Collaborative Learning of Deep Reinforcement Pushing and Grasping based on Coordinate Attention in Clutter

The ability of robots to perceive position information of objects is important for pushing and grasping tasks in clutter. We propose a Coordinate Attention Grasping and Pushing Network (Ca-GPNet) to learn synergistic grasping and pushing strategies in cluttered environments. We use both fully convolutional networks to predict actions: One predicts the grasping direction and position of the gripper, and the other predicts the initial push position and direction of the gripper. We propose to use the coordinate attention module to capture position information of objects along the horizontal and vertical directions of space and aggregate the features. The attention module extracts long-range interdependencies in one dimension while keeping the information about the location of objects in the another dimension. Then a pair of attention maps that perceive the location and orientation features of objects are generated to reinforce the perception of objects' locations by the network. The grasping success rate of the system is 73.7% and 71.4% of the action efficiency in the simulation. Our approach can be applied to real-world pushing and grasping tasks. Ca-GPNet grasping success rate is 79.2% and action efficiency is 75.3 % in the real world. Compared with baseline methods, our system can rapidly learn the pushing and grasping cooperative policy in a cluttered environment, with higher sample efficiency, grasping success rate and action efficiency.