IJCAS

With the increasing complexity of the robot grasping environment, it puts forward higher requirements on the grasping strategy of manipulators. However, grasping cluttered multi-class objects is a challenging task because the objects are stacked and occluded from each other, and it is often difficult for robots to find a suitable grasping position. Deep reinforcement learning with DQN algorithm has been used to study the pushing and grasping strategy to manipulate cluttered multi-class objects. However, there exists long learning time and low success rate. To solve this problem, we adopt two fully convolutional networks (FCN) to map the color and depth maps to actions: pushing and grasping. These two networks are trained by an improved soft actor-critic algorithm that includes auto entropy regularization, regularized objective function and clipped double Q learning. Pushing and grasping synergies has been learned with the dense reward feedback. The simulation experiment demonstrate that the learning process converges quickly and stable with a grasp success rate up to 83.3%. We further demonstrate the generalization ability and well performance of our models when novel objects appear in the scenes that the robot has never grasped before. Finally, real-world experiments with trained models from simulations are conducted to test the grasping performance on manually arranged scenes.

Keywords: Autonomous grasping, deep reinforcement learning, robot, soft actor-critic.