Reinforcement learning for hierarchical and modular neural network in autonomous robot navigation

This work describes an autonomous navigation system based on a modular neural network. The environment is unknown and initially the system does not have ability to balance two innate behaviors: target seeking and obstacle avoidance. As the robot experiences some collisions, the system improves its navigation strategy and efficiently guides the robot to targets. A reinforcement learning mechanism adjusts parameters of the neural networks at target capture and collision moments. Simulation experiments show performance comparisons. Only the proposed system reaches targets if the environment presents a high risk (dangerous) configuration (targets are very close to obstacles).