Discrete Motion Planner based on Deep Recurrent Neural Network for Mobile Robot Obstacle Avoidance in Dead-End Environments

Abstract

For autonomous mobile robots, obstacle avoidance is a necessary capability for navigation. We have thus far proposed a discrete motion planner based on a fully-connected feedforward deep neural network. The motion planner resulted in a sufficient capability with discrete motions, such as straight, right turn, and left turn. Once a robot stops, however, it is impossible to restart navigation toward a goal destination. This problem occurs when obstacles form a dead end in an environment. In order for the robot to escape from the dead end, further motions are required. Moreover, the robot is required to take into account a series of motions for the dead end. In this paper, therefore, we propose a discrete motion planner based on a fully-connected deep recurrent neural network. For network training, we use imitation learning from motion instructions. In the experiment, two of the motion planners are applied to the robot. Finally, we show that the robot based on the proposed motion planner successfully moves toward the goal destination while avoiding obstacles even in a dead-end environment.