Research on the autonomous mobile robot has been done steadily. Recently, robots employ various locomotion and sensors [Khazanov14]. These researches assign robots by task to perform a specific task [Amigoni05], or use semantic information to service in a limited space indoors [Lim10]. In addition, there are studies that use task management, environmental awareness, trajectory planning, decision making and terrain classification using semantic maps and ontology for robot mapping [Liu12], [Li12]. However, these robots mainly carry out autonomous study on predefined areas in a way limited to locomotion and sensor system. In this paper, we propose a novel autonomous navigation method that can travel a wide area through a map suitable for semantic DB-based robot motion and sensor systems developed for intelligent robots. The proposed method simulates Figure 1: Semantic DB based Reliable Autonomous Navigation the navigation method used by human beings and drives similarly. Robots travel and recognize objects around them as humans do. To do this, the robot registers and recognizes the objects necessary for navigation, as shown in 1. After that, if the recognized objects are obstacles, they will be avoided or waited after depending on the properties of the motion defined by the semantic information. If the recognized object is an object that affects driving (e.g. a rough road), the driving speed is adjusted according to the drivability defined in the object's properties.
Meanwhile, a wide-area spatial map based on 3D LiDAR sensor and vision sensor is constructed to drive wide-area space suitable for various robot motions and sensors. Robots can use this map to recognize the exact location from the multi-sensor data or even a single sensor. Therefore, the proposed driving method is useful in the wide area where many people walk because they drive according to semantic information. This is expected to bring dramatic developments to the autonomous driving of robots.
The semantic information used for autonomous driving of robots is based on objects. The robot recognizes the object while driving and uses semantic information of the object to secure driving ability similar to humans. This can be divided into two abilities: First, the motion property is used as semantic information of the object. Using this, an object, such as a person or a chair occupied by a person, is recognized as a movable object and waits for 5 seconds when the robot meets a moving obstacle while driving. After that, it is avoided in the same manner as a fixed obstacle, as shown in 2. drive. The degree of roughness is divided into 5 levels and it is possible to secure the driving stability of the robot. 3 shows the difference in running speed on gravel and asphalt roads. The semantic DB is created by modeling object information necessary for driving from the viewpoint of the robot. Based on this, a map that is suitable for the locomotion and sensor system of the robot is build. This is a key factor for driving the robot in a wide space. 4 and 5show the entire flow chart for creating the map and the robot platform used for building the map. The platform was used to build a map based on the point cloud of the 3D LiDAR sensor and the ORB features of the vision sensor [Mur17]. 6 shows wide-area spatial maps for each sensor. It is possible to generate a semantic map including semantic information in a map suitable for the robot sensor. Based on this semantic map, the robot performs autonomous Navigation. The robot sets up the driving strategy according to the properties of the recognized object (e.g. movable, rough etc.). Using this method, the autonomous mobile robot can flexibly move in the wide area.
Using the proposed method, autonomous driving was carried out in a wide area of about 6ooo m2, as shown in 7.
This was tested during the exhibition at the convention center. First, the robot received information about the semantic map and mission (patrol) from the local server and started driving. Next, the robot traveled to pass between the people standing in line for entry and waited a while when it could no longer drive. The robot then continued to drive through the gaps caused by people's movement. As shown in the results, it was possible to drive smoothly in an environment with dense crowds and to travel between crowds using semantic information.
In this paper, we propose autonomous navigation method in wide area by generating semantic map suitable for locomotion and sensor system of various robots based on the driving method performed by humans. For this purpose, driving using multi-sensor based semantic navigation map and defined semantic information was performed. In the future, we will apply semantic information-based autonomous driving methods that can be used for more types of locomotion and sensors.
Acknowledgment This work was supported by the Korean Evaluation Institute of Industrial Technology and conducted by the Ministry of Industry and Commerce in 2017 (Industrial Core Technology Development Project, Project Number 10080489) and 2018 (Industrial Core Technology Development Project, Project Number 20000683).