Sim-to-Real Image Transfer for RL-Based Bin Picking

Robotic bin picking has become increasingly important in the field of automated manufacturing and logistics, where precise handling of objects in complex settings is essential. Reinforcement learning (RL) has emerged as a powerful tool for developing policies to manage such tasks. However, a significant challenge persists in transferring these policies from simulation to real-world environments due to visual discrepancies, often resulting in reduced performance when applied to actual robotic systems. To overcome this issue, we introduce an innovative approach that incorporates spatial transformer networks (STN) into the RL framework. This technique specifically addresses the scale variations between simulated and real-world images, enabling the policy to adapt more effectively and enhancing both the robustness and generalizability of the learned behaviors. Our approach has shown a marked improvement in real-world bin picking tasks, with the integration of STN leading to an 11.5% boost in grasping accuracy, achieving a high overall success rate of 95.5%. By effectively narrowing the visual domain gap, our method reduces the reliance on extensive real-world data collection and training, facilitating quicker and more efficient deployment of robotic bin picking solutions in a variety of industrial contexts.