26th International Symposium on Automation and Robotics in Construction (ISARC 2009)

Framework to Improve Mobile Robots Navigation Using Wireless Sensor Modules

Diego Martinez 1 and Yong K. Cho 2 Research Assistant, Computer Engineering and Electronics Department, College of Engineering, University of Nebraska-Lincoln, 1110 S. 67th St. Peter Kiewit Institute (PKI) 118A, Omaha, NE 69182; PH(402)554-3277;FAX (402)554- 3850; email: [email protected] 2Assistant Professor, Construction Engineering and Management Division, Charles Durham School of Architectural Engineering and Construction, College of Engineering, University of Nebraska-Lincoln, 1110 S. 67th St. Peter Kiewit Institute (PKI) 104C, Omaha, NE 69182; PH(402)554-3277;FAX (402)554- 3850; email: [email protected]
1Graduate

Abstract
As on-going research, this paper presents a framework to improve wireless mobile robots navigational accuracy in diverse indoor environments where the signals are affected by various types of interference including electromagnetic, multi-path, and fading and scattering. In particular, indoor construction environments pose unique challenges to accurate wireless navigation due to their relative complexity and inherently dynamic nature. Several integrated location and orientation sensors including a digital compass, a gyroscope, wheel encoders, an accelerometer, and Ultra Wideband (UWB) position tracking sensors are introduced in this paper. A distinct cause of error for each sensor is studied based on location, traveling distance, and rotational angle. To improve the position data accuracy, statistical methods such as outlier analysis and the Kalman Filter are applied in this research. A framework for position and orientation error compensation between relative and absolute sensors is described with preliminary research results indicating that position and orientation errors can be statistically adjusted in real time.

Keywords: mobile robot, navigation, dead reckoning, kalman filter, wireless sensor, error
1 Introduction The competitive, market oriented, and rationalized construction of tomorrow will require developing automated and robotized construction system today [4]. This includes indoor construction applications such as interior finishing, piping, excavation, mining, and earth moving [5] among others. In particular, implementing suitable indoor localization in construction processes will lead to an increase in productivity and improvement in work quality and working conditions [4]. The requirement to have reliable positioning is becoming increasingly important, and can be used for indoor position application such as implementing robots with mobile platform for construction tasks [4]. However, implementing mobile robots for construction tasks has proven to be difficult due to the dynamic and uncertain nature of the construction site [5]. It is critical that mobile robots absolute and relative positions are accurately determined in both outdoor and indoor environments. Global Positioning Systems (GPS) are widely used to determine the absolute position in outdoor environments where the signals are not obstructed by nearby buildings or trees. It is possible to determine the mobile robots position with an accuracy of 2-3 cm with well-equipped GPS systems. However, GPS signals cannot travel through walls and thus suffer from signal attenuation, making GPS systems unsuitable for indoor applications. One of the most prominent technologies used for indoor applications is ultra wideband (UWB). UWB provides good performance within the boundaries of a small area, generally within 10-15 cm accuracy. This paper describes on-going research into the development of a new method of autonomous navigation as applied to wireless mobile navigation. This will provide wireless autonomous mobile navigational functions to a robot on construction sites. The main goal of this research project is to integrate mobile robots inertial navigation control unit into a UWB indoor positioning system, and identify and
553

ation and Computational Technology Informa correct the so ource of errors s using well-kn nown statistic cal methods, su uch as the Ka alman Filter an nd outliers analysis. 2 System Ov verview The robot t control unit should maxim mize the chanc ce to reach its goal. The mobile robot sho ould be able to o measure prog gress in its rela ative frame in order to comp pare these me easurements with w its absolut te frame. This s is shown in Figure F 1.

Absolute frame f

Relativ ve frame

1. Absolute Fr rame vs. Relati ive Frame Figure 1 Data comi ing from a gyr roscope, an ac ccelerometer and a wheel enc coders are com mbined to provide the robot with orientation and positio on in its relativ ve frame. This is known as Dead Reckon ning, or propio oceptive navigation. Th etc. These he position ac ccuracy is affected by slippa age, uneven su urface, wheel disalignment, d errors can inc crease in sever rity over time, however they y can be mode eled, thus pred dicted and cor rrected, using statistical anal lysis. Data obtai ined from UW WB sensors an nd a digital com mpass provide es the absolut te position and d orientation for the robot s absolute fra ame. This is kn nown as exteroceptive navig gation. The ac ccuracy of these devices is limited by the e resolution of f the devices. 2.1 Localizatio on systems Estimating g the position of a robot in its environme ent requires th he knowledge of the geome etry configuration n (x, y, ) for th he frame attac ched to the robot, with resp pect to a local coordinate fra ame. This is done by deter rmining two different d refere ence frames. The first one is a relative fr rame that mak kes use of wheel encode ers and inertial l sensors such h as a gyroscop pe or an accele erometer. Th he second is th he absolute frame, which requires the use u of external l sensors, such h as a UWB an nd a digital co ompass. te Localization n 2.1.1 Absolut The absolu ute frame will l be relative to o the space. Th his frame will have also a 0, ,0 position, wh hich must be related to the relative frame e of the robot t. It has to be updated u from m the robot geo ometry inform mation. The 2D D re determined d by the readin ngs of the UW WB sensors in this research. The heading will be coordinates ar determined by y the digital co ompass.

UWB System m

The Ubise ense system is an ultra wide-band (UWB) ) position system used in th his research wh hich measures s Time of Diffe erence of Arri ival (TDOA) a and Angle of Arrival A (AOA A) to achieve positioning[6]. p The system consists of a number n of fix xed sensors wh hich receive UWB U pulses fr rom active bat ttery-powered tags. The fixed sensors are networked d over Ethern net, and the measurement m d is processed on the Ubi data isense softwar re ive the real tim me track of th he mobile tag [8]. [ platform to gi

Digital Comp mpass

This senso or provides th he absolute ori ientation in th he navigation control. c The main m disadvantage of the digital compa ass is that the earths e magne etic field is affe ected by electr romagnetic fie elds. This mak kes the use of this sensor ha ard to implement in indoor environment for absolute position. p
5 554

26th International Symposium on Automation and Robotics in Construction (ISARC 2009) 2.1.2 Relative localization The relative frame supplies the dimensions of the robot. It will be relative to the robot. This frame will have a (0,0) position at the beginning and an angle of 0. The heading of this frame will be determined by the gyroscope, and the translation will be determined by the encoders. Odometry is defined as the use of encoder measurements at the wheels to estimate the configuration of the robot state (position and orientation). To achieve successful autonomous mobile robot navigation, accurate odometry is essential. Localization, mapping and path-planning algorithms are all fundamental for robot navigation and all use odometry information [3]. A wheel encoder is the essential sensor used in odometry. It is a device that is used to convert the angular position of a shaft to a digital code. It provides the distance in which a wheel has travel by basically measuring the relative distance. The purpose of inertial sensors is to calculate the relative change of a moving target between two consecutive sampling times, based on the measurement of acceleration and angular velocity from the inertial sensors [1]. In order for the inertial sensors to function properly, the gyroscope must be set up parallel to the direction of motion of the robot.

Odometry

Wheel Encoders

Inertial Sensors

Gyroscope

The gyroscope measures an angular rate by picking up the signal generated by an electromechanical oscillating mass as it deviates from its plane of oscillation under the Coriolis force effect when submitted to a rotation about an axis perpendicular to the plane of oscillation. Gyroscope errors come from bias drift and noise. They are of particular importance for robot positioning because they can compensate the weakness of odometry.

Accelerometer is used to measure the rate of acceleration. This sensor is used to measure the accelerations of the mobile robot. It enables the control system to know when the mobile robot is at rest. Gyroscopes and accelerometers are used to measured rotation and acceleration. These sensors have the advantages to be self-contained, meaning that they do not need external references. The bias drift caused by the gyroscope can be fixed by using the accelerometer readings and the Kalman filter. Once the absolute localization is obtained, it can be fused with the estimation of the robots relative position so that it can correct its trajectory. 3. Source of errors

Accelerometer

Gyro + Accelerometer

Relative Location + Absolute Localization

UWB

The measurement errors increase even when there is a clear open path between the UWB pulse transmitter and receiver. This system requires careful calibration before use. The signal levels for the installed environment must be calibrated. A measurement of the background noise level is required, so signal below that threshold can be excluded. [3].

Digital Compass
There are two problems associated with the digital compass as heading sensor. First, the body orientation changes either during locomotion or while standing on uneven terrain. This produces the pitch and roll of the compass, making its read-out unreliable. Second, the earth field at the compass level may be disturbed by other electromagnetic fields or distorted by nearby ferrous materials. These deterministic interferences can be categorized in two types. First, hard iron effects are caused by magnetized objects, which are at a fixed position with respect to the compass. This relative closeness should be avoided. Second, soft iron effects are caused due to the distortion of the earth field by ferrous materials [10].

555

Informa ation and Computational Technology

Odometry

Odometry y errors fall int to two categories: systemati ic odometry errors e and non n-systematic er rrors. Usually internal system matic factors cause c a rise of f systematic er rrors, which sh how a biased characteristic. . In contrast, non-systemati tic errors are in ndependent sy ystematic featu ures and have e an unbiased characteristic [8].

Inertial Sens sors

The stochastic errors pr resent on inert tial sensors ca ause the subse equent numeri ical integration ns of the ts to exhibit an n ever increasi ing variance. That T is, when a gyro or acce elerometer ou utput is measurement numerically in ntegrated in a dead-reckonin ng navigator, the variance in n the resulting g position and d velocity outputs grow unbounded in n time [9]. Th his degradation n of measurem ment accuracy propagates in nto the navigation sol lution at rates dependent on n the integrity y of the compo onent sensors s, the algorithm ms employed, and the durati ion of the un-aided navigat tion [9]. 4. Proposed d Error Correction

UWB

The main problem with h UWB is that when there is s not a clear path between the t tag and sen nsors, it create es ts that are con nsidered to be outliers. Con nsequently, out tlying points have h to be rem moved by usin ng random point an outlier ana alysis. This ana alysis has to be in real time. Outlier Remova al In statistic cs, an outlier is s an observati ion that is num merically distant from the re est of a data se et. In our system, this is s caused by an n indirect path h between the sensors and tag located in the t wireless ro obot system. There is not a mathematica al definition th hat determines s what constit tute an outlier, , rather it is a subjective exercise due to t the variation of different samples. Man ny methods ar re used to dete ermine whether or not an observation is s an outlier. These T methods s are based on n the mean and d standard dev viation of the sample. The method used as a foundatio on for our out tlier removal is i based on Grubbs test for r outliers and Rosners Test t for Outliers. Kalman Filter The secon nd method use ed to correct U UWB reading is Kalman Fil lter. The Kalm man Filter was s introduced in n the early 1960 0s and since then t it has fou und widesprea ad use. The pu urpose of the discretetime d Kalman Filter r is to provide the t closed for rm recursive solution for th he estimation of o linear discre ete-time dynam mic systems. The Kalman Filter F has two o steps: the pre ediction step, wherein the next n state of th he system is predicted given n the previous measurement; m ; and the upda ate step, where e the current state s of the sy ystem is estima ated given the e measurement t at that time step. s The furth her study of th hese equations s is left to the reader.

Digital Comp mpass

In order to o compensate e the compass errors, the reg gression analy ysis is used to find a feasible e pattern in a defined test bed b area. This area must be located away from ferrous materials and relative closen ness to power r closes, which generate mag gnetic fluxes (B Beauregard, 20 006).

Wheel Enco oders

This resea arch analyzes non-systemati n c errors, those e that result fr rom the intera action with the e surface with h the wheels. The University of Michigan B Benchmark (U UMBmark) me ethod is empl loyed in the te estbed so that the robot is programmed p to o follow a pre e-programmed d 4x4 square path p and four spots for 90 degrees d turns. This has to be e completed 4 times in cloc ckwise directio on and five tim mes in counter r clockwise dir rection [3].

Gyroscope + Accelerome eter (Inertial l sensors)

The probl lem with this system s is that path deviation n at constant velocity v canno ot be correcte ed. The axis of f a gyroscope also a tends to drift d with time, giving rise to o errors. Inerti ial sensors allo ow a high rate e of computation of the robot configurations c s, but they are not sufficient t because erro ors are accumu ulated. A thod is require ed to reduce th hese sources of o errors. Here, the Kalman n Filter is used d to get rid of statistical met the notorious s gyroscope dr rift with the pr resented. The e gyro input is s a voltage me easure by the sensor s [12].

Where is the angle; is the ang gular velocity; is the sam mpling period; bias is the gyr ro bias in angular veloci ity; u is the gy yro output. To o convert stat te to measurem ment is usually y the easiest part,
5 556

26th Internat tional Symposiu um on Automati ion and Robotics s in Construction n (ISARC 200 09)

It is im mportant to ap pply these step ps in exact ord der to correct the t gyroscope e readings. 5 System Model 5.1 Principl le The ro obot has suffic cient informati ion and know wledge concern ning its enviro onment. The id dea is to gener rate a local pat th using the kn nown informa ation, then atte empt to locall ly, in relation to t the robot, avoid a any obstacles detected d by th he onboard sen nsors and retu urn to the orie entation relativ ve to its frame e.

Dead Rec ckoning (Rel lative frame) )

The displacement es stimates can b be in the form of changes in n Cartesian coo ordinates (i.e. x and y es) or, more ty ypically, in hea ading and speed or distance e. With suffici iently frequent t absolute pos sition coordinate updates, Dead D reckonin ngs linearly gr rowing positio on errors can be b contained within w pre-def fined bounds [10]. 5.2 Path Planning P (Absol lute Frame) The pa ath-planning problem p is usu ually defined as a follows, Given a robot and a a descripti ion of an environme ent, plan a pat th between tw wo specific loc cations. The pa ath must be co ollision-free (f feasible) and satisfy cert tain optimizat tion criteria. In other word ds, path planni ing is generati ing a collision n-free path in an a environme ent with obsta acles and optim mizing it with h respect to some criterion [11]. [ The robo ot has a short sensing ran nge compared d to the size o of the region of o interest. It radially r senses from its posit tion. Obstacle es can block the sensing in n some directions [11]. This study assum mes that the robot knows its coordinates and a n via UWB an nd the digital c compass. orientation 5.3 Positio on Algorithm (R Relative + Absol lute) Once the t first point is obtained in n the path, this s data is conve erted into an angle a and dist tance relative to t the robot frame. Whene ever the robot t reaches the target t point pr rovided by the e path plannin ng, its absolute e s calculated by y using UWB r readings and then t passed th hrough Kalma an estimation. Then the pos sition position is algorithm compares rela ative values of f the robot, su uch as distance e traveled and d angle provided with the e, with the pre-defined boun nds estimated from the UW WB readings. If f the robots relative r positio on gyroscope trajectory is not within the t pre-define ed bounds pro ovided by the absolute locat tion system, th hen the system m has he robot positi ion to the cen nter of the bou und. to reset th 6 Mobile e Robot Navi igation Contr rol Architectu ure Figure 2 shows the mobile m robots s navigation control archite ecture designed d in this resea arch.

Figure 2. M Mobile Robots s Navigation System S Archit tecture

557

Informa ation and Computational Technology

UWB

This system m uses a netw work of sensor rs installed at known k positio ons and a set of o tags located d in the mobil le robot. The co ommunication n between diff ferent element ts has two opt tions. The first option is usi ing an Ethernet netw work that basi ically connects s all different sensors using a Cat5 Ethern net cable to a router. The second option n is using a wi ireless interfac ce. This option n uses a set of f wireless brid dges connected d to each sensor and respectively to the t main route er.

Micro contro oller board

The board d used in this research r is on ne of shelf solu utions from Olimex O (Figu ure 3). This bo oard will be connected wit th the WiFi Latronix L modu ule on the mob bile robot plat tform later. This T micro con ntroller offers a uses low-cost effec ctive platform m to interface w with the senso or by using dif fferent interfac ces. The first interface i I2C interface that connects s with the digital compass. The second in nterface is an Analog to Dig gital Converte er (ADC).

Figure 3. M Micro controlle er board for in nertial sensors

Inertial Sens sors

The analog g signals comi ing from the a accelerometer r and gyroscop pe are fed into o the ADC inp put port from the PIC16F87 77A to be dec coded.

Digital Comp mpass

The comp pass used in th his robot contr rol is the CMP PS03 built by Devantech. It uses two methods of operation, wh hich can be easily integrated d with our mic crocontroller (PIC16F877A ( A). The first op ption is a PWM signal that t outputs a square wave. The second option o is an in ntegrated circu uit interface, or o I2C. This method is faster than the previous p one, a and allows a faster fa integratio on with the micro m controlle er.

WiFi Module le

This is a compact embe edded solution n that connect ts a UART port into its two o inputs and th he data can be e d and controll led over a net twork. Mobile robots contr rol module use es Port A for its easily accessed communicatio on. Port B is connected c to t the Micro con ntroller board. .

Mobile Robo bot

The robot t includes a ser ries of sensors on board such as infrared d, human detec ction, microph hone, and camera amon ng others. The one mainly u used for this pr roject is the wheel w encoders s that provides the data for the dead reck koning algorith hm and regression analysis solution. s Figu ure 4 summari izes the frame ework for this study describe ed previously. . 7. Experiments The UWB B system requi ires calibration n within the ab bsolute and re elative frames such that a co ommon point of origin (0,0) ) is established d. Using a To otal Station an absolute reference point is established. Each sensor has to be calib brated indepen ndently to this point of orig gin. Test # 1 UW WB- Average e + Kalman: A buffer arra ay, which has 5 slots for ea ach data X and Y, is created d. Data enters th he system eve ery 0.2 second ds. The data is s then put into o the 5 slot ar rray, an averag ge is calculated d, and the array y shifts to acc commodate incoming data a. Each calc culated averag ge is stored as a a temporar ry variable and fed f into the Kalman K filter. T The results are e shown in Fig gure 5. Test # 2 UW WB-MidValue e + Kalman : This method d is similar to the t previous, however, h the median value is chosen rather r than the mea an. The result can be observ ved in Figure 6.
5 558

tional Symposiu um on Automati ion and Robotics s in Construction n (ISARC 200 09) 26th Internat

UWB B Receiv ver

UWB Receiver

POSITION ERORR P BOUNDING BOX MOBILE ROBOT

VE RELATIV ABSO OLUTE Acccelerome eter
Position Compensation

U UWB Tran nsmitter Dig gital Com mpass

Wheel Encod ders Gyroscope

Orientation Compensation

UWB B Receiv ver

UWB Receiver

Figure 4. Framework F for r error compe ensation betwe een relative an nd absolute se ensors

F Figure 5. UW WB- Average + Kalman

Figur re 6. Test # 2 UWB-MidVal lue + Kalman n

559

Informa ation and Computational Technology Test # 3 UW WB- Outliers Removal + K Kalman: Whi ile similar to th he previous methods, m this test t includes th he calculation of f a threshold where outlie ers are remov ved based on n Outlier anal lysis described d in Ronsner rs method(Figur re 7). Three met thods tested sh how a signific cant improvem ment in UWB localization, and a Test # 3 provided p the best results.

Figure 7. UWB- Outli iers Removal + Kalman 8 Conclusion ns This paper r introduced an a on-going re esearch projec ct developing indoor i robot wireless navig gation leading to the implem mentation of a suitable local lization system m within indoo or constructio on environmen nts. A framework ha as been introd duced to impro ove the mobil le robots navi igation in con nstruction proc cesses. The proposed fram mework introd duces method ds to detect an nd correct erro ors from vario ous sensors inc cluding both internal and external e sensor rs. Each senso ors error attri ibutes are iden ntified through h intensive an nd extensive la ab tests. Initial re esults from th his research ha ave shown an improvement i t in real time for f the UWB localization l system. This improvement i was achieved d by combining g different sta atistical analysi is. Three diffe erent experiments were w conducte ed with results s supporting the t above conclusions. Further in nvestigation is being conduc cted to address additional se ensor related errors e based on o the framework wh hich has been n established in n this study. It t is expected to t lead to the improvement i t of the remaining sen nsors integrate ed within the s system. These e improvemen nts will lead a better, b more suitable s localization sy ystem that can n be applied to o various indo oor constructio on automation n applications s.

References
[1] Tao. Y, Hu. H H, Zhou. . H. (2005).In ntegration of Vision V and Inertial Sensors for Home-ba ased Rehabilit tation. IEEE E Internationa al Conference on Robotics and a Automation, Barcelona a, Spain InerV Vis the 2nd Workshop W on n Integration o of Vision and Inertial Senso ors,. [2] Barshan, , B. and Durra ant-Whyte, H.F. (1993). A An Inertial Nav vigation System m for a Mobil le Robot. Proceedi ings of the 1993 IEEE/RSJ J Internationa al Conference on Intelligent t Robotics and d Systems, Yokoham ma, Japan, July 26-30, pp. 2 2243-2248. [3] Borenste ein, J. and Fen ng, L. (1994). UMBmark & A Method fo or Measuring, , Comparing, and a Correctin ng Dead-rec ckoning Error rs in Mobile R Robots. Tech hnical Report, The University of Michigan n UM-MEAM M94-22, December. D [4] Bock, T and Kreupl, K. K (2004). Pr rocedure for th he implement tation of Auto onomous Mob bile robots on n struction site, Proceedings s. 21st Interna ational Symposium on Auto omation and Robotics R in the Cons Construc ction, Jeju, Ko orea, 304-309. .

5 560

26th International Symposium on Automation and Robotics in Construction (ISARC 2009) [5] Williams, C., Cho, Y., Youn, J., Jung, U. (2007) Wireless Sensor-driven Intelligent Navigation Method for Mobile Robot Applications in Construction, ASCE, 2007 International Workshop on "Computing in Civil Engineering, ASCE, Pittsburgh, PA, ASCE, 640-64. pages [6] Ringstad, G and Enge.P. (2008)"GPS and UWB for indoor navigation". Stanford University. <http://waas.stanford.edu/~wwu/papers/gps/PDF/guttormion01.pdf.> [7] Martinelli, A. and Siegwart, R. (2003). "Estimating the odometry error of a mobile robot during navigation. In European Conference on Mobile Robots, Warsaw, Poland, Sept. 4-6. [8] Moore, T. Hill, C., Hide C., Ochieng W.Y., Feng S., Aguado E., Ioannides, R., Cross P. and Lau L. (2007). End-to-End Testing of an Integrated Centimetric Positioning Test-Bed, Proceedings of ION GNSS 20th International Technical Meeting of the Satellite Division, Sept. 25-28.Fort Worth, TX, pp.1128-1138. [9] Wall. J. and Bevly, Y .(2008). Characterization of Inertial Sensor Measurements for Navigation Performance Analysis. <gavlab.auburn.edu/paper.php?id=22> (Dec. 12, 2008). [10] Beauregard, S.(2006). A Helmet-Mounted Pedestrian Dead Reckoning System, 4th International Forum on Applied Wearable Computing. [11] Sedighi, K., Ashenayi, K. Manikas, T.W. Wainwright, R., Heng-Ming Tai . "Autonomous local path planning for a mobile robot using a genetic algorithm." Evolutionary Computation, CEC2004.Volume: 2, pages: 1338- 1345. [12] T. Tong. "Kalman Filter Made Easy". http://openuav.astroplanes.com/library/ docs/writeup.pdf> (Dec. 10, 2008).

561