Robotic Grasp Control Policy with Target Pre-detection Based on Deep Q-learning

Abstract

We present an accurate, real-time approach combined target pre-detection with robotic grasp based on deep Q-learning. Skilled robotic manipulation benefits from learning approach between target pre-detection and robotic grasp actions: target pre-detection recognizes the object and finds a good grasp rectangle in a single step; meanwhile, grasping can help displace objects to make target detection more accurate and disturbed-free. During grasping experiments in simulation scenarios, our approach rapidly learn complex actions amid challenging cases of clutter, especially achieves better grasping success rates and performs significantly better.