Orientation — server: robots have been taught to find their way according to the principle of human memory

Russian engineers have developed a new technology to determine the location of robots without using navigation systems. It is based on the machine's fixation of basic landmarks and the relationships between them, as the human brain does. This approach makes it possible to avoid building cumbersome detailed maps, save computing power and find the way even in conditions of continuous changes in the position of objects. According to experts, the development opens up the possibility for the widespread introduction of autonomous robots.

How a robot looks for a way

MIPT specialists, together with colleagues from the Federal Research Center for Informatics and Management of the Russian Academy of Sciences and the AI Research Institute, have developed a new technology for robot orientation in space, which allows the machine not to use global positioning systems, plan a route faster, depend less on errors and save memory. The same principle that the human brain uses is used to find the path. The robot does not build detailed maps of the area, but remembers the main landmarks and the connections between them.

— It's like how a person remembers a new place. We don't memorize every detail, but rather highlight the main landmarks and the connections between them. It is this principle that underlies the technology called PRISM-TopoMap, which makes it a practical solution for autonomous robot navigation in real conditions," Dmitry Yudin, head of the MIPT Laboratory of Intelligent Transport, told Izvestia.

The robot represents the terrain in the form of a graph, a mathematical abstraction of any natural system. For proper orientation, it is important to correctly determine the location of the car on it. However, existing machine learning-based methods often make mistakes when recognizing locations. The new PRISM-TopoMap (Place Recognition and Integrated Scan Matching for Topological Mapping) topological mapping method combines several modern data processing technologies to solve this problem.

First, camera images and lidar data are analyzed using an improved recognition algorithm. This allows the robot to recognize places even when the lighting and angle have changed. Before adding a new location to the map, the robot compares it with what is already known, and then builds a diagram of nodes and connections.

Miscalculation: the inspection of microelectronics was assigned to a robot inspector with AI

How automatic complexes can help improve the quality of functions and the cost of final products

The technology allows robots to build and update the map right as they move. To evaluate the effectiveness of the technology, the scientists tested it on five large rooms in a computer simulation and on a real wheeled robot, and then compared it with other metric and topological methods.

— Our experiments in virtual 3D environments and tests on a real robot have shown that the new method successfully builds accurate and coherent maps, even in the presence of sensor measurement errors. It not only provides full coverage of the space, but also works much faster, cheaper and more efficient than existing analogues," said Alexander Melekhin, an engineer at the intelligent transport laboratory.

In the future, the developers plan to teach the system to understand the type and purpose of rooms: to distinguish between kitchens, corridors, warehouses, as well as to improve the algorithms for laying routes according to the created schemes. This will make robot navigation even more accurate and meaningful in real conditions.

— The maps created by our method allow you to quickly and easily plot routes up to several kilometers. By adding recognition of the types of rooms and objects inside them, we can ensure the fulfillment of various navigation tasks, including automated delivery between buildings," said Kirill Muravyov, a junior researcher at FITZ IU RAS.

The technique will find application where it is not possible to use global positioning systems. For example, indoors or to explore other planets.

When will robots become autonomous

— A very promising project that expands the possibilities of navigation in non-deterministic environments. Despite the fact that the data requires more experiments, even at the early stages of the project, its effectiveness is visible through comparison with analogues. Teaching a system to navigate in space, while using methods that people usually use, is the most effective way to develop, but there are many difficulties along the way that still need to be overcome," said Alice Sotnikova, market expert at NTI Neuronet, Deputy CEO of robot manufacturer Degree of Freedom LLC.

The system can be used in the future not only for ground-based robotics, but also in other applications of mobile robots, she added.

"People will stop performing routine operations and start training robotic assistants"

Andrey Zabelin, CEO of the IEK GROUP, talks about the Fifth Industrial Revolution, universal automation and the rethinking of the human role

Humanity is approaching an era where robots will become an integral part of everyday life, from delivering goods to medical care. However, the key technical problem remains the orientation of robots in rooms without GPS/GLONASS signals, explained Dmitry Vetoshkin, market expert at Neuronet.

— Existing methods such as Wi-Fi markers, sound sensors, or QR codes have limitations. I believe that the developers are offering an innovative method that mimics human perception: the robot identifies familiar objects and determines its position relative to them. This requires the integration of AI and neural networks trained for specific locations.: warehouses, apartments, industrial areas. After training, the system can be scaled to any number of devices, from drones to humanoid androids. The solution is relevant for logistics, medicine and other fields, but requires further development: testing in VR and real conditions," he said.

In real life, the position of objects in rooms or the urban environment is constantly changing, so creating a model capable of capturing static and dynamic elements is necessary for the development of autonomous robots. For example, cleaners or delivery people. Right now, most of their power is spent trying to navigate a changing environment. If they can compare mobile and static items, this will allow these devices to focus on their main tasks and better cope with them, summed up Alexey Karfidov, Head of the Department of Technological Equipment Engineering at NUST MISIS.