Goal-Directed Planning by Predictive-Coding based Variational Recurrent Neural Network from Small Training Samples

The existing goal-directed planning methods using predictive-coding-based variational recurrent neural network (PV-RNN) need many training samples. In this study, we propose a goal-directed planning method by PV-RNN from small training samples. To achieve generalization from small training samples, the proposed PV-RNN integrates the following two ideas: (1) To have internal models based on kinematics or dynamics (2) To have diversity in small training samples. We evaluated our proposed PV-RNN model for its ability to generate sufficiently generalized goal-directed plans from small training samples using an 8-DOF robot arm in a simulator environment. Furthermore, we evaluated the advantages and disadvantages of the proposed PV-RNN compared to general reinforcement learning and imitation learning using a benchmark environment for a robot arm task.