use this semantic map partition to navigate to a region with ... label for any point in the partition. ..... Experiences with an interactive museum tour-guide robot.
The 2010 IEEElRSJ International Conference on Intelligent Robots and Systems October 18-22, 2010, Taipei, Taiwan
Semantic Map Partitioning in Indoor Environments using Regional Analysis Carlos N ieto-Granda, John G. Rogers III, Alexander J. B. Trevor, and Henrik I. Christen se n
Abstract- Classification of spatial regions based on semantic information in an indoor environment enables robot tasks such as navigation or mobile manipulation to be spatially aware. The availability of contextual information can significantly simplify operation of a mobile platform. We present methods for 3utomHted recognition and classification of SI)aceS into separHte semantic regions and use of sueh information for generHtion of a topological map of an environment. The association of semantic labels with spatial regions is based on Humal/ Augmented Mappil/g. T he methods presented in this paper are evalu ated both in simu lation and on real data acquired from an office environment.
I. INTRODUC TIO N Humans are constantly trying to make their lives easier. Se rvice robots capable of operating in human environments have the pote nti a l to improve daily li fe by assisting humans in a variety of tasks. Endowi ng these robots with the abil ity to understand and reason about spatial regions such as indi vidual room s, as well as understanding the semantic labels of sllch spaces cou ld faci litate tasks such as navigation and mob ile manipulation in human envi ronments. Human env ironme nts are typically partitioned into di sc rete spaces, such as offices, corridors, li ving roo ms. etc. Such a partitioning allows human s to organize and enable their everyday acti vities, and these spaces typicall y have specirk: purposes and labels. Service robot s that understands the partitioni ng of human envi ronments can uti lize thi s information to better assist hum ans in everyday tasks. For example, if a robot is given the command "fetch the red mug from the kitchen" , having an understanding of the location and exte nt of the region considered "kitchen" is beneficial. In thi s paper, we prese nt a method for buildi ng a se manti c map partition in cooperation with a human guide. Given a metric map that the robot can loca li ze in. the system creates a se manti c map partition . The resulting semantic map part ition provides a probabili sti c classifi cati on of the metric map into a set of labe ls provided by the hum an guide. The robot can then use thi s semantic map part ition to navigate to a region with a specific label, and can determine the maxi mum like lihood label for any point in th e partition. Addi ti ona lly. if the robot is not confident that it knows a likel y semantic label for its current pose in the map, it wi ll prompt the guide to provide one. C. Nieto-Granda. J. Rogers Itl and 1\. Trevor and :Ire Ph .D. students at Georgia Tech College of Computing {car l os . nieto , atrevor ,
j g roge r s }@gate ch . e d u H. Christensen is the Kuka Chair of Robotics al Georgia Tech College of Computing hic@cc . gatech . edu
978-1-4244-6676-4/10/$25.00 © 201 0 IEEE
The paper is organi zed as follows. We briefl y describe related work in Section II , followed by the motivation of our researc h in Section III. We then present our Gauss ian probabilistic regions approach in Section rv. In Section V. we present so me experiments and resu lts both in simulation and on a real robot. Finally, conclu sions and future work are given in Section VI. II .
RELATED WORK
In rece nt years, we have seen important develo pments in service and assist ive robots for domestic app li cations and tasks. These works focu s o n the understanding of the environment using semantic information in order to create a sy nergisti c interaction between human s and robots. Dellaert and Brue mmerl 3 ) proposed ex tending FastSLAM to add se mantic infonnat ion of the environment (Q each particle's map. Several approaches have been presented fo r map partitioning, using topological and geometri c represe ntation s of the environment. For example. OberHinder Ill] proposed a SLAM algorithm based on FastSLAM 2.0 [91 that maps features represe nting region s with a se manti c type, topo logical properties. and an approximate geometric extent. The res ulting maps e nable spatial reasoning on a semanti c leve l and provide abstract in fo rmati on allowing effic ien t se mantic planning and a co nve nient interface for humanmachine interaction. Thrun 1"16] integrated grid-based maps to learn the environment using artificialncu ral networks and narvc Bayesian in tegra ti on to generate a lOpulugic~ll map by partitioni ng the latter into co herent regions. Another body of work focuses on extracting se manti c spatial properties of th e environment from 2D and 3D data. Donsu ng and Nevatia [71 introduced a new spatial represe ntation. s-map, for an indoor navigation robot. This map represents the locations of visible 3D-surfaces of obstacles in a 20 space. O'Callaghan [ 121 developed a new statistical modeling technique for bui lding occupancy maps by providing both a continuous representat ion of the robot's surroundin g and an associated predictive variance e mploying a Gaussian process and Bayesian learning. Ek va ll [41 applied an automatic st rategy for map partitio ning based on detecting borders between rooms and narrow open in g to denote doors or gateways usi ng different types of features (lines, points. SlIT). 'R hino' [21 is an example of a service robot which integrates localization. mapping. collision avoidance. planning. and va ri ous modules concerned with user interaction telepresence giving lo urs on a mu seum. BIRON, a mobile Home Tour Robot 115] . uses integrated vision based localization a modular architecture and extending a spoken dialog system for o n-line labeling
1451
and interaction about different location s in a real, full y furni shed home environment where it was able to learn the names of different rooms. Th e approach presented by Topp and Christensen ll 8J and [1 7J, provides a separation of regions that relare to a users view on the environment and detection of transitions between them . They assumed an interac ti ve sctup ror the spedfkation of regions and showed the applicability of their method in terms of di stin ct ive ness for space segmentation and in term s of locali sation purposes.
IV. A t' PROAC H
III. SEMA NTt C SLAM As robots have to cooperative with hum ans it is advantageous that they have a shared representation of the space, preferab ly a model that is simple for the human to use as part of commanding the robot and understanding feedback. Semantic mapping li terature has focused on developing robotic mapping techniques capable of functionally supporting these types of interaction s. To perform the se tasks, o ne of the strategies that is used is to portray the relationship between a place and the knowledge that is associated with it e.g.(functionality. obj ective location), is semantic mapping. Kuipers l8J proposed the Spalial Semantic Hierarchy (55 H), which is a qua litative and quantitative model of knowledge of large-scale space consisting of multiple interacting representati ons. This map also informs the robot of the control strategy that should be used to traverse between locations in the map. This represent atio n is based on the relationship of objects. actions and the dependencies from the environment. More recently. Beeson et al. f II provided a more specific framework representation of spatial know ledge in small scale space. This framework is focused on the robot 's sensory hori zon e.g.(global and local sy mbo lic, and metri ca l reasoning o f the space), but also human interaction. Existi ng approaches for robot indoor navigation bu i Id an occ upancy grid map using range dara from its se nsors. These maps, however, on ly provide geometric inform at ion such as obstacl es and open areas in the env ironment without a se mantic understa ndi ng of it. Martfnez-Mozus and Kottmann 1101 l l4J introduce a scmantic understand ing of the environment creati ng a conceptual representation refe rring to functiona l properties of typical indoor environments. Providin g se mantic info rmation enab les a mobile robot to more effici ently accomplish a variety of tasks sllch as hum anrobot interactio n, path-planning, and localization. Ekvall [41 integrated an augme nted SLAM map with information based on object recognition , providing a richer representation of the environment in a servi ce robot scenario. In this work , we focus on provi ding a semantic partition of a metric map using semantic labels provided by a human . We believe thi s representation cou ld be used to support semantic reasoning for a variety of mobil e robot tasks in indoor environments. As an example. we present navigation to the nearest puint in a metric map that has a spccific semantic label.
Our goa l was to des ign a syste m capable of reasoning about spaces. In contrast to work such as flO] . whi ch builds a topologi ca l map o n top o f a metric map. we instead provide a co ntinuous c lass ifi cati on o f th e met ric map into scmanti ca ll y labeled regions. The semantic map layer o f our system is a multi va riate probability di stribution on the coordinates of our metric map to a set of se manti c labels. This multivariate di stribution is modeled as a Gau ss ian mode l. Each of the Gaussians in the model is based on the robot's sensor data when it was provided a label by a human g uide. Eac h spatial reg ion is represent ed usin g one or more Gallss ians in our met ric map's coordinate fram e. So, a reg ion with label L and 11 Gaussians. each with mean I-t and covariance E , is rep rese nted as;
Region = {L. {litl , E l }, {1'.2, E z }, .... {IL" , E,, }}} A semantic map is then just a co ll ecti on of such regions. so a se manti c map with m region s wou ld be represented as;
Our system builds these maps part itio ns of our met ri c maps through human guidance. The human takes the robot on a tour of th e space (either by driving the robot manuall y, or lI sing a perso n fo ll owing behavior), and teaches the robot typing the appropri ate label for the space that it is currently in. The regi onal analysis technique is to take a laser scan mcasurcmt:nt , fi t a Ga uss ian tu the resulting puints. and store the mean and covariance in th e map along with th e labe l provided by the human. Using thi s se manti c map partition , the robot can be asked for its belief of the name of the region it is currently occupying. This is done by evaluati ng the Maha lanobis di stance of the robot' s current pose :r close by labels coded as Gau ss ian region mode ls (Equation I), and choosing the region that is cl osest using thi s metric.
DA/(.c) =
J(,. - ltrrE- l(:< - IL)
( I)
Thi s map represe ntati on all ows for prohahili sti c class ifi cation of the map by region label. Additionally, while navigat ing throu gh the envi ron ment , the robot continuously checks its posi ti on with re spec t to the semantic Illap partition. If it is not sufficiently confident (more than a ce rtain thresho ld ) that it is in a region with a known labe l, it prompts the user to inp ut the name of the current region . Once the robot has a se mantic map partition, use rs can requ est that the robot navi gate to onc of the regions. such as " livi ng roo m" . The robo t can then find the region in the map with label " living room", and calculate the Mahalanobi s di stance from its current position 10 the mean of each Gaussian in the region. The robot selects the closest of these as the goal, and se nd s thi s to its path planner in order to autonomou sly navigate to that region. While traveling. the robot continuou sly calculates its confidence or which region
1452
it is. and stops when it is con fidem that it is mo re likel y to be in the goa l region than any other region as follows:
in orde r to test the system's effectiveness ,)(
