Learning Individual Features to Decompose State Space for Robotic Skill Learning

Abstract

Due to suffering from the diversity and complexity of robotic tasks in continuous domains, robotic skill learning is the most challenging issue in this area, especially for robots with high-dimensional state spaces. To learn structured policies for continuous control, the graph neural networks (GNN) was previously applied to incorporate explicitly the robot structure into the policy network. In this work, we tackle the problem of robotic skill learning in high-dimensional state space with the help of graph neural networks. Instead of utilizing a general purpose multi-layer perceptron (MLP) as a unified controller to output actions for all joints of the robot, we construct a separate controller for each joint of the robot by using the individual features that have been extracted by GNN model. Empirical results on simulated continuous systems, including applications to PR2 task and Centipede task, demonstrate that the proposed framework can achieve satisfactory learning performance, and more importantly, it significantly reduces the parameters of the policy network.