Deep Reinforcement Learning with Convolutional Networks Applied to Autonomous Navigation of Real Robots Using Virtual Scenario Training

摘要

Industrial applications of autonomous mobile robotic systems have demanded the development of autonomous navigation algorithms using artificial intelligence. This paper proposes a method based on deep reinforcement learning algorithm and convolutional network (DQN-CNN: Deep Q-Network with Convolutional Network) to allow autonomous navigation of robots equipped with RGB cameras. The training phase of this algorithm is performed in the Coppelia VREP simulation environment with real practical implementation based on the NVIDIA Jetson Nano development board. The proposed methodology includes training in a simulation environment and integration with a real system. For prototyping, a navigation track is used, where the robot is controlled by a DQN reinforcement learning algorithm. The results confirm the effectiveness of the proposal, as the real robot was able to successfully perform the planned mission. It is expected that this work can contribute to the advancement of mobile robotics and reinforcement learning, as well as provide a useful methodology for researchers and developers working with these technologies. The video of the practical experiment is available at: https://bit.ly/3IK1Rb5.