Local Motion Planning for Ground Mobile Robots via Deep Imitation Learning

Abstract

Imitation learning have been recently applied to a number of robotic-related tasks. In this work, a novel approach based on imitation learning will be applied to the local motion planning problem for mobile robots. Based solely on a monocular RGB camera and set of expert demonstrations, our learned model can predict an accurate steering angle actions to be performed by the local motion planner of a mobile robot. In order to train and validate our model, we collected large amount of labelled RGB images from a monocular camera mounted on a ground mobile robot platform while driving the robot manually by human demonstrator. The proposed approach has achieved resilient results in capturing the behavior of a human demonstrator and provided a smooth and safe steering trajectories. Quantitatively, the model has also achieved lower mean absolute error of on a step-by-step basis of steering angular velocity with only 0.1 radians/s error over all the testing dataset.