Automatic Generation of Feedback Stabilizable State Space for Non-holonomic Mobile Robots

Learning approaches to robot control problems generally require a lot of trials, which is crucial to make the approaches available in wider applications. As a means to improve learning efficiency, it is promising to introduce methodologies and ideas developed in the control theory. Aiming at introducing an idea of nonlinear control method to learning approach, this paper presents an acquisition of a state space that allows control to reach a target for the two-wheeled mobile robot with non-holonomic constraints. In the proposed framework, it is assumed that knowledge of the sensor is not available in advance. An adaptive grid distribution algorithm to cope with a non-holonomic controller scheme is proposed. It was experimentally confirmed that the robot could reach the target point stably by the proposed method. The proposed method presents an idea to effectively integrate machine learning and control theory and it has the potential to become a unified learning method that can be applied to various control targets with fewer samples or trials.