CAPABILITY ITERATION NETWORK FOR ROBOT PATH PLANNING

Path planning is an important topic in robotics. Recently, value iteration based deep learning models have achieved good performance such as Value Iteration Network(VIN). However, previous methods suffer from slow convergence and low accuracy on large maps, hence restricted path for agents with complex kinematics legged robots. Therefore, we propose a new method called Capability Network(CIN). CIN utilizes sparse reward maps encodes the capability of agent state-action transition probability, rather than convolution kernel models. Furthermore, two training including end-to-end module alone are proposed, both which speed up greatly. Several experiments various scenarios, 2D, 3D grid world real robots different map sizes conducted. The results demonstrate that has higher accuracy, faster convergence, lower sensitivity to random seed compared VI-based models, more applicable robot planning.