This paper discusses the development of the multi-functional indoor service robot PSR (Public Service Robots) systems. We have built three versions of PSR systems, which are the mobile manipulator PSR-1 and PSR-2, and the guide robot Jinny. The PSR robots successfully accomplished four target service tasks including a delivery, a patrol, a guide, and a floor cleaning task. These applications were defined from our investigation on service requirements of various indoor public environments. This paper shows how mobile-manipulator typed service robots were developed towards intelligent agents in a real environment. We identified system integration, multi-functionality, and autonomy considering environmental uncertainties as key research issues. Our research focused on solving these issues, and the solutions can be considered as the distinct features of our systems. Several key technologies were developed to satisfy technological consistency through the proposed integration scheme.

Fig.1 shows the hardware configurations of three PSR systems. The PSR-2 is an upgraded version of the PSR-1, so they have similar configurations. The Jinny is a specialized version aiming at a commercial guide robot with a human friendly appearance.

Figure 1. Hardware configurations of three PSR platforms. (a) PSR-1 (b) PSR-2 (C) The Jinny.

Fig.2 shows an example of navigation in a conventional office building. The start point is node0, and goal is node2 as shown in Fig.2.(a). Fig.2.(b) presents a planned path and an actual trajectory during the navigation. The robot's actual trajectory has the discontinuity which mainly results from position updates by the localizer. The robot can move smoothly since the behavior also contains several schemes for stable tracking such as an acceleration filter. Fig.2.(c) shows the results of localization. It represents the local map, laser scan data, reference data, sample distributions, and estimated position. The data for path planning and localization is gathered simultaneously with a single experiment. The reference measurements of an estimated robot location are mostly consistent with the scanned measurements of an actual robot position. It means that the accurate position estimation is accomplished. Although the environment is slightly changed and a user disturbs the Jinny's way, the proposed localization algorithms work successfully.

(a) An experimental environment	(b) A path planning example	(c) A localization example
Figure 2. Navigation examples.

Four target tasks were successfully implemented to three PSR platforms.

A. A delivery task
Initially, a user commands a task by giving initial and destination room numbers through the remote computer or the interface device on the PSR. Then, the PSR navigates to the room where the target box is. When it reaches in front of the room, it releases the trailer and enters the room alone. After the PSR picks up the box, the PSR leaves the room and places the box on the trailer. Next, the PSR docks the trailer again and moves to the target room to place the box in the pre-determined position. The last step is to return to the standby state with trailers.

B. A guide task
The guide task was implemented into the Jinny platform since a guide task inherently requires a human friendly appearance. The guide robot Jinny is developed toward installation in the National Science Museum of Korea. The Jinny autonomously navigated the crowded environment and explained exhibits to visitors. A user can select the scenario by selecting the sequence of exhibits to be explained. The Jinny also provided several interesting services including a simple game with visitors, and dance to the music, and following visitors using laser range data.

C. A patrol task
The patrol task can be achieved without extra effort since only navigation capability is required. The navigation behaviors developed for other tasks are directly reused. For omni-directional surveillance, a pan tilt camera and four web cameras are set up on the top of the PSR-2 as shown in Fig.3. The robot can broadcast camera images to the monitoring station.

D. A cleaning task
The scenario of a floor cleaning is as follows. The robot is initially loads the grid map of a target workspace, and then, divides it into several sections if it is a large open space. It is more efficient to divide the space into several sections than to cleaning it at one process since the former is less affected by uncertainties of environments like localization errors. Each section is swept by the full-coverage cleaning algorithm based on the wall following technique. The robot repeats these steps until the assigned workspace is fully covered. Then, the robot moves to the next section to be cleaned. These processes iterate alternately until the whole workspace is covered.

Figure 3. Four target tasks. (a) Delivery (b) Guide (C) Patrol (4)Cleaning

Publication