Floor-by-floor strategy: Scientists have taught AI robots to ride an elevator

Russian scientists have developed an AI-based software solution that allows wheeled robots to use an elevator to move between floors. Algorithms help machines recognize and press the right buttons, wait for the doors to open, assess whether the cabin is free, enter it, select the desired floor and exit. Solving these tasks requires precise navigation, understanding of the world around you, and the ability to plan multi-step actions. The development paves the way for the creation of autonomous service robots for work in multi-storey spaces.

Why is it difficult for robots to ride an elevator

Scientists from the Moscow Institute of Physics and Technology (MIPT) and the AI Institute of Artificial Intelligence have developed a system that allows robots to use a passenger elevator. This ability is especially relevant for vehicles on wheels that cannot climb stairs on their own. The Ministry of Education and Science of the Russian Federation told Izvestia about the development.

— Modern mobile robots already perform many useful tasks, but their operational radius is often limited to one floor. Ladders for most wheeled platforms remain an insurmountable obstacle. The only way for such machines to gain freedom of movement in a multi—storey building is to learn how to interact with elevators," explained Dmitry Yudin, one of the developers and head of the Laboratory of Intelligent Transport at the Central Research Institute of Physics and Technology.

According to him, this task is simple for humans, but it is a difficult challenge for artificial intelligence. The robots must find the elevator and figure out how to use it. This includes recognizing and pressing the necessary buttons, waiting for the doors to open, assessing whether the cabin is free, entering it, selecting the desired floor, driving and finally exiting.

Performing these actions requires machines to navigate accurately, understand the world around them, and plan several steps ahead.

— The proposed approach combines two components. First, the robot builds a three-dimensional scene diagram using data from the cameras. In it, nodes are objects (elevator, door, buttons, etc.), and edges are the connections between them. The diagram gives an understanding of where the robot is and what surrounds it," said Dmitry Yudin.

Investing in robots: the Russian Federation has launched the production of smart technology components

How import-substituted parts will help the development of medicine and the aerospace industry

Then the second component comes into play, he continued. This is a large language model — a program similar to the technology that underlies Chat GPT. She acts as an internal planner for the robot.

The model receives a command from a human and, correlating with the constructed "map", transforms the task into a chain of simple actions understandable to the robot. Then the machine sequentially executes the constructed algorithm.

How to teach machines to interact with humans

To test the development, the specialist said, a realistic virtual environment was created — a digital twin of a three-story building with an interactive elevator. Experiments in it have shown the high efficiency of the proposed approach.

In the future, the research team intends to improve the program in order to teach machines to work in dynamic environments. For example, you can use an elevator together with humans and other robots, giving way or waiting for your turn.

According to scientists, the development is an important step towards the creation of autonomous robotic service complexes adapted for operation in a complex human environment — offices, business centers, hospitals, restaurants and other public places.

— The task of using an elevator, which is intuitive for humans, is difficult for machines, because in addition to the algorithm of the machine, the robot needs to take into account many related factors. For example, the location of buttons, speed of movement, safety, passenger convenience, waiting time, and others," explained Sergey Bazykin, Doctor of Technical Sciences, Professor, Head of the Department of Instrument Engineering at Penza State University.

We translate it into understandable language: robots will be taught to speak humanly to each other

How new technology will make it easier to control machines with artificial intelligence

To overcome these difficulties, robot engineers create standardized instructions. However, the movement of people is less structured and more unpredictable. Therefore, navigating in a crowd is more difficult than traffic algorithms, he added. In order for the robot to navigate freely in a crowded space, it is necessary to refine the relevant functions, as well as prepare amendments to laws and take into account a number of ethical issues. At the same time, a universal crowd simulation model has not yet been created.

The ideal service robot is an assistant, not a substitute for a human. Such systems are designed to help and perform routine tasks. They simplify people's lives, but they do not compensate for human communication and warmth, the expert said.

— The ability to navigate is now lacking not only for robots, but also for humans. Therefore, semantic networks that describe models of the world and link objects, goals, and actions into a single logical structure are in demand," said Anna Pyataeva, head of the Artificial Intelligence Center at Siberian Federal University.

This achievement, when the system not only responds, but helps to find a way, is one of the most important shifts in modern AI, she added. In management, in science, and in complex industrial processes, it makes everything "clearer." Today, clarity is the most scarce and most valuable function of AI.