Action Selection Based on Prediction for Robot Planning

In this work we focus on the action selection process of a robot by equipping the robot with the ability of internal prediction. A novel approach with internal simulation is proposed, in which Conditional Generative Adversarial Nets (CGANs) provides the possibility of action selection and allows the robot to choose an optimal action based on the prediction. This leads to robots that can perform tasks better. In addition, a structure containing recurrent neural network (RNN) is used to further predict the sequence of actions for robot planning. A key feature of this model is the incorporation of sensorimotor prediction, where the robot generates corresponding actions based on the current context and anticipates the sensory consequences of currently executable actions in internal simulation. Experiments have been conducted on PKU-HR6.0 to verify the effectiveness of our approach, showing that it improves the accuracy and speed of robot arm reaching.