Cognitive Divergence Correction

The main articles are:

• Video analytics (terms, applications, technologies)
• Computer vision (machine vision)
• AI training

2025: Introduction of Neural Network Training Technology

Cognitive Pilot on December 18, 2025 introduced neural network training technology .

One of the key problems in training systems computer vision for autonomous control transport until recently was the discrepancy in the perception of the scene between man and. neuronet We are talking about cases when a person computer and vision see the same picture in different ways. As a rule, such events occur in difficult conditions (dirt, snow, rain, partially closed boundaries of objects, etc.).

For example, an unmanned tractor moves across a field where pieces of treated earth or turf cover the edge along which it is carried out. navigation Or a smart tram moves along tracks partially hidden under water, snow or ice, or robot focuses on visual markers distorted by shadow or lighting. As a result, AI the system sees the trajectory of movement in its own way, and the operator-driver in its own way.

Even with high detection accuracy, the network may misinterpret the context, "comments Gennady Savitsky, lead developer of Cognitive Pilot.

In such cases, a person usually determines the trajectory of movement unmistakably based on his experience and intuition, and the neural network "sees" it based only on visual signs. Therefore, in the process of collecting training data, systematic errors accumulate, which leads to their mismatch. AI literally ceases to understand "what they want from it," and begins to give erroneous predictions. As a result, the control accuracy of the autopilot system decreases, false positives, and the overall safety level of the autonomous system drops.

For the first time, Cognitive Pilot specialists were able to offer a comprehensive solution to this difficult problem. We could not find anything similar in foreign teams. Its decision, according to experts, will bring control accuracy to a qualitatively new level, where errors are practically excluded, and ensure the maximum possible traffic safety, "Savitsky continued.

Cognitive Pilot specialists have developed Cognitive Divergence Correction technology, which allows you to identify and quantify divergence between how a neural network and a person see the boundaries and semantics of a scene (understand the meaning of what is happening).

The developers of "Cognitive Pilot" have learned to automatically highlight the signs of the scene that a person determines based on his intuition and available experience, and on the basis of which he understands how the trajectory of the vehicle should go. A whole mechanism has been created - a divergence analyzer that highlights such features. Among them can be, for example, the shape and location of pieces of soil on the field, the intensity and contrast of borders, the spatial structure of the scene.

Next, the system automatically analyzes the discrepancies, calculates the metrics of the discrepancies between the boundaries highlighted by the model and ground truth (human). Consistent personnel (where the person and the network see the same) remain in the training set, and inconsistent personnel is sent to the'quarantine' for further analysis and decision on their use.

As a result, data consistency and, as a result, the quality of training and security of autonomous management systems are improved. Without solving this problem, it is impossible to ensure the necessary accuracy of agricultural operations of unmanned agro-equipment, as well as the complete safety of smart trams and other autonomous vehicles, - said Gennady Savitsky.

𝓲

Фото: Cognitive Pilot

The market segment of technologies of this class for December 2025 is only being formed, but according to experts, its prospects and significance are excessively high. According to Precedence Research, the global AI development market for unmanned vehicles amounted to $4 billion in 2024 and will reach 15.23 billion by 2030 with a dynamics of 24.1%.

Since the end of spring 2025, autopilot systems with integrated Cognitive Divergence Correction technology have already begun to be installed on smart tractors, completely unmanned and cab-free mini-tractors, as well as on Cognitive Pilot smart trams in St. Petersburg and in practice confirmed its effectiveness.

We have been providing the level of detection of objects of road and field scenes above the industrial one for quite some time. However, there were still rare situations in which errors could occur. But the fewer they remain, the more serious the efforts of developers require their correction. The results obtained in solving such problems are sometimes close to serious scientific discoveries. The latest steps to achieve maximum traffic safety are the most difficult. We are constantly working in this zone to avoid even exceptional cases of errors, "said Olga Uskova, General Director of Cognitive Pilot.