Research on Perception and Control Technology for Dexterous Robot Operation

Robotic grasping in cluttered environments is a fundamental and challenging task robotics research. The ability to autonomously grasp objects scenes crucial for robots perform complex tasks real-world scenarios. Conventional based on the known object model structured environment, but adaptability of unknown complicated situations constrained. In this paper, we present robotic architecture attention-based deep reinforcement learning. To prevent loss local information, prominent characteristics input images are automatically extracted using full convolutional network. contrast previous model-based data-driven methods, reward remodeled an effort address sparse rewards. experimental results show that our method can double learning speed series randomly placed objects. real-word experiments, success rate robot platform reaches 90.4%, which outperforms several baselines.