Mobile robot navigation using evolving neural controller in unstructured environments

Autonomous mobile robot navigation using only local sensory data in an unstructured environment poses a challenge for mobile robots. Our approach aims to evolve a single neural controller that simultaneously learns obstacle avoidance and target seeking without an explicit behavior switching scheme unlike some other recent approaches. We train the robot in a static, unmapped environment comprising of multiple obstacles of different shapes and sizes. We study 4 different training schemes that emphasize the learning objectives differently. The trained controller is validated by simulations across a diverse range of environments different from the training environment. A comparative study of the robot behavior under different objective functions has been done. We discuss the performance measured on the basis of several metrics.