Attention-Based Policy Network for Sensor-Free Robotic Arm Control With Deep Reinforcement Learning

This paper proposes a novel attention-based convolutional neural network (CNN) for sensor-free robotic arm control, aiming to improve six dimensional (6D) pose estimation and end-effector operation in an end-to-end manner. Unlike traditional methods that rely on explicit feature engineering or sensor feedback, our approach leverages a sophisticated attention mechanism within the convolutional backbone to enhance spatial awareness. The proposed localization sub-module scores each prior regime through a weighted average of activation maps, allowing the network to focus on the most informative regions of the input. Additionally, we introduce a two-phase training methodology requiring only image-level annotations. In the first phase, the network learns to extract discriminative features from synthetic images, which are crucial for accurate 6D pose prediction. In the second phase, a reinforcement learning agent, equipped with the trained vision model as its sensory module, is optimized using a sparse reward function to refine action policies. Experimental evaluations in two virtual scenarios demonstrate that our method outperforms popular CNN-based approaches in terms of both accuracy and efficiency. Specifically, our method improves task success rates by 52.9% and reduces position error by 72.3% compared to baseline models, showcasing its effectiveness in sensor-free robotic arm control.