Knowledge Induced Deep Q-Network for Robot Push and Grasp Manipulation Skills Learning

In the process of robotic push and grasp manipulation skills learning, the prior knowledge of object shape and environmental constraints can be employed to make corresponding manipulation policies. In this paper, a DQN decision model is proposed for learning multi-step push and grasp manipulation of different types of objects. The key of the proposed method is that the prior knowledge of object shape and environmental constraints is introduced into the reward system by a SVM classification model. Therefore, the reward system will continuously motivate the robot to approach to the target manipulation during training process. Simulated and actual verification of the trained model is carried out. The experimental results show that the trained model has a high success rate in grasping cube objects, and it can also successfully complete the task of pushing cylindrical objects to the constraint boundary.