Use of an Autonomous Mobile Robotfor Elderly Care

Karsten Berns, Syed Atif MehdiRobotics Research Lab, Department of Computer Sciences

University of Kaiserslautern, Kaiserslautern, GermanyEmail: {berns,mehdi}@cs.uni-kl.de

Abstract—In our societies, there is a steady increase in theelderly population. These people need a constant supervision toperform simple tasks at their homes, which is sometimes notpossible. This lack of permanent attention sometimes resultsin the late detection of emergency situations. Therefore, it isjudicious to promote technology that helps to detect and reactin case of emergency situations. In the last decade, monitoringdevices have been installed in the home environments for thesurveillance of the inhabitant. However, due to poor privacythere will be an unfavourably low acceptance of such systems.Another approach to the problem is to develop mobile robotsfor elderly care. In this scenario, a mobile robot has to navigatefrom place to place to search for the elderly person. Navigatinginefficiently in the home environment may delay the processof searching. This paper focus on the aspect of enhancingnavigational efficiency of the robot that would speed up thetask of finding human being in the unstructured and dynamichousehold environment. To achieve precision in searching thehuman, a behaviour based Markov decision process (MDP) alongwith a detailed representation of the environment as a grid maphave been developed on a small sized indoor autonomous mobilerobot, ARTOS.

Index Terms—Elderly Care, Indoor Robot, Service Robot

I. INTRODUCTION

The demographic situation in many developed countries isshowing a steady growth in the population of senior citizens.Elderly people are often not able to perform all activitiesof their daily life without the help of caregivers and face ahigher risk of experiencing a medical emergency in unattendedsituations. Therefore, they usually have to move to assistedliving facilities where they are looked after by the nursing staff.Due to limited number of nursing staff and increasing costs ofsuch facilities, many research groups use modern technologiesto assist elderly people at their homes. The aim is not only tolower the nursing costs but also to increase the quality of lifeof senior citizens.

Autonomous Robot for Transport and Service (ARTOS),Figure 1, is an initiative to provide services to the elderlypeople, living alone in their homes. It is equipped with laserrange finder, ultrasonic sensors, RFID1 reader, tactile sensorsand a pan-tilt-zoom (PTZ) camera. Figure 2 shows the sensingrange of ARTOS. The control system is based on the MCA-KL2. Many components are being developed as behaviours of

1Radio Frequency IDentification2MCA-KL: Modular Controller Architecture - Kaiserslautern Branch

(http://rrlib.cs.uni-kl.de/)

Fig. 1. Autonomous Robot for Transport and Service (ARTOS)

the behaviour based control architecture iB2C3 [1].

Bumper

Laser Range Finder

UltrasonicSensors

Fig. 2. Sensing Range of ARTOS

ARTOS has been specially designed for indoor living en-vironments and is able to navigate through narrow corridorsand closely placed furniture in the living environment. Atelecommunication link can be established between the elderlyperson and the caregivers using wireless Internet of the robot.Moreover, it can be tele-operated in the environment by thehealth care personnel to detect medical emergencies.

To enhance the privacy of the elderly person and relievethe caregivers from extensive supervision all the time, ARTOSshould be able detect and analyse the emergency situation onits own as a first step. In case any such situation is detected,as a second step, an emergency call to the caregiver should beestablished for immediate help and analysis of the situation. Toachieve this goal, primary task is to search the human being inthe environment while navigating autonomously through the

3iB2C: integrated Behaviour-Based Control

2010 Advanced Technologies for Enhancing Quality of Life

978-0-7695-4280-5/10 $26.00 © 2010 IEEE

DOI 10.1109/ATEQUAL.2010.30

121

obstacles in the home.In the following, we present an autonomous methodology

for searching the human being in the environment. The ap-proach takes into account the obstacles in the environmentand tries to reach the inhabitant at places where his presenceis most likely at that time.

To explain this concept and implementation, the paper isorganised in the following way. First of all a short summary ofrelated work is presented. Section III presents the methodologyof the approach along with the major modules required forsearching the human being in the environment. The workingenvironment of ARTOS and its simulation are being discussedin Section IV. Finally at the end, conclusion and future workare presented.

II. RELATED WORK

Observing the elderly person at home can be accomplishedby installing a dense video supervision framework within theapartment to cover every corner, as [2] has developed a falldetection system using multi-cameras. The Aware Home atGeorgia Tech in Atlanta [3], house n [4] as part of Placelabat MIT, the Assisted Living Lab in Kaiserslautern, Germany[5] and HomeLab from Philips in Eindhoven, NL, are onlysome of the examples where a variety of sensors have beenmounted to observe the elderly person. Installation of sensorsystem in the home environment is not only expensive, interms of cost of installation, but also requires manipulationin the living environment for instance installation of wiresand mounting of different sensors. Even more important,this setup critically affects the privacy of the inhabitant. Theresult of study conducted by [6], revealed that elderly peoplerequire that an ambient assisted home should do a lot of goodthings for them but there should not by any surveillance ofinhabitants. Moreover the question, “Are there any hiddencameras or hidden microphones in the home?”, always comeahead whenever there is a reference to the ambient assistedliving (AAL).

A possible solution to the above problem is to use mobilerobots that can be controlled remotely by the helping personnel[7], [8] and [9]. In this way, resident of the home preciselyknows when he is being watched.

A mobile robot at home should offer services to the elderlypeople besides being only an observation channel for the care-givers. The Robocup@Home proposes and runs competitionsconcerning the abilities of home service robots [10].

In this context, we have developed an autonomous mobilerobot, ARTOS, that not only being a channel for communi-cation between elderly person at home and caregiver but alsocapable of being a transportation unit and a service robot.Besides that, certain measures with respect to privacy andsecurity of the elderly person have also been taken into accountin ARTOS. Among many others, the need of the robot beingautonomous and detecting the emergency situation on its ownwill be some of the features of ARTOS.

III. METHODOLOGY

Besides primary tasks of transportation and service, ARTOSalso needs to detect the medical emergency situations that anelderly person may encounter. Figure 3 shows one of suchexample where an elderly person has been detected lying onthe floor of the apartment by the remote caregiver using webinterface of ARTOS.

Fig. 3. An application scenario for ARTOS - An elderly person in emergencysituation

In order to facilitate monitoring the human being in theenvironment for medical emergencies, ARTOS has to becapable of predicting the location of the elderly person inthe apartment at particular times. Searching randomly for theinhabitant in the home can effect the time required for thisprocess. Moreover, the home environment does not providean easy to navigate place for robots. Therefore, preferenceshould be given to the places where it is more likely to findthe human at that particular time.

Fig. 4. Overview of the Methodology

To address the issue, an approach similar to [11], whereMDP model has been used to schedule tasks in the office

122

environment, and [12], where a variant of MDP (namely RN-POMDP) is described, is being proposed.

In this paper, a behaviour based Markov decision process(MDP) has been developed to build up cognitive informationof the human. It will help in reaching the inhabitant todetermine health situations. The idea is to enable the robotto comprehend when it is required to find the human being onits own. To achieve this task, behaviour based approach hasbeen adapted where behaviour represents the desire to find thehuman and the policy for finding human is determined usingMDP. Figure 4 gives an overview of the proposed methodologyof searching the human in the environment and the followingsubsections explains the main components of the system.

A. Mapping and Localisation

The main objective of the mapping system is to supportprecise navigation in the environment. Therefore grid mapapproach has been chosen for building the map. A grid maprepresents the world around the robot as an array of (usuallyuniform) grid cells. Each cell stores information about the areait covers, with the most important information usually beingwhether the area is occupied or not. The current occupancybelief is represented by an occupancy counter. Positive valuesare used for occupied cells, negative values for free cells andan occupancy counter of zero reflects an unknown occupancystate. The occupancy counters are limited so that a belief inan occupancy state cannot get too strong.

The laser range finder as well as the two chains of ultrasonicsensors are used as sources for the grid map creation process.The sensor data can be obtained in one of the two followingformats:

1) Polar format: The sensor values are stored as a seriesof distance-angle-pairs. This format is used for the datafrom the laser range finder.

2) Sector map format: The polar sector maps are used hereto access the data generated by the ultrasonic sensors.

The integration of sensors information is carried out usingdifferent arrays of occupancy counters for different sensors,otherwise an obstacle that is only visible to sensor 𝑆1 wouldbe removed when processing the data of sensor 𝑆2, whichdoes not see it. Therefore, in every execution cycle, the dataof each sensor is processed and the corresponding array ofoccupancy counters is updated. Then the data of these arraysis aggregated to build one combined grid map according tothe following rules:

1) If a cell is occupied in at least one sensor’s array, thenits counter in the combined grid map is set to +1.

2) If a cell is not occupied in any of the sensors’ arraysand is free in at least one sensor’s array, then its counterin the combined grid map is set to −1.

3) If a cell is not occupied or free in any of the sensors’arrays, then its counter in the combined grid map is setto 0.

Further details of the mapping can be found in [13].Being a service robot, ARTOS should precisely know its

location in the environment. Therefore, the localisation in the

Fig. 5. RFID laid under the carpet to be used by localisation module

environment is achieved using differential odometry and isimproved by reading passive RFID tags that are installed underthe carpet. As can be seen in Fig. 5, a grid of about 4, 000passive RFID tags with a grid size of 12.5 cm by 12.5 cm(5 in by 5 in) has been implanted under the carpet of testapartment (see Fig. 7).

𝑁 = number of RFID tags in range (1)

Pos𝑥,𝑦 =1

𝑁

𝑁∑𝑖=1

(𝑥𝑖

𝑦𝑖

)(2)

Equations 1 and 2 show how a position estimate Pos iscalculated. It is simply the mean value of the 2D positionsof all 𝑁 RFID tags in range. The orientation estimation isbased on detecting several tags while the robot is moving. Acombined arithmetic and heuristic calculation is performed toestimate the robot’s orientation. Many experiments have shownthat this approach is sufficient for indoor navigation.

B. Navigation

ARTOS can navigate autonomously in the environment. ThePath Planner searches the path between the obstacles usingA* algorithm to generate the smallest possible path betweentwo points. This algorithm processes cell by cell, starting with𝑠, until a path to 𝑑 has been found. In each processing step,the cell with the lowest cost is chosen as the next cell to beprocessed. For a cell 𝑐, the cost 𝑓(𝑐) is

𝑓(𝑐) = 𝑔(𝑐) + ℎ(𝑐) (3)

where 𝑔(𝑐) denotes the cost for the shortest known pathgoing from 𝑠 to 𝑐 and ℎ(𝑐) denotes the estimated cost of apath going from 𝑐 to 𝑑. The Euclidean distance is used asheuristic function. Before planning a path, the obstacles inthe map are enlarged by marking the cells close to them as“neighbours”. High costs are assigned to these neighbours sothat other free cells may be preferred. As a result, paths donot lead the robot close to obstacles unless it is necessary.

The path planned by the Path Planner is the shortestpath from the source to the destination. Following exactlythose points might result in getting the robot too close tosome obstacles which will cause reduction in speed and thusthe robot might take a longer time to reach its destination.

123

To overcome this problem, the elastic band approach [14]has been used. The algorithm the uses the path informationgenerated by the Path Planner and optimises it with respectto a smooth robot motion, a short driving time, and a safedistance to obstacles. The Point Approacher receives targetcoordinates and the current robot pose as input, and calculatesa desired velocity 𝑣des and angular velocity 𝜔des dependingon the distance 𝑑 and absolute angle ∣𝛼∣ to the target (seeEquations 4 and 5):

𝑣des =

⎧⎨⎩0 ; 𝑑 ≤ 𝑑min

1 ; 𝑑 ≥ 𝑑max12 + 1

2 ⋅ sin(( 𝑑−𝑑min𝑑max−𝑑min

− 12 ) ⋅ 𝜋) ; else

(4)

∣𝜔des∣ =

⎧⎨⎩0 ; ∣𝛼∣ ≤ 𝛼min

1 ; ∣𝛼∣ ≥ 𝛼max12 + 1

2 ⋅ sin(( ∣𝛼∣−𝛼min

𝛼max−𝛼min− 1

2 ) ⋅ 𝜋) ; else(5)

By comparing the robot’s orientation to 𝛼, 𝜔des can be calcu-lated from ∣𝜔des∣. 𝑑min, 𝑑max, 𝛼min, 𝛼max mark the distances andangles at which 𝑣des and 𝜔des, respectively, take their extremevalues. If 𝑑max is reduced, for example, the robot will drivelonger at the maximum speed when approaching a target. Ifit should decelerate earlier, 𝑑max has to be increased. To makethe changes of 𝑣des and 𝜔des smoother, sigmoid functions areused.

C. Searching Human

For searching the elderly person in the environment, certainnumber of points are used as reference points. When therobot ”feels” to find the human being, it evaluates the costof reaching those points from the current location and theprobability of presence of person at reference points at thatparticular time and evaluates a policy to drive to one ofthe reference point. The person will be detected using theimages from camera. In the following a brief description ofimplementation is given.

1) Markov Decision Process: The Markovdecision process (MDP) is defined as a five-tuple(𝑆,𝐴, 𝑆′, 𝑅(𝑆,𝐴), 𝑇 (𝑆,𝐴, 𝑆′)), where 𝑆 is the currentstate, 𝐴 is the action performed at 𝑆, 𝑆′ is the resulting state,𝑅(𝑆,𝐴) is the reward of performing 𝐴 at 𝑆 and 𝑇 (𝑆,𝐴, 𝑆′)is the probability of reaching 𝑆′ when 𝐴 is performed at 𝑆.The value function for the state is defined by Eq. 6.

𝑈(𝑆) = max𝐴

⎛⎝𝑅 (𝑆,𝐴) + 𝛾 ∗

∑𝑆′

𝑇(𝑆,𝐴, 𝑆

′) ∗ 𝑈 (𝑆′)⎞⎠(6)

The policy for an action to be performed at 𝑆 is determinedby Eq. 7.

𝜋(𝑆) = argmax𝐴

(∑𝑆

𝑇(𝑆,𝐴, 𝑆

′) ∗ 𝑈 (𝑆′))(7)

Currently, the reference points are used to monitor thehuman presence. The probability of finding human, 𝑃 (𝑆), at aparticular time is computed at these points. Here 𝑅(𝑆,𝐴) is theprobability of finding the person, 𝑃 (𝑆′), when 𝐴 is performedat 𝑆. 𝑇 (𝑆,𝐴, 𝑆′) is normalised value of 𝑃 (𝑆′) times the in-verse of Navigational Cost to reach 𝑆′ . This cost incorporatesthe cost of avoiding obstacles in the environment and is basedon the A*- algorithm computed by Path Assessment module.

In the current scenario, the 𝑆 is taken as the current positionof the ARTOS where human has not been found, 𝐴 are theactions that can be performed to find human from 𝑆, 𝑆′ is theresultant state after performing 𝐴 and is most likely the placewhere human can be found. The information of presence ofhuman being in apartment at different places at different timeshas been generated and shown in the Fig. 6. The apartmenthas been divided into five zones, namely Bedroom, TV room,TV-Kitchen, TV-corridor and Corridor based on the presenceof the person. TV-Kitchen is the area between the kitchen andthe TV room and a person in the kitchen can be seen fromthis area. Similarly, TV-corridor is the area between the TVroom and the corridor. These probabilities clearly shows themovement of human being from one room to the other atdifferent times of the day.

Fig. 6. Probability of presence of an elderly person in living apartment atdifferent times

2) Behaviour Based Control Architecture: The behaviourhas been defined as a three-tuple

𝐵 = (𝑓𝑎, 𝑓𝑟, 𝐹 ) (8)

where 𝑓𝑎 is the activity function, 𝑓𝑟 is the target ratingfunction and 𝐹 is the transfer function. The behaviour basedarchitecture has been discussed in detail in [15].

In the current implementation 𝑓𝑟 is computed by Eq. 9

𝑓𝑟 =𝑇 𝑖𝑚𝑒𝐸𝑙𝑎𝑝𝑠𝑒𝑑𝑆𝑖𝑛𝑐𝑒𝐿𝑎𝑠𝑡𝑆𝑒𝑎𝑟𝑐ℎ

𝐴𝑙𝑙𝑜𝑤𝑒𝑑𝑇 𝑖𝑚𝑒(9)

where, TimeElapsedSinceLastSearch is the time when therobot has last seen the person and AllowedTime is the max-imum time the robot can stay contented for not seeing or

124

observing the person. Input vector is the probability of findingthe human at a particular time at all the places. The activityfunction, 𝑓𝑎 , depicts that the robot is engaged in searchinghuman and the transfer function, 𝐹 , is based on the MDPwhich results the policy for the current state. Inhibition willbe used in future to inhibit the behaviour.

IV. EXPERIMENTS

A real home environment has been established at IESE,Fraunhofer, Kaiserslautern, Germany, to evaluate the perfor-mance of the overall system. It is a fully furnished apartmentwith an area of 60𝑚2, comprising of a living room, a bedroom, a kitchen and a bath room. Figure 7 shows an overviewof the apartment and Fig. 8 shows the map of the environmentas generated by ARTOS using laser scanner and ultrasonicsensors. It can be seen easily that the environment is clutteredwith obstacles and there is not much space for robot naviga-tion. The red grids shows the obstacles in the environment andthe orange grids are marked as neighbours of the obstacles.The white elements are used for finding the path from sourceto the destination, while pink grids is the navigational path thatARTOS has to follow to avoid collisions. The blue circles arethe elastic bands to evaluate a safe distance between the robotand the obstacles.

DN

A

Elderly P

erson

DE

MO

CE

NT

ER

CareG

iver

Relative

PC

MONA

2961mm

835mm

4444mm

1140mm

2112mm

iCup

Set Top B

ox Cam

era

TV

RF

ID

Am

iCooler

Fig. 7. The testing apartment at IESE-Fraunhofer

Fig. 8. Map and Navigational Path of ARTOS in the apartment

A diagram of its velocity during navigation from corridorto the kitchen is shown in Fig. 9. As can be seen, the robot

did not stop or halt due to obstacles while following the path.This is an important result as it demonstrates that the mappingand obstacle avoidance components detected the obstructionsso early that a way leading around them could be calculatedbefore the robot got stuck.

-0.10

0.00

0.10

0.20

0.30

0.40

0.50

0.60

0.70

0.80

0.90

1.00

5 15 25 35 45 55 65 75

Time [s]

Sca

led

Vel

oci

ty

Fig. 9. ARTOS’ velocity during its navigation from corridor to the kitchen(scaled to [−1.0; 1.0]).

The robot’s path during the experiment is shown in Fig. 10.The gaps in the path are caused by the RFID-based posecorrections, which made the estimated pose “jump”.

0

1

2

3

4

5

6

7

0 1 2 3 4 5 6 7

Position X [m]

Po

siti

on

Y [

m]

StartLocation

GoalLocation

Fig. 10. The path of the robot from the corridor to the kitchen

The experiments to prove the correctness of human searchhave been done in simulation. The simulation consists of anapartment like IESE, Fraunhofer. The simulated environmentfeatures necessary furniture, a human being that moves inthe environment and the ARTOS with simulated sensors,see Fig. 11. An image obtained by the camera attached toARTOS in simulation is shown in Fig. 12, which will beused for detecting the human being. Experiments conductedin simulation shows that ARTOS successfully reaches thelocation where the presence of human being is most likelyat that particular time.

V. CONCLUSION AND FUTURE WORK

This paper focuses on the components required for search-ing the human being in the environment. Considering thedynamics of living environments, different sensor systems are

125

Fig. 11. Simulation of IESE-Fraunhofer

Fig. 12. View from Simulated Camera on ARTOS

being used as basis for the map creation so that a large varietyof obstacles, like furniture, are recognised. The mappingmodule generates a detailed information of the environmentin the form of a grid map. The map is updated to consider theunstructured and dynamic obstacles in the environment. Theprecise location of the robot in the environment is achievedusing the RFID landmarks by the localisation module. TheA* algorithm is used to find the path to the destinationwhich uses the grid map information to generate the bestpossible path. To follow the generated path, Point Approacheris used. The mapping and navigation systems developed hereenable the robot, ARTOS, to move autonomously in the livingenvironments.

The searching of human is carried out using the probabilisticanalysis of presence of human being in the environment.Markov decision process is being used to generate the policyfor navigating autonomously to next location in searching thehuman being. Currently, the searching is being performedin the simulation but the scope of the experiments will beenhanced to the real apartment. Moreover, human detectionbased on the camera images will be implemented to verifythe human being in the environment.

ACKNOWLEDGEMENT

We are thankful to Higher Education Commission of Pak-istan and DAAD Germany for funding of Syed Atif Mehdi. We

also like to thank IESE, Fraunhofer for support in conductingexperiments in Assisted Living Lab.

REFERENCES

[1] M. Proetzsch, T. Luksch, and K. Berns, “The behaviour-based controlarchitecture iB2C for complex robotic systems,” in Proceedings ofthe 30th Annual German Conference on Artificial Intelligence (KI),Osnabruck, Germany, September 10-13 2007, pp. 494–497.

[2] R. Cucchiara, A. Prati, and R. Vezzani, “A multi-camera vision systemfor fall detection and alarm generation,” Expert Systems, vol. 24, no. 5,pp. 334–345, November 2007.

[3] G. Abowd, A. Bobick, I. Essa, E. Mynatt, and W. Roger, “The awarehome: Developing technologies for successful aging,” in Procceedingsof the Workshop on Automation as a Care Giver at the AmericanAssociation of Artificial Intelligence (AAAI), Alberta, Canada, July 2002.

[4] S. Intille, K. Larson, and E. M. Tapia, “Designing and evaluating tech-nology for independent aging in the home,” in International Conferenceon Aging, Disability and Independence (ICADI), Washington DC, USA,December 2003.

[5] J. Nehmer, A. Karshmer, M. Becker, and R. Lamm, “Living assistancesystems - an ambient intelligence approach,” in Proceedings of the 28thInternational Conference on Software Engineering (ICSE), Shanghai,China, May 20-28 2006.

[6] W. L. Zagler, P. Panek, and M. Rauhala, “Ambient assisted living sys-tems - the conflicts between technology, acceptance, ethics and privacy,”in Assisted Living Systems - Models, Architectures and Engineering Ap-proaches, ser. Dagstuhl Seminar Proceedings, A. Karshmer, J. Nehmer,H. Raffler, and G. Troster, Eds., no. 07462. Dagstuhl, Germany: SchlossDagstuhl - Leibniz-Zentrum fuer Informatik, Germany, 2008.

[7] P. Deegan, R. Grupen, A. Hanson, E. Horrell, S. Ou, E. Riseman, S. Sen,B. Thibodeau, A. Williams, and D. Xie, “Mobile manipulators forassisted living in residential settings,” Autonomous Robots, Special Issueon Socially Assistive Robotics, vol. 24, no. 2, pp. 179–192, February2008.

[8] F. Michaud, P. Boissy, H. Corriveau, A. Grant, M. Lauria, D. Labonte,R. Cloutier, M.-A. Roux, M.-P. Royer, and D. Iannuzzi, “Telepresencerobot for home care assistance,” in AAAI Spring Symposium on Mul-tidisciplinary Collaboration for Socially Assistive Robotics, Palo Alto,USA, March 2007.

[9] A. Tapus, M. Mataric, and B. Scassellati, “The grand challenges insocially assistive robotics,” Robotics and Automation Magazin, vol. 14,no. 1, pp. 35–42, 2007.

[10] T. van der Zant and T. Wisspeintner, “Robocup@home: Creating andbenchmarking tomorrows service robot application,” in Robotic Soccer,P. Lima, Ed. Itech Education and Publication, December 2007, ch. 26,pp. 521–528.

[11] M. Beetz, T. Arbuckle, T. Belker, A. B. Cremers, and D. Schulz,“Integrated plan-based control of autonomous robots in human envi-ronments,” IEEE Intelligent Systems, vol. 16, 2001.

[12] A. F. Foka and P. E. Trahanias, “Probabilistic autonomous robotnavigation in dynamic environments with human motion prediction,”International Journal of Social Robotics, no. 1875-4805, pp. 79–94,2010.

[13] S. A. Mehdi, C. Armbrust, J. Koch, and K. Berns, “Methodologyfor robot mapping and navigation in assisted living environments,” inProceedings of the Workshop on Robotics and Automation in AssistiveLiving Systems at the International Conference on Pervasive Technolo-gies Related to Assistive Environments 2009 (PETRA 2009), Corfu,Greece, June 9-13 2009.

[14] S. Quinlan and O. Khatib, “Elastic bands: Connecting path planningand control,” in Proceedings of IEEE Int. Conference on Robotics andAutomation, Atlanta, 1993, pp. 802–807.

[15] M. Proetzsch, T. Luksch, and K. Berns, “Development of complexrobotic systems using the behavior-based control architecture iB2C,”Robotics and Autonomous Systems, vol. 58, no. 1, pp. 46–67, 2010.

126