Leveraging reinforcement and curriculum learning for flexible robot-based snap-fit assembly automation

Abstract This work proposes a novel approach combining reinforcement learning (RL) and curriculum learning for flexible robot-based snap-fit assembly. The proposed system entails a simulation environment to train RL models for snap-fit connections, adapting part geometries and friction coefficients during training. The results show significant improvements in assembly success rates over the training stages, reaching over 95 % after iterative adjustments. Future work will focus on transferring the trained models to real robots and addressing the Sim2Real-Gap.