“Smart Barrier” Developed at the NSU Artificial Intelligence Center

An automatic barrier for parking, equipped with artificial intelligence, which allows vehicles to be admitted to a closed area without human involvement, not only cars with license plates from the “white list”, but also special transport and guest vehicles with single-entry passes, was developed by employees of the Research Center in the field of artificial intelligence (AI Center) of Novosibirsk State University.For a year already, on the territory of the Higher College of Informatics (HCI) of NGU, a barrier with a detector has been operating, which is equipped with an algorithm developed by them for raising the barrier when cars with numbers entered into the database pass through. But now the barrier will be controlled by an AI agent. This will expand the functionality of the device, increase its reliability, and minimize human interference in its operation.

II-agent— is an intellectual system capable of independently setting goals, building a step-by-step plan, making decisions, and performing actions to solve a task, using external tools (browsers or databases) without human involvement.

—In the new barrier gate, instead of a simple algorithm, an AI agent was used by us. The reason is that the algorithm works well in normal mode, but when non-standard situations arise, which we did not include in the algorithm, the agent is precisely effective. It can be assigned to perform any task that previously involved a human: for example, when the barrier gate lifting device malfunctions, to inform the developer or the customer about it, or to try to solve this problem independently if such an opportunity is available.II-agent can obtain the necessary data for this from various sources: databases or logs — automatically created files in which the system, program, or server chronologically records all occurring events,— said the programmer of the Center of Information Technologies of NSU, a lecturer at VKI NSUNikita Doroshenko.

The usual algorithm works as follows: it has access to a “white list” of car numbers, located opposite the barrier gate with video cameras and detectors. The camera transmits frames, the detector analyzes the camera image for the number, recognizes the license plate, compares it with the database of allowed cars, and, if matched, sends a command to the controller to open the barrier gate.

Agents work somewhat differently and perform more tasks. The detector sends them an event, and the agent does not simply check against a list of numbers and decide whether to raise the barrier or not; it acts more complexly. It can allow a car whose number is not in the database to pass through — for example, a special-purpose or service vehicle. It will issue a command to raise the barrier in front of an “ambulance” or a fire truck. Or it will notify the dispatcher that a car is parked in front of the barrier and obstructs passage. If the algorithm runs according to a ready-made scenario, then agents make decisions independently.

—Our agents are divided into zones of responsibility. Among them is an “evader” agent, who only raises and lowers the barrier in front of cars whose numbers are on the “white list.” The senior agent is engaged in more complex work — he records in his memory the car numbers and collects statistics about who is currently on the territory, who was let in, who is not on the “white list,” how many times each car entered the territory, and how long each time they stayed there.This same higher-level agent can “see” that for some reason the barrier gate was not raised in front of the car from the “white list” or has not been working for some time. In this situation, he first tries to solve the problem independently, referring to databases or logs, checking the system’s operation to understand whether it is working or not. Then he makes a decision: to inform the dispatcher that a car has approached and the barrier gate needs to be raised, or to warn the developer that a malfunction has occurred in the system that requires prompt elimination.— explainedNikita Doroshenko.

The local neural network is deployed locally on the VKI NGU server. Initially, the developers used models from the Qwen family of models, but then switched to Gemma-4 with 31 billion parameters.

The program is written using the LangChain frameworks, which simplify the development of applications based on large language models (LLM), and LangGraph, designed for creating complex multi-agent AI systems. Creating this system involves building the program in the form of graphs. Initially, it collects data from a “white list,” extracts events from the database, then sends information with a request to the neural network. Primarily, the LLM issues a verdict on whether to open the barrier or not.It returns this response in the format expected by the developers, forms the response, after which post-processing occurs, responsible for raising the barrier.

At the moment, the use and testing of an agent has already begun, which the developers called the “agent-boss,” responsible for “whitelists.” It is precisely from it that the agent-waiter currently receives the “whitelist.” The agent-waiter does not know when passes for someone run out, how many times to let through; the “boss” is responsible for this. The agent-boss has a web interface, thanks to which you can, using chat (in this case, communication occurs directly with the LLM) or forms, set passes, collect statistics, view which vehicles passed through the territory. The agent knows all this thanks to receiving data from the waiter.This part of the project is handled by the engineer of the Center II NGUAleksandr Shovkoplyas. The development of an agent who will be responsible for checking the operation of detectors and systems, as well as assigning tasks to the detector so that it begins analyzing the camera indicated by the developer, is handled by an employee of the II NU Center.Vasily Babushkin.

—On the territory of the VCI, the agent-‘washer’ successfully passed the regular situation check. It was not yet possible to check his identification and pass skills for access to the special transport area. The function for allowing vehicles that are permitted only a one-time pass into the territory is undergoing testing. When we bring our development to a final state that allows it to be deployed on another server, it is planned to deploy it at the parking lots near NGU buildings. In general, our ‘smart barrier’ can find application in various institutions, residential complexes, and closed parking lots,— saidNikita Doroshenko.Â

—For us, the barrier is a tool on which it is relatively safe to observe how AI agents will behave in the real world if they are trusted with the “red button.” After the pilot testing period, we will begin to expand the zone of responsibility of agents to include unmanned transport, climate control, and other applications,— commented the leading researcher of the Center for Artificial Intelligence at NSU, director of the Institute of Intellectual Robotics at NSUAlexey Okunev.