Supporting robot application development using a distributed learning approach

Robot application developers program industrial robots during the commissioning or re-configuration of a production line. Not only do they require precise locations of work pieces and respective movements but also need to optimize program parameters manually. This often results in high engineering effort and stress. This paper introduces an approach to support robot application developers using distributed learning and machine learning to optimize robot application parameters in simulation. Contrary to end-to-end robot learning approaches which are often infeasible to adopt in practice, this approach leverages domain knowledge of application developers and their familiarity with writing robot programs. The developer is only additionally required to specify bounds for parameters to be optimized and an evaluation criterion. A central machine triggers as many worker machines as required to run simulations in parallel with different parameters to learn the best policy. Our approach works independent of the robot task and robot simulation software and can be configured with different algorithms. To validate the approach, it is applied to a robotic insertion task in an assembly scenario using a genetic algorithm and ABB RobotStudio.