Bending obstacles when moving a mobile robot

摘要

The issues of modeling when navigating around obstacles of a mobile robot using machine learning methods are considered: Q-learning, SARSA algorithm, deep Q-learning and double deep Q-learning. The developed software includes the Mobile Robotics Simulation Toolbox, Reinforcement Learning Toolbox, and the Gazebo visualization package for environment simulation. The results of the computational experiment show that for a simulated environment with a size of 17 by 17 cells and an obstacle 12 cells long, training using the SARSA algorithm occurs with better performance than for the others. An algorithm for avoiding obstacles without the use of machine learning is proposed, and it was shown that the speed of avoiding obstacles using this algorithm is higher than the learning speed using deep Q-learning and double deep Q-learning, but lower than using the SARSA and Q-learning algorithms. . For the proposed algorithm, a numerical experiment was carried out using the robot movement simulation environment in Gazebo 11 and it was shown that cubic obstacles are being avoided faster than cylindrical ones.