The route has been adjusted: the new algorithm will reduce the probability of accidents in robots and drones by 17 times

Russian scientists have developed a new algorithm that will allow service drones and unmanned vehicles to avoid accidents even in difficult and unforeseen traffic situations. The system splits the route into small intermediate targets, so that drones do not have to slow down to choose the right path. According to experts, first of all, the technology will be in demand when operating autonomous devices in closed territories. In the future, after accumulating enough data for training, it can be adapted for driving on public roads.

An algorithm for accident-free driving

MIPT specialists have created the SG-Safe algorithm, which trains warehouse robots and unmanned vehicles to reach their targets in difficult situations without collisions. Thanks to him, cars reach the goal in 90% of cases with an accident rate of only 3%, which is 17 times less than the best existing approaches. At the same time, they make decisions ten or more times faster, because the system does not need to build a route on the go.

Photo: IZVESTIA/Sergey Lantyukhov

— Using our training method, a loader robot will be able to maneuver between racks, and an unmanned vehicle will be able to carefully drive through narrow courtyards and parking lots. Tests have shown that when the number of obstacles increases, the algorithm does not lose accuracy and still rarely makes mistakes," said Grigory Gorbov, a junior researcher at the Center for Cognitive Modeling at the MIPT Institute of AI, a graduate student at MIPT.

Safety on autopilot: The Ministry of Transport will update the rules for unmanned vehicles

The classification of automation levels will be expanded

When moving in the corridors of the warehouse, the robot must avoid many obstacles. The more complicated the route, the more options he has to explore.

Risky actions can lead to collisions, and if the algorithm is too careful, it can start avoiding mistakes. As a result, the robot will simply stop before reaching the goal. This problem is especially common on long routes with obstacles, multiple turns, and limited maneuvering space.

The MIPT scientists' solution forces the robot to be bolder during training — it breaks down the difficult task of navigation into shorter stages. To do this, the system generates intermediate subgoals and uses them to guide machine learning to the endpoint.

Photo: IZVESTIA/Sergey Lantyukhov

— SG-Safe changes the principle of environmental research. Instead of trying to find a way to a distant goal right away, the robot first learns to reach simple and understandable landmarks. This allows him to gain experience in safe driving without giving up on uncertainty due to the risk of making a mistake," said Alexander Panov, director of the Center for Cognitive Modeling at the MIPT Institute of AI.

The system implements two related strategies. The first one offers intermediate subgoals. It tells you where you should try to get to first, so that you can eventually master the whole path. The second one is responsible for safe behavior and uses these subgoals to guide the exploration of the surrounding space.

Hints from the first strategy are needed only at the training stage. When the training is completed and the robot starts real work, the already honed safe movement program remains in its control system. Therefore, the robot makes decisions almost instantly, bypassing complex calculations. By comparison, algorithms that build a route on the go spend 10 to 30 seconds on each step.

Technology implementation

The effectiveness of the development was tested using computer simulation, including a test simulating the movement of a car through a maze of narrow corridors with sharp turns.

SG-Safe successfully achieved the goal in 90% of cases, while other methods of reinforcement learning often fell into a stupor. Using only sensor data, the algorithm made decisions ten or more times faster than analogs that build a route on the go. The accident rate was only 3%, which is 17 times less than that of one of the best existing approaches.

Photo: Global Look Press/David Balogh

The next step of the developers is to simplify the architecture of the system and explore the possibility of transferring this technology from computer simulations to real robots.

— In my opinion, this development has a high potential for practical application in unmanned vehicles and warehouse robotics. The indicators obtained in the simulations look encouraging, but for an objective assessment it is necessary to confirm them in real conditions — with unstable operation of sensors, unpredictable behavior of people and other road users," said Oleg Kivokurtsev, co—founder of Promobot.

If the technology successfully passes such tests, it can become an important element of a multi-level security system for autonomous devices and help significantly reduce accidents, the specialist added.

They drive themselves: fully autonomous cars will appear on the roads of the Russian Federation

In two years, unmanned vehicles are planned to be released on all highways of the country.

The approach looks more promising for warehouse and intralogistic robotics. "For road unmanned vehicles," Artyom Bogatyrev, market expert at Aeronet Research Institute, is confident, "this is more of a foundation for the future."

— Reinforcement learning algorithms often slip into excessive caution: the robot stops trying risky maneuvers and, along with the risk, loses the ability to reach the target. Splitting a long route into intermediate landmarks allows you to gain experience of safe movement in short steps, and this is exactly what is missing when learning on long paths with turns and narrow passages," he said.

Photo: Global Look Press

So far, 90% of successful races and 3% of accidents have been achieved in the simulation, said Leonid Drobyshevich, NTI technology expert.

— For a warehouse with a well-known layout and low speeds, such indicators are already close to working ones. The requirements for road transport are orders of magnitude stricter, and the transition from a virtual environment to a real road is almost always accompanied by a decline in quality.: Sensors make noise, pedestrians and other road users behave unpredictably, and the scene changes faster than the warehouse. Therefore, I see the nearest application in closed circuits: in warehouses, distribution centers, industrial sites, and port territories," the specialist said.

There, the technology can be validated on a real fleet of vehicles and collect statistics that will be needed for more complex scenarios on public roads, he concluded.