Sberbank and Microsoft Research have developed an AI robot control system. It allows robots to be trained to interact with the physical world.
The key goal of the research project was to develop a solution that can interact with people, freeing them from mechanically complex routine work and providing a high level of security. The system can be applied in various fields - from logistics to rescue operations.
Work on the project was carried out at the Robotics Laboratory of Sberbank in Moscow and at Microsoft in Berkeley and Redmond in the USA and lasted from May 2019 to May 2020. The developers had to create a technology for unloading cash collector bags with coins weighing up to 6 kg using a robotic arm from mobile carts for their subsequent processing in calculating machines. Real-world tests of the system have shown that it can operate with an accuracy of 95%.
Work on the project took place in three stages. The first stage involved the development of a simulation of a robotic installation and deformable bags of coins, taking into account their physical characteristics. The second stage involved integrating simulators with a machine learning framework and conducting experiments in a simulation environment. At the final third stage, the intelligent agent trained in simulators was transferred directly to a physical robot.
Original (in Russian)