Semantic Mapping and Autonomous Navigation for Agile Production System

Abstract

A typical task for mobile robots in production logistics is to transport objects from one location to another. This requires the robots not only to locate the objects, but also to design a collision-free transport path. Currently, many mobile robots operate in an occupancy map, require a predefined goal pose as a destination, and lack the availability of high-level navigation. With the aid of RGB-D cameras, semantic objects can be detected and added to the map, providing more opportunities for scene understanding and flexible navigation. In this paper, we extend the current 2D mapping and navigation framework with object segmentation and fine position navigation to achieve better performances in task-level navigation. First, we propose a framework for creating and maintaining a hypermap by recognizing semantic objects in the environment and integrating them into an existing 2D occupancy map. Second, we present a coarse-to-fine navigation strategy on this hypermap. The coarse navigation receives object information and designs a global path towards the destination, while the fine navigation utilizes the local information to ensure precise docking to the workstation. A field experiment demonstrates that the proposed system can achieve high performance in a production logistics environment.