding paths on graphs in relation to unmanned vehicles include A *. However, it also has disadvantages, for example, in high-dimensional spaces, the speed of work decreases. In this case, you need to understand that on the road itself there are neither vertices nor edges of cells and you will have to compose the graph yourself, taking into account the position of the car in space. But the more complex it is, the more loaded the graph becomes.

2. Optimization methods.

Optimization techniques allow you to impose constraints and penalties using functions. Penalties can be for speeding or getting too close to other objects. With this method, the space is explored aggressively. The algorithm does not look for an optimal way to avoid an obstacle, but simply “drives” in different directions.

3. Stochastic algorithms.

This method is characterized by high speed of work in large spaces, however, the optimal choice remains a big question. However, the data obtained with this method can be sent directly to the vehicle control system. For more details on these algorithms, see the collaborative works [PDF] of Bader Alshamari (University of Kuwait | KU) and Ovidier Kalina (Eastern Michigan University | EMU).