The 3rd International Conference on Ubiquitous Robots and Ambient Intelligence (URAI 2006)

Preliminary Study on the Environment Map for Mobile Robots Seunghwan Park,

Yucheol Lee, Moonhoen Lee, Sangik Na, Wonpil Yu Intelligent Robot Research Division Electronics and Telecommunication Research Institute 161, Gajeong-dong, Yuseong-gu, Daejeon, 305-700, KOREA {sinkyv, yclee, mhlee, nsi, ywp}@etri.re.kr

Abstract - Generally, a robot needs the spatial data or map to offer services to a human. As a human want to receive more intelligent services from a robot, it is required that the concept of the environment map should be extended with considering the properties of a human and a robot. In this paper, the environment map is subdivided according to the subjects and the related map categories are proposed. Function Map, which is used for the motion of a robot, and Interface Map, which is used for the convenience of a human, are defined firstly. For the connection of these two maps, Interaction Map is also designated. Each ‘Space,’ which is defined as the minimum unit of the spatial environment, has a map set composed by these three kinds of map. This set is treated as ‘0-layer’ maps. Another map is needed for the management of the 0-layer maps and for the task accomplishment among ‘Spaces.’ Network Map is introduced for this purpose and it is treated as the ‘1-Layer’ map. The organized operation of the robot services can be achieved using this hierarchical map system. Keywords - Environment Map, Mobile Human-Robot Interface, Hierarchical Map

Robot,

1. Introduction As the importance of robot mobility is increased, more accurate environment maps are required. When an environment map is constructed, US (ultrasonic) sensors are used frequently [1-3]. US sensor has the benefits that it is relatively cheap and easy to use. But it also has some defects such as poor directionality, frequent misreading, and specular reflections [4-6]. Thus, many methods are used to make up with the weak points when sensor data should be interpreted. Laser range finder is used for the construction of an environment map if accurate data are needed [7-10]. Laser sensor can offer very precise data and even 3-dimensional data can be acquired by this sensor. But its price is relatively high, thus laser sensor is not used as frequently as other sensors like US. If laser sensor is made cheap with the sacrifice of its function except the range detection, its usage may be extended. For example, there is a research that laser finder is used only for following a guide and map building is accomplished with US sensors [11]. An environment map is constructed with sensor measurements. Generally, sensor data have some spurious

values. Thus, various methods are researched and applied to data to acquire more accurate results. The probability-based methods are most frequently used for this purpose [9, 12, 13]. Especially, there are many cases based on the Bayesian Rule when the positions and/or number of obstacles should be known. In addition, Fuzzy Logic [14] or Neural Network [3] is employed to improve the accuracy of data or to make the global map which is composed by the local maps. On the other side, these various methods are examined in the same environment to compare their merits and demerits [2]. There are some methods in which landmarks are used for the localization of robots or map building [10, 15, 16]. A landmark is a kind of reference by which a robot can recognize the environment. It can be categorized as an artificial landmark and a natural landmark. While artificial landmarks are made artificially and locate on specified points, distinctive feature points, which originally exist in the environment, are assigned as natural landmarks. Relatively accurate data can be acquired with low-priced sensors if we use these landmarks. Imperfect sensor data are refined by the methods mentioned above. An environment Map is now made using these refined data. There are two representative methods for constructing an environment map. One is the Occupancy Grid Map method and the other is the Topology method. The Grid Map method is easy to use and update map data. But its data size is relatively big, thus there is a large computational load in the dynamic environment. To overcome this defect, there is a research about the Hybrid Map [16]. In this research, a static environment map and a dynamic environment map are made separately. And these maps are combined to construct a whole map. If some dynamic changes happen, only the dynamic map is updated. Thus, the calculation load can be reduced. Feature points in the environment are assigned as nodes or lines and other regions are abbreviated in the Topology method. Thus, the environment map becomes simple and computational load is small in this method. Also, robot motion can be made easily. On the other side, detail recognition of the environment is difficult and the optimal path planning is not guaranteed in the Topology method. The Topology method is used in various ways. For example, there is a research that Topology-based Map is extracted from Grid data [17]. Additionally, the geometric model of the environment or the Metric map is combined

107

The 3rd International Conference on Ubiquitous Robots and Ambient Intelligence (URAI 2006)

with the Topology Map to construct the Hybrid Map [18, 19]. The studies mentioned above are somewhat limited to the environment which a robot recognizes. But the services which a human wants to receive are more and more subdivided and required to be intelligent as technologies are developed. Additionally, it is predicted that the cases, which a human participates in these service tasks and interacts with robots, are increased [20]. Thus, it is needed to subdivide robot tasks and the environment effectively and to grant meanings to them. In other words, the task which a robot can do and the data which a robot can interpret should be divided from the task and data which a human can process. Additionally, the information device which connects these two categories is also needed. Up to the present, the researches of this viewpoint are not progressed deeply. Only the studies, which are about the effective acquisition of environment data [21, 22], and the method, which gives some meanings to the environment regions using the prescribed information of objects [23], are proposed. Thus, needed is the separate research about the information offer method that a robot can recognize and the information offer method which is familiar with a human. For this purpose, we want to subdivide and categorize the environment map on the subject of usage, in this paper. The presentation of this paper proceeds as follows. In section 2, subdivision method of the environment map on the subject of usage is presented. The proposal is discussed in section 3, which explains what items are required to implement these subdivided maps effectively and what formats are needed to satisfy these items. Finally, section 4 contains brief concluding comments and some future works.

2. Environment Map Subdivision When a mobile robot, which provides services such as security or fire detection, executes its tasks, the data like images and/or sensor measurements captured by the robot are inputted to the user’s terminal through the transition server after some processing. If a user inputs some commands into the terminal, these data are transmitted to a robot through the server and the robot accomplishes the assigned tasks. In these cases, the server generally corresponds to the transition computer which has large capacity. But, a personal computer or a robot-imbedded computer may have the function of server according to the situation. When a robot performs its tasks, it is influenced by its circumstances. Thus, the map is needed, which includes accurate data about the circumstances. The most important reason that a map is needed for a mobile robot is the offer of the environment information for stable and self-controlled motions. We define ‘Function Map’ as the map for this purpose. Function Map offers the information of position of the surroundings to a robot. With Function Map, a robot can estimate its own real position, build the optimal path, follow the path, and avoid collisions with obstacles. If a user wants to control a robot and/or to observe the state of a robot, a means is needed. This means should

Fig. 1. Relations of the maps include the region information as well as the command set for a robot. Thus, we need the map which represents the positions and states of a robot or its surroundings in various terminals. This map exists for user’s convenience. We define ‘Interface Map’ as the map for this purpose. Intelligent service means not only to execute the express commands excellently but to do adequate actions according to the circumstances and situations. Namely, the intelligent service should include a kind of common sense that human beings generally have. Through this property, a human and a robot can interact with each other. We define ‘Interaction Map’ as the map which contains this interaction and the information of the environment which needs the interaction. This Interaction Map has a part of the properties of Function Map and Interface Map, and exists as an independent map which connects two previous maps. In Fig. 1, these relations of the maps are shown. Generally, the maps mentioned above save information of each space which has pre-defined size. Thus, another map is required, which manages the map information about the relations among spaces and includes the information for the application of maps. We define ‘Network Map’ as the map for this purpose. Network Map is the map which connects and manages several spaces, thus it is made with the maximum scope of the environment which robots act.

3. Subdivided Map Format 3.1 Representation of the geometrical data Generally, data can be simply considered as the collection of information materials. The kinds and collection types of data vary according to the gathering intention. Among these varieties, space data are defined as the data which are used to represent the spatial relations among surroundings and/or to model the surroundings [24]. The space data can be divided into two categories, figure data and attribute data. A map is generally considered as the figure data which represent the properties appeared in a geometric space. According to the model of figure data, the basic approach method for a map can be divided to the vector model and the raster model. The model classification is decided by the representation type of the surroundings. Additionally, the object of data collection, save, and interpretation also affects the selection of data model. Real world is categorized as the regions enclosed by points and lines in the vector model. In contrast, grids or pixels are used to represent surroundings in the raster

108

The 3rd International Conference on Ubiquitous Robots and Ambient Intelligence (URAI 2006)

Vector

Raster

Point Line Polygon

Fig. 2. Vector Model and Raster Model model. In the vector model, each object is considered to have a different property. Thus, the information about regional boundary is important to represent the spatial data. SVG, DXF, and the Feature Map are examples of the vector model. While SVG and DXF model have standardized formats, the Feature Map does not have a defined format. This is due to the fact that the Feature Map is almost used in the research fields. Space is divided with grids or pixels of same size in the raster model. Each grid has a value. According to this value, the existence of an object is decided. Because these grids have equal size and shape, their boundaries are not considered. The rules of determining grid values are different to the model types. The graphic file formats, which we use generally, are produced in the concept of the raster model. BMP, JPG, and PNG are examples of this model. The Grid Map is also an example of the raster model. Like the Feature Map, the Grid Map does not have a standard of determining grid values. In Fig. 2, some examples of the vector model and the raster model are represented. 3.2 Subdivided Map Format In the previous section, the aim and the necessity of maps are mentioned. Now we propose the map formats for Function Map and Interface Map. Function Map is used for various self-controlled motions. Thus, the map format should contain the adequate information for self-controlled movement. For this purpose, Function Map has to offer the precise information about the existence of objects and the information changed should be easily updated in this map.

These are the basic conditions for the generation of an optimal robot path. Thus, the raster model is more adequate for Function Map than the vector model, in which the update of object position information is somewhat difficult. Among the various raster model formats, BMP, JPEG, GIF, and PNG formats can contain only the existence information of object, and thus the accuracy of map information is not satisfactory. The Grid Map is one of the raster model candidates which can solve this problem. Consequently, we decide the ‘Grid Map’ as the format of Function Map. The concept and example of the Certainty Grid Map is shown in Fig. 3. If we use only Grid Map for Function Map, a mobile robot or a server may waste their computing power every time when the re-planning of path accomplishes because of the environment changes. Additionally, the difficulty of robot control can produce the oscillation situation in robot movements. Because of this reason, the path planning and movement technology based via-point is developed, which guarantees the smooth robot motions and the low dissipation of computing power. The reasons that the via-point technology needs are the requirement of following three functions in robot motions. - Safe robot motions in the region where it is difficult to move. - Simplification of robot path planning by an offer of intermediate points to move. - Stable robot movements with the via-point to via-point motions. We define ‘Via-Point Map’ as the map which contains the via-point information to be used in this via-point technology. Of course, we construct Via-Point Map using the frame of Grid Map. For covering confusion, we want to refer that, in this paper, Grid Map means the map represents surroundings, and Via-Point Map means the map only represents via-points which do not exist in real circumstances. Via-Point Map is related directly to the robot movements, thus this map is included in Function Map. Consequently, Function Map is composed with Grid Map and Via-Point Map. Interface Map should offer the position information of robots and surroundings to a user. Additionally, it is desired that the service results are shown with familiar and well-formed shapes to a user. If we focus the condition of Interface Map, which should be able to offer information through various terminals, it is easy to decide that the raster model is not adequate because its format using grids is relatively restricted. Thus, we first select the vector model as the category of Interface Map. In the case of DXF format, it contains only the figure data and its size is relatively big. Thus, DXF lacks the compatibility to

Fig. 4. SVG Examples (left) and Related Codes (right) Fig. 3. Certainty Grid Map

109

The 3rd International Conference on Ubiquitous Robots and Ambient Intelligence (URAI 2006)

Table 1. Map Format for Mobile Robots 3.3 Network Map Map

Needs Accurate Information Easy Update

Function Map

Interaction Map

Various Terminal Scalability & Compatibility Intelligent Service Environment – adapted Service

Format

Raster Model

Grid Map

Via- Point

Via-Point Map

Vector Model

SVG Format

Interaction Rule

Rule-based Map

Smooth Motion Minimum Calculation Load

Interface Map

Category

various operating systems and it is difficult to use this format for searching information. The Feature Map is used frequently in studies and many researchers define the features for their research aim. But these studies are progressed only for the scientific purpose, and thus the standardization of determining and using the features is not discussed. Because of this reason, the Feature Map is not an adequate format for Interface Map. In contrast, SVG format has relatively small data size and is easy to be transplanted to many operating systems. Thus, ‘SVG’ can be selected for the format of Interface Map. In Fig. 4, some examples and related codes of SVG are represented. Interaction Map targets the intelligent service which a robot offers various and adequate services automatically without users’ frequent requests. For this purpose, a robot should have the ability of providing services adapted to its circumstances and situations. Namely, the environment-adapted services should be bestowed to users. For implementing the intelligent service actually, the Interaction Rules, which illustrate what users want to be offered, should be defined. These rules contain not only the services in the general environments but the services specialized to an individual user. The rules for the specialization can be inputted as a part of the initial commands for a robot and/or updated or added for the reflection of the environment changes. The basic principle for establishing these rules can vary according to various factors, but we recommend that the rules are based on the structure of natural human language. The Subject, the Object, and the Action may be essential elements within this principle. It is also recommended that the robot actions based the rules are assigned with the specified regions. Thus, ‘Rule-based Map’ is defined for this purpose. In Table 1, the abbreviation of map formats is represented.

The subdivided formats of environment map are mentioned in the previous section. Now we will discuss the management of these maps in this section. Before real discussing about the management, we should designate the spatial category of the robot environment. The maximum scope should be assigned firstly, and then the lower categories are progressed until the categorization reaches the basic unit for the environment map. We assume that robots act in indoor environment. In this case, the whole building may be the maximum scope about the robot tasks. This is the highest category. Every floor will be the next category. And ‘Spaces’ produced by the division of a floor become the lowest category. This word, ‘Space,’ is the specified one that represents a region divided by pre-defined method. Consequently, the ‘Space’ becomes the minimum unit for the environment map. Several ‘Spaces’ are gathered to construct higher hierarchy. If a floor has small size, the ‘Space’ in this floor may be this floor itself. A ‘Space’ has its own Grid Map. Additionally, Via-Point Map is added for the smooth motions of robots. As mentioned previously, these two maps belong to Function Map. SVG file is also required to show the shape and/or state of ‘Space’ to users’ terminals. If we use a large SVG file which contains all the information about all ‘Spaces’ of the building, some problems may arise. For example, processing time will be large because of large data size, and also the memory usage will increase. This makes it difficult to update the map even if relatively small change occurs in an individual ‘Space.’ Thus, we use an individual SVG file for each ‘Space.’ One map for rules, Rule-based Map, is added to these 3 maps. Namely, total 4 map files are assigned to each ‘Space.’ We call them as ‘0-Layer’. In the viewpoint of managing the robot motions, all maps included in a Space should be controlled simultaneously. We designate Network Map to do this role. One Network Map exists in the highest category of the environment map. This file is called as ‘1-Layer,’ and contain all information needed for the connections among ‘Spaces.’ Additionally, Network Map has Interaction Rules and the related command sets of robot actions. The role of Network Map can be divided into two tasks. First, it connects the spatial maps for the movements among ‘Spaces.’ Let assume that a robot receives the command that designates the movement from a point of ‘Space’ 1 to a point of ‘Space’ 4. Then the path planning is accomplished by Network Map as follows.

110

- Step 1 : Receive a command for a robot movement among ‘Spaces.’ - Step 2: Search a connection from the ‘Space’ that includes the start point to the ‘Space’ that includes the end point and then assign the path between two ‘Spaces.’ - Step 3: Assign the path from the start point to the exit point of the ‘Space’ that includes the start

The 3rd International Conference on Ubiquitous Robots and Ambient Intelligence (URAI 2006)

Fig. 5. Concept of the System point. Also assign the path from the exit point of the ‘Space’ that includes the end point to the end point. - Step 4: Move a robot following the pre-assigned paths. Second task of Network Map is the management of Interaction Rules. When a user designates some rules to the specified regions, Network Map receives this information. If robots come into these regions, Network Map orders the user-defined actions to robots. With this procedure, the processing load, which can be generated if the user’s terminal has the overall rules and related action data, can be greatly reduced and thus, smooth task achievement will be done even with the low-capacity terminal. The system, which has well-defined map formats, has the management scenario as follows; - Acquire information about the environment using various devices connected to system. - Update the low-layer maps. - Manage of the low-layer maps using the high-layer Network Map. - Interchange data among a server, robots and users and achieve the assigned task. The concept of overall system is represented in Fig. 5.

4. Conclusions According to the technology development, the demand for intelligent services increases more and more. Also, users want to receive the services which is not mechanize but humane. Thus, it is required to exclude the simple environment map and to make the map adequate for a human and a robot, respectively. For this purpose, the environment map formats are proposed. This proposal is based on the properties of a human and a robot are considered. Grid Map is chosen as the map that a robot can interpret and use, and this map has the merits that data processing and update are easy and fast. Additionally, Via-Point Map is defined to smooth robot motions. Function Map contains these two maps and is used for the robot motion control. A SVG format based XML is selected for Interface Map which is used for user’s easy understanding of the robot state and the environment. This format has the

benefits of scalability and compatibility, thus it can be applied easily to various terminals. Interface Map and Function Map are related to a human and a robot, respectively. And Interaction Map is employed to connect these two maps. For the implementation of Interaction Map, Rule-based Map is selected. Intelligent inference is needed for this Interaction Rules. ‘Space,’ which is the minimum unit for containing the spatial information, has Function Map, Interface Map and Interaction Map. These maps construct the low-layer of the environment map. The high-layer is needed to manage the low-layer. In our research, Network Map does this role. It covers not only the connect management among the low-layer maps, but management among ‘Spaces.’ The necessities and the conceptual roles of each map are discussed in this paper. In future we will implement the proposed maps with the real map data. Additionally, the in-depth study about the Interaction Rules is promising and needed for a future work.

Acknowledgements This work was supported in part by MIC & IITA through IT Leading R&D Support Project.

References [1] H. Zha, K. Tanaka, T. Hasegawa, “Detecting changes in a dynamic environment for updating its maps by using a mobile robot,” Proceedings of the IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS '97), Vol. 3, pp. 1729 – 1734, 1997. [2] R. HoseinNezhad, B. Moshiri, M. R. Asharif, “Sensor fusion for ultrasonic and laser arrays in mobile robotics: a comparative study of fuzzy, Dempster and Bayesian approaches,” Proceedings of IEEE Sensors, Vol. 2, pp. 1682 – 1689, 2002. [3] F. J. Toledo, J. D. Luis, L. M. Tomas, M. A. Zamora, H. Martinez, “Map building with ultrasonic sensors of indoor environments using neural networks,” IEEE International Conference on Systems, Man, and Cybernetics, Vol. 2, pp. 920 – 925, 8-11 Oct. 2000. [4] A. Elfes, “Sonar-Based Real-World Mapping and Navigation,” IEEE Journal of Robotics and Automation, Vol. 3, Issue.3, pp. 249-265, June 1987. [5] J. H. Lim, D. W. Cho, “Physically Based Sensor Modeling for a Sonar Map in a Specular Environment,” Proceedings of the IEEE International Conference on Robotics and Automation (ICRA ‘92), Vol. 2, pp. 1714-1719, May 1992. [6] S. Lacroix, G. Dudek, “On the Identification of Sonar Features,” Proceedings of the IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS ‘97), Vol. 2, pp. 586-592, 1997. [7] D. Wolf, A. Howard, G. S. Sukhatme, “Towards geometric 3D mapping of outdoor environments using mobile robots,” Proceedings of the IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS ‘05), pp. 1507 – 1512, 2005. [8] K. Nagatani, H. Ishida, S. Yamanaka, Y. Tanaka, “Three-dimensional localization and mapping for

111

The 3rd International Conference on Ubiquitous Robots and Ambient Intelligence (URAI 2006)

mobile robot in disaster environments,” Proceedings of the IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS ‘03), Vol. 4, pp. 3112 – 3117, 2003. [9] S. Thrun, C. Martin, Liu Yufeng, D. Hahnel, R. Emery-Montemerlo, D. Chakrabarti, W. Burgard, “A real-time expectation-maximization algorithm for acquiring multiplanar maps of indoor environments with mobile robots,” IEEE Transactions on Robotics and Automation, Vol. 20, Issue 3, pp. 433 – 443, June 2004. [10] S. M. Gordon, “Cognitive mapping from 3-D laser scans for use by humans and robots for localization and navigation,” IEEE International Workshop on Robot and Human Interactive Communication (ROMAN 2005), pp. 402 – 407, 13-15 Aug. 2005. [11] P. Althaus, H. I. Christensen, “Automatic map acquisition for navigation in domestic environments,” Proceedings of the IEEE International Conference on Robotics and Automation (ICRA '03), Vol. 2, pp. 1551 – 1556, 14-19 Sept. 2003. [12] R. Biswas, B. Limketkai, S. Sanner, S. Thrun, “Towards object mapping in non-stationary environments with mobile robots,” Proceedings of the IEEE/RSJ International Conference on Intelligent Robots and System (IROS ‘02), Vol. 1, pp. 1014 – 1019, 2002. [13] D. Hahnel, R. Triebel, W. Burgard, S. Thrun, “Map building with mobile robots in dynamic environments,” Proceedings of the IEEE International Conference on Robotics and Automation, Vol. 2, pp. 1557 – 1563, 14-19 Sept. 2003. [14] G. Oriolo, G. Ulivi, M. Vendittelli, “Real-time map building and navigation for autonomous robots in unknown environments,” IEEE Transactions on Systems, Man and Cybernetics, Part B, Vol. 28, Issue 3, pp. 316 – 333, June 1998. [15] S. Takezawa, G. Dissanayake, “Simultaneous localisation and mapping problems in indoor environments with stereovision,” 32nd Annual Conference of IEEE Industrial Electronics Society (IECON 2005), pp. 1896-1901, 6-10 Nov. 2005.

[16] D. Wolf, G. S. Sukhatme, “Online simultaneous localization and mapping in dynamic environments,” Proceedings of the IEEE International Conference on Robotics and Automation (ICRA '04), Vol. 2, pp. 1301 – 1307, Apr 26-May 1, 2004 [17] E. Fabrizi, A. Saffiotti, “Extracting topology-based maps from gridmaps,” Proceedings of the IEEE International Conference on Robotics and Automation (ICRA '00), Vol. 3, pp. 2972 – 2978, 2000. [18] A. C. Victorino, P. Rives, “Global consistency mapping with a hybrid representation,” Proceedings of the IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS ‘05), pp. 2554 – 2559, 2-6 Aug. 2005. [19] P. Buschka, A. Saffiotti, “Some notes on the use of hybrid maps for mobile robots,” 8th International Conference on Intelligent Autonomous Systems (IAS), Amsterdam, NL, pp. 547-556, 2004. [20] S. Coradeschi, A. Saffiotti, “Symbiotic Robotic Systems: Humans, Robots, and Smart Environments,” IEEE Intelligent Systems, Vol. 21, Issue 3, pp. 82 84 Jan.-Feb. 2006. [21] C. W. Nielsen, B. Ricks, M. A. Goodrich, D. Bruemmer, D. Few, M. Few, “Snapshots for semantic maps,” 2004 IEEE International Conference on Systems, Man and Cybernetics, Vol. 3, pp. 2853 2858, 2004. [22] C. Galindo, J.-A. Fernandez-Madrigal, J. Gonzalez, “Improving efficiency in mobile robot task planning through world abstraction,” IEEE Transactions on Robotics, Vol. 20, Issue 4, pp. 677 – 690, Aug. 2004. [23] C. Galindo, A. Saffiotti, S. Coradeschi, P. Buschka, J. A. Fernandez-Madrigal, J. Gonzalez, “Multi-hierarchical semantic maps for mobile robotics,” Proceedings of the IEEE/RSJ International Conference on Intelligent Robots and Systems(IROS ‘05), pp. 2278 – 2283, 2005. [24] J. P. Lee, Y. S. Kim, S. H. Oh, J. O. Lee, S. I. Park, S. Y. Jung, K. Y. Sung, H. K. Cho, S. H. Jung, N. S. Jung, “The Development & Conversion of the National Geography Institute Format,” Ministry of Construction & Transportation, Nov. 1999.

112