Manipulators with Machine Learning-based Control with Reinforcement

摘要

This industry is created for a number of tasks in which there is no unambiguous solution algorithm. Accordingly, these tasks cannot be solved in a fully automated way. Such tasks include welding of steel with a fuzzy weld, cutting of the forest, sorting, packaging, conveyor tasks with variable environment, all these tasks are currently solved by a person. As a result of using algorithms, it is possible to achieve results up to 70% savings on debugging smart manipulators and on automation tasks for the entire industry 4.0. The size of this market, i.e. the market of robot manipulators, is approximately 35 billion dollars in global terms, and in particular in Russia, with a low level of automation, 15 billion rubles. The target audience is a part of the production facilities that implement automation, robotics and, accordingly, they are engaged in the processing of parts. The business model by which we are developing is the creation of a hardware complex to upgrade customers’ equipment. The DDPG algorithm is used for training. The training itself is carried out in two versions -using HER (Hindsight Experience Replay) and without. The results are compared and the best model is taken for work.