Acquiring Robot Navigation Skill with Knowledge Learned from Demonstration

Abstract

The ability for robots to acquire navigation skill is a hotspot in the field of mobile robot research. In this work, to accelerate the process of learning mapless navigation skill for mobile robots, knowledge from demonstrations are used. What’s more, not only the success rate of navigation, but also the time cost and motion stability are taken into consideration. We propose a mobile robot navigation approach, which first learns the navigation knowledge from demonstrations, and then learns the navigation skill through exploration based on the learned knowledge. The knowledge is learned from demonstrations sampled from existing algorithms, rather than experts, to reduce the complexity of sampling. The reward function is specially designed to guide the robot to move towards the target efficiently and stably. Experimental results show that the proposed method can not only improve the learning efficiency, but also improve the navigation performance.