How reinforcement learning can reduce programming costs
Large amounts of data are required to optimize robot movements, reduce programming costs and open up new automation applications. In addition, generating this data with robots is very costly and time-consuming, as several robots have to work in parallel. However, in order to save costs, the data can also be collected in a simulation environment. The only drawback is that the action strategies learned in the simulations cannot be easily transferred to real robots. This effect is known as the "reality gap".
In the "TransLearn" project, the model errors are identified so that the simulative results can be transferred to real robots. The project aims to significantly reduce the programming costs of industrial robots. On the one hand, programming can be carried out better and faster in simulation, on the other hand, the robots can learn parameters independently in simulation and in the real production plant and optimize task execution. Whereas programming experts have been required for programming robots up to now, they will in future be able to be instructed by process experts. Shorter cycle times and lower energy consumption could also be targets for optimization using reinforcement learning. This not only increases productivity, but also leads to cost savings and greater sustainability in production.