Object-Oriented Navigation with a Multi-layer Semantic Map

Abstract

In this paper, a novel architecture is proposed for object-oriented navigation of a mobile robot. The idea is to autonomous navigate and docking with respect to a given object in an indoor environment. For this purpose, the navigation scheme is designed based on a novel multi-layer semantic map. The map includes an occupancy layer for path planning, a localization layer for control and an object pose estimation layer for docking control. The two-dimensional (2D) occupancy map is generated from a 3D RTAB-Map method. The multi-layer semantic map integrates the outputs of a deep-learning model of real-time 3D object pose estimation for robot localization. Experimental results show that the proposed navigation system can plan a collision-free path for the mobile robot to reach and dock to the target object. The LiDAR-based navigation control runs at 15 Hz and the object-based localization is updated at 4 Hz on the Jetson TX2 embedded controller. In the current implementation, the mobile robot can dock to an object with an accuracy of 8 cm in position and 5 degrees in orientation. The navigation design has the potential to be applied for daily life tasks of mobile robots in an indoor environment.