Attention-Enhanced Multimodal Robotic Arm Lift Reinforcement Learning

Abstract

Reinforcement learning (RL) has made significant progress in end-to-end grasping tasks, but existing methods mostly rely on single-modal perception or limited feature modeling, which limits the robustness and generalization ability of the strategies. To this end, this paper proposes an attention enhanced multimodal reinforcement learning framework. Based on the Isaacsim simulation environment and the rsl_rl framework of Isaaclab, it studies the completion of the lift task by the UR5 robotic arm equipped with the Robotiq 85 griper. This framework integrates the robot's body perception state and visual observation into a unified policy network. And embed multiple lightweight attention mechanisms in the visual network to enhance feature representation. A single-object grasping experiment with random initial postures was designed. The Mean reward, Pose position value, and Reward lifting were adopted as core indicators to comparatively analyze the performance of four types of attention mechanisms: channel, spatial, hybrid, and multi-branch. The experimental results show that the attention enhancement model significantly improves the training efficiency and grasping stability of the grasping task. Among them, Global Attention Mechanism (GAM) and Split-Attention Networks (ResNest) performed the best, while Efficient Channel Attention (ECA) could also stably outperform the Convolutional Neural Network (CNN) baseline at a lower computational cost. This study comprehensively compared the attention mechanisms in multimodal grasping, confirmed their effectiveness in optimizing reinforcement learning operation strategies, and provided new ideas for visual feature modeling of basic operation tasks of robotic arms.