An Autonomous Mobile Inspection Robot for an Electric Power

Sub-station

Simon Thompson

, Satoshi Kagami

and Masafumi Okajima

Digital Human Research Center, National Institute of Advanced Industrial Science and Technology, Tokyo, Japan

Research and Development Department, Kansai Electric Power Company, Inc., Osaka, Japan

Keywords:

Autonomous Navigation, Mobile Robot, Localisation, Inspection.

Abstract:

In this work, we describe the development of an outdoor, autonomous mobile robot that performs inspections

of various facilities within an electric power sub-station. A segway-based robot was developed that can per-

form autonomous navigation along a given set of waypoints and perform inspection tasks (taking photographs

at set locations). A retractable leg system was developed to allow the robot to enter/exit self-balancing mode

and achieve a stable rest position from which to perform inspection tasks. The robot platform, localisation and

control systems, and the inspection process are described, and a real world experiment consisting of navigation

over a 1km path with 5 inspection points is reported. All inspection tasks were completed to the satisfaction

of plant operators.

1 INTRODUCTION

This paper describes the development of an outdoor

autonomous mobile robot platform that performs in-

spections of various facilities within an electric power

substation. Speciﬁcally the robot must navigate along

a given path, deﬁned as a series of waypoints, stop-

ping at certain waypoints to take photographs of

plant facilities using a zoom-able pan/tilt camera.

The environment of the electric power substation is

large, 400x400 meters, complex, containing irregular

shaped machinery, and having areas of rough and/or

uneven terrain.

The main requirement for such a system, is an

autonomous platform capable of physically moving

through the sub-station environment while being able

to perform localisation with sufﬁcient accuracy to

navigate through the plant and to perform the inspec-

tion task. Accurate localisation at inspection points

reduces the need for image processing to identify tar-

get locations which could prove difﬁcult in outdoor

environments. A further requirement was the localisa-

tion system should not use GPS, as experiments with

hand-held GPS units within the sub-station proved

unreliable, perhaps due to the presence of powerful

electro-magnetic ﬁelds. To allow ﬂexibility in assign-

ing the inspection route, it is also desirable to elimi-

nate the need for artiﬁcial landmarks.

Typical approaches to mobile robot localisation

use bayesian techniques to estimated a robot’s lo-

cation within a geometric map. Such systems usu-

ally use laser range ﬁnder range measurements and

wheel based odometry information to localise the

robot within 2D occupancy grids (Thrun et al., 2001).

The localisation process then becoming an iterative

process of predicting motion from the robot’s odom-

etry measurements and evaluating the predicted esti-

mates by matching the current sensor data with the

expected sensor data calculated from the map.

Moving from indoor robots to outdoor robots, be-

cause of the relatively unstructured nature of the envi-

ronment, the localisation problem typically changes

from 3DOF (2D position and azimuth orientation)

in 2D maps, to 6DOF (3D position and roll, pitch,

yaw) in 3D maps. This increase in the dimensions

of the state space to be estimated (and in the en-

vironment map) results in extra computational re-

quirements compared to indoor robots. With outdoor

robots typically moving faster and being larger, this

presents some problem in terms of safe and accurate

motion control based on feedback from the localisa-

tion process.

This increase in computation for outdoor localisa-

tion has led to introduction of additional localisation

cues such as GPS signals and artiﬁcial landmarks in

order to reduce the need for matching between sen-

sor data and large, complex environment maps (Thrun

et al., 2006) (Ohno et al., 2004) (Moralez and Tsub-

300

Thompson S., Kagami S. and Okajima M..

An Autonomous Mobile Inspection Robot for an Electric Power Sub-station.

DOI: 10.5220/0004480503000306

In Proceedings of the 10th International Conference on Informatics in Control, Automation and Robotics (ICINCO-2013), pages 300-306

ISBN: 978-989-8565-71-6

 2013 SCITEPRESS (Science and Technology Publications, Lda.)

ouchi, 2007). This can be a problem, however in

environments where GPS signals are unavailable or

unreliable, or when the installation of artiﬁcial land-

marks within the environment is unacceptable or im-

practical. In such cases, systems typically revert back

to 3DOF localisation as in (Madhavan and Durrant-

Whyte, 2004), who report a laser range ﬁnder based

3DOF localisation system for ground vehicles using

full laser scans (with 1m error) and landmarks (0.5m

error). Likewise, (Levinson et al., 2003) use the re-

ﬂectivity of road surfaces to build image like maps

of urban environments and localise in 3DOF by cor-

recting GPS readings with map based matching of LI-

DAR range and reﬂectivity data, and report an impres-

sive 10cm localisation accuracy. Without GPS they

report a maximum error (measured at 50m intervals -

presumably while stopped) of 35cm. This approach,

however requires regular lane markings. A similar

system based on visual sensing of lane markings is

reported by (Xia et al., 2006) but can suffer from oc-

clusions due to a limited ﬁeld of view.

As for mobile inspection robots, the majority of

reported systems are for inspections of hard to ac-

cess environments such as pipe inspection (Suzumori

et al., 1999) or for rescue activities (Osuka, 2003),

with research focused on speciﬁc mobility or sensing

environments required by such environments. (Davi-

son and Kita, 2003) report a pair of cooperating in-

door inspection robots with the application of inspect-

ing the outside of pipes in a nuclear power plant.

Stereo vision and sparse feature maps are used to ac-

curately localise the robots in small scale environ-

ments.

(Thompson et al., 2010) (Thompson et al., 2011b)

report the development of a 6DOF localisation sys-

tem that can perform accurate localisation within

dense 3D polygon maps of large outdoor environ-

ments without the use of GPS. In this work we use this

localisation system to develop a mobile robot plat-

form capable of performing an autonomous inspec-

tion task as described above.

Below, the components of the autonomous inspec-

tion robot are described, including the robot platform,

the retractable legs, the 6DOF localisation system, the

method used for motion control, and the implementa-

tion of the inspection task. Also a real world inspec-

tion task in a power sub-station consisting of 5 inspec-

tion points along a 1km path is reported. The main

contribution of the paper is the successful demonstra-

tion of an autonomous outdoor inspection task over

a wide area within an operational electric power sub-

station.

Figure 1: The inspection robot, based on a Segway RMP

base.

2 AUTONOMOUS ROBOT

SYSTEM

2.1 Robot Platform

The robot platform used is the Segway RMP 200/ATV

(Nguyen et al., 2004) which is an inverted-pendulum

type, self-balancing two-wheeled mobile robot. The

RMP was chosen as it provides a reliable base that can

operate outdoors, and has the power to traverse rough

terrain and the maneuverability to turn on the spot.

This base has been modiﬁed to mount front and back

facing Sick LMS laser range ﬁnders, and a bumper

ring. A box computer running ART Linux performs

motion control and communicates with the RMP

base, while a laptop computer (Linux, Core 2 Duo

2.26Ghz CPU, nVidia Quadro FX 770M graphics

card) runs the localisation software. ( Figure 1 shows

the robot platform. The robot also has retractable

legs which it uses for stability when not moving,

and which allows it to start/stop self-balancing mode

without human intervention. For inspection purposes

the robot is equipped with a zoom-able pan/tilt cam-

era. A 3D spherical microphone array for detecting

irregular sounds within the sub-station is also shown,

but is outside the scope of this paper (for details see

(Sasaki et al., 2012)). Stacks of different colour lights

are used to communicate the robot state to nearby hu-

mans for safety purposes.

2.2 Retractable Leg System

The Segway RMP base is a self balancing wheeled

inverted pendulum type of robot and typically must

maintain some degree of motion to stand upright.

AnAutonomousMobileInspectionRobotforanElectricPowerSub-station

301

Figure 2: Active trapezoidal linked legs in a) retracted po-

sition, and b) extended position.

This complicates any inspection task requiring a sta-

ble base, such as taking a picture from a camera

mounted on the robot platform. To overcome this,

a pair of retractable legs was developed (Kabasawa

et al., 2012), to enable the robot to exit self balanc-

ing mode and assume a stable position. Figure 2

shows the legs mounted on the robot base in a) the

retracted (or ”up”) position, and b) the extended (or

”down”) position. The legs are two pairs of active

trapezoidal linked legs which also allows the robot to

stop on inclined surfaces, or with an inclined body.

From a starting position, legs down, the system initi-

ates leg retraction an immediately switches the Seg-

way control into self balancing mode. To stop, legs

are extended and when they touch the ground the self-

balancing control mode is switched off.

2.3 Dense 3D Polygon Maps

In order to localise accurately within a large, com-

plex environment such as a power sub-station, we use

dense 3D polygon maps constructed from 3D laser

Figure 3: Part of the 3D polygon map of the power sub-

station.

scans as reported in (Kagami et al., 2010). Range

data from each scan location are associated with ad-

jacent scans, and the transformation matrices were

solved. Aligned points were then converted into dense

3D polygon maps. Such realistic 3D maps allow for

knowledge of traversable terrain, as well as providing

ﬂexibility in choice of navigation route, as most of the

structure of the environment is captured. In this study,

a map covering approximately 400 × 400m was con-

structed which contained 2, 018,418 polygons. Part

of the map is shown in Figure 3, ground surface is

shown in blue, while the grey shows various struc-

tures within the sub-station. Such complex structures,

interspersed with wide open spaces are typical of the

plant environment. This image shows approximately

120 × 80m of ground area.

2.4 6DOF Localisation

In this study a particle ﬁlter framework is adopted to

perform 6DOF localisation within the dense 3D poly-

gon maps, using odometry information from the Seg-

way RMP base, and two robot mounted laser range

ﬁnders (front and back facing). In such an approach,

a set of state estimates, or particles, are evaluated for

the likelihood that the states they represent would pro-

duce the current laser range sensor data, by comparing

it with expected range data generated using the map.

For 3DOF localisation this is relatively easy, as a 2D

map can be used. For 6DOF localisation the extra di-

mensionality of the state space, and the extension of

the map to 3D mean the computational requirements

for localisation can become unmanageable for the real

time constraints of autonomous navigation.

To perform accurate, efﬁcient localisation in dense

polygon maps, several techniques were developed to

reduce the computational requirements involved in

matching sensor data with expected map views in the

particle ﬁlter based localisation process.:

• Constrained Motion Model: use of vehicle dy-

namics and dense 3D polygon map to limit disper-

ICINCO2013-10thInternationalConferenceonInformaticsinControl,AutomationandRobotics

302

sion of localisation belief distribution during mo-

tion (Thompson et al., 2010). Knowing the sur-

face of the current location, height pitch and roll

dispersion can be bounded.

• Indexed Polygon Map: simple grid index and

polygon classiﬁcation scheme to limit number of

polygons used for to generate expected sensor

views within a given map area (Thompson et al.,

2011a)

• Selection of Polygon Sets: precompute the best

sub-set of polygons to use at each index grid based

on likely frequency of observation, angle of inci-

dence and angle of elevation of the observation

ray (Thompson et al., 2011b). This further re-

duces the number of polygons to be used for lo-

calisation within a given map area.

• GPU Based Expected View Generation: use spe-

cialised graphics hardware to reduce cost of gen-

erated expected sensor views from polygon maps.

Using these techniques, fast, accurate localisation

can be performed in large scale, dense 3D poly-

gon maps. Typical localisation results are an aver-

age of 10cm localisation error with a cycle rate of

approximately 10Hz. Experiments within the sub-

station environment have produced average localisa-

tion error of 3cm over simple paths that travel close

(within 10m) to plant machinery and thus have strong

localisation cues. For details of experimental re-

sults on localisation accuracy see (Thompson et al.,

2011a)(Thompson et al., 2011b). Localisation perfor-

mance was not signiﬁcantly affected by observed en-

vironmental changes such as sparse pedestrians, oc-

casional passing vehicles, addition/removal of stored

equipment, and even the addition of demountable

buildings.

One problem with the localisation system that was

observed in development, was failure due to tempo-

rary loss of odometry information. In certain areas

of the sub-station, the robot pass through areas with

strong electro-magnetic ﬁelds which caused the USB

connection from the RMP base to the navigation con-

trol system to disconnect, resulting in loss of odom-

etry information while the robot comes to a stop. A

simple system to monitor the connection, and in the

case of disconnection to ”fake” approximate odome-

try (with appropriate uncertainty) of the stopping mo-

tion while attempting reconnection, was developed.

2.5 Path Following Control System

To perform autonomous navigation the robot system

employs a path following method known as the pure

pursuit algorithm (Heredia and Ollero, 2007). The

path is deﬁned by a series of way points, and line seg-

ments connecting the way points form the reference

path. The pure pursuit algorithm attempts to follow

the path by computing a circular arc from the esti-

mated robot position to a point some distance along

the reference path. This ”look ahead” distance varies

depending on current velocity and whether a way-

point or goal location is near. In this work, no ob-

stacle avoidance is implemented as it is a controlled

environment and the robot should not deviate unex-

pectedly from its path. If an object is detected in its

paths, the robot simply stops and waits until the path

is clear.

3 INSPECTION TASK

The inspection task involves taking a photograph

of various facilities for later viewing by plant per-

sonel. Each inspection consists of the following sub-

behaviours:

1. navigating to a predeﬁned point within the sub-

station

2. turning to a desired orientation if required

3. lowering legs for stability

4. camera performing pan/tilt and zoom as required

5. taking the photograph

6. raise legs

7. proceed to next way-point.

In order to give a set of inspection tasks to the robot,

a high level command language was deﬁned:

• START x y: deﬁnes a start location by x and y

coordinates within the given map.

• GOTO x y: deﬁne a navigation way-point within

the given map

• TURN theta: turn the robot to a given global ori-

entation.

• INSPECT id x y z zoom: deﬁnes an inspection task

in terms of an identiﬁcation number, a three di-

mensional location, and a camera zoom parame-

ter.

For any speciﬁc inspection task, the robot performs

the following sequence of commands (with appropri-

ate parameters): GOTO, TURN, INSPECT. First the

robot navigates to the given way-point in the GOTO

command parameters using the localisation and con-

trol systems described above. Then the robot turns

to the given orientation theta, again using the lo-

calisation estimate for feedback. The legs are then

lowered, and the robot-self-balancing mode disabled.

AnAutonomousMobileInspectionRobotforanElectricPowerSub-station

303

a) b)

c) d)

e) f)

Figure 4: An example of the inspection procedure: a) robot

moves towards the inspection point, b) stops at point, c)

turns to required orientation, d) put s legs down to stabi-

lize platform, e) activates pan-tilt camera to take inspection

photograph, f) retracts legs and continues to next way-point.

Once the robot is motionless and stable, the current

6DOF localisation estimate and the 3D location given

in the x,y,z parameters to the INSPECT command

(deﬁned in relation to global map coordinates), are

used to calculate a pan/tilt conﬁguration which orien-

tates the camera towards the inspection point. Once

correctly orientated, the camera’s zoom control is ad-

justed according to the zoom parameter, and the pic-

ture is taken.

4 EXPERIMENTAL RESULTS

Real world experiments demonstrating autonomous

inspection were performed at an operational electric

power sub-staion operated by Kansai Electric Power

Company (Japan).

Figure 5: An example of an inspection photograph taken

by robot, images show view from pan-tilt camera: a) view

when legs are lowered, b) pan to required orientation, c) tilt

to required elevation, d) zooming to target, e) photograph

taken for inspection task.

4.1 Single Inspection Task

Figure 4 shows the robot during an inspection task,

sub-ﬁgure a) robot moves towards the inspection

point, b) stops at point, c) turns to required orienta-

tion, d) put s legs down to stabilize platform, e) acti-

vates pan-tilt camera to take inspection photograph, f)

retracts legs and continues to next way-point. Figure 5

shows images from the pan-tilt camera during the in-

spection task. Sub-ﬁgure a) shows the view after the

ICINCO2013-10thInternationalConferenceonInformaticsinControl,AutomationandRobotics

304

Figure 6: Inspection task within polygon map. Naviga-

tion path (red squares (waypoints) and red lines (connecting

paths), and inspection points (blue squares) labelled one to

ﬁve.

Figure 7: The estimated path traveled by the robot during

the inspection task.

robot has, stopped turned to the required orientation

and lowered it’s legs; b) after the camera has per-

formed a pan motion; c) after tilt motion; d) in the

process of zooming; e) the photo taken for the in-

spection task. In this example the inspection point

was over 15m from the robot stopping position, and a

gauge of approximately 15cm was photographed us-

ing just the robot’s estimated position for guidance.

4.2 Inspection of an Electric Power

Sub-station

The inspection task required the robot to navigate

over a 1km path throughout a power sub-station

within one hour, performing inspection tasks at ﬁve

separate places within the environment (Fig. 6). In-

spection points are things like transformers, gauges

and latches. The inspection points were manually de-

ﬁned by operators, with the 3D coordinates of the

points being recovered from the polygon map. The

robot traveled over the path with a maximum speed

of 0.6m/s The inspection task was completed within

23 minutes, with all ﬁve inspection points captured

to the satisfaction of plant operators. The estimated

path the robot traveled is shown in Figure 7. Average

localisation error was approximately 10cm.

5 CONCLUSIONS

In this paper we have presented an autonomous mo-

bile robot capable of performing an inspection task

within the grounds of an electric power sub-station.

Fast, accurate 6DOF localisation, using onboard-

sensors only, in conjunction with a retractable leg

system which stabilizes the robot base, allowed the

mobile robot to take precise photographs of inspec-

tion targets. A high level command language was de-

scribed, and a successful, real-world inspection task

was reported.

Although in this study, the inspection task was

completed satisfactorily with just the use of the es-

timated robot position, ideally the localisation result

would provide an initial estimate for an image pro-

cessing system to perform a local search for the in-

spection target, and subsequently to give autonomous

conﬁrmation that the task was achieved. Also in this

case, the robot did not perform any obstacle avoid-

ance. Subsequent work aims at 3D sensing and avoid-

ance of potential obstacles, although this should oc-

cur in a predictable, safe and operator approved man-

ner. Further study is also needed to evaluate long term

autonomous operation in terms of inspection success

rate and localisation performance in various weather

conditions.

REFERENCES

Davison, A. and Kita, N. (2003). Active visual localisation

for cooperating inspection robots. In IEEE/RSJ Inter-

national Conference of Intelligent Robots and Systems

(IROS).

Heredia, G. and Ollero, A. (2007). Stability of autonomous

vehicle path tracking with pure delays in the control

loop. In Advanced Robotics, Vol 21, No. 1-2.

Kabasawa, M., Thompson, S., Kagami, S., and Okajima, M.

(2012). Development of trapezoidal linked legs for a

wheeled inverted pendulum. In JSME Robotics and

Mechatronics Conference.

Kagami, S., Hanai, R., Hatao, N., and Inaba, M. (2010).

Outdoor 3d map generation based on planar feature

AnAutonomousMobileInspectionRobotforanElectricPowerSub-station

305

for autonomous vehicle navigation in urban environ-

ment. In Int. Conf. on Intelligent Robots and Systems

(IROS).

Levinson, J., Montemerlo, M., and Thrun, S. (2003). Map-

based precision vehicle localisation in urban environ-

ments. In Proceedings of Robots, Science and Sys-

tems.

Madhavan, R. and Durrant-Whyte, H. F. (2004). Terrain-

aided localization of autonomous ground vehicles. In

Automation in Construction.

Moralez, Y. and Tsubouchi, T. (2007). Gps moving perfor-

mance on open sky and forested paths. In Proc. of the

Int. Conf. on Intelligent Robots and Systems.

Nguyen, H. G., Morrell, J., Mullens, K., Burmeister, A.,

Miles, S., Farrington, N., Thomas, K., and Gage, D.

(2004). Segway robotic mobility platform. In SPIE

Proc, 5609: Mobile Robots XVII,.

Ohno, K., Tsubouchi, T., Shigematsu, B., and Yuta, S.

(2004). Proper use of gps for outdoor navigation by

an autonomous mobile robot. In Proceedings of Int.

Conf. on Intelligent Autonomous Systems.

Osuka, K. (2003). Development of mobile inspection robot

for rescue activities: Moira. In IEEE/RSJ Interna-

tional Conference of Intelligent Robots and Systems

(IROS).

Sasaki, Y., Kabasawa, M., Thompson, S., Kagami, S., and

Oro, K. (2012). Spherical microphone array for spatial

sound localisation for a mobile robot. In Proc. of the

Int. Conf. on Intelligent Robots and Systems.

Suzumori, K., Miyagawa, T., Kimura, M., and Hasegawa,

Y. (1999). Micro inspection robot for 1-in pipes. In

IEEE/ASME Transactions on Mechatronics, Vol. 4,

No. 3.

Thompson, S., Kagami, S., and M.Okajima (2011a). Facet

classiﬁcation in 3d polygon maps for autonomous ve-

hicle localisation. In Proc. of the Int. Conf. on Intelli-

gent Unmanned Systems.

Thompson, S., Kagami, S., and M.Okajima (2011b). Se-

lection of polygon sets for 6dof localisation of au-

tonomous vehicles. In Proc. of the Int. Conf. on Sys-

tems, Man and Cybernetics.

Thompson, S., Kagami, S., and Okajima, M. (2010). Con-

strained 6dof localisation for autonomous vehicles. In

Proc. of the Int. Conf. on Systems, Man and Cybernet-

ics,.

Thrun, S., Burgard, W., and Fox, D. (2001). Robust monte

carlo localisation for mobile robots. In Artiﬁcial Intel-

ligence, vol. 128.

Thrun, S., Montemerlo, M., Dahlkamp, H., Stavens, D.,

Aron, A., Diebel, J., Fong, P., Gale, J., Halpenny,

M., Hoffmann, G., Lau, K., Oakley, C., Palatucci, M.,

Pratt, V., Stang, P., Strohband, S., Dupont, C., Jen-

drossek, L., Koelen, C., Markey, C., Rummel, C., van

Niekerk, J., Jensen, E., Alessandrini, P., Bradski, G.,

Davies, B., Ettinger, S., Kaehler, A., Neﬁan, A., , and

Mahoney, P. (2006). Stanley: The robot that won the

darpa grand challenge. In Journal of Field Robotics,

Vol. 23, No. 9, June.

Xia, T. K., Yang, M., and Yang, R. Q. (2006). Vision based

global localization for intelligent vehicles. In Pro-

ceedings of the Intelligent Vehicles Symposium.

ICINCO2013-10thInternationalConferenceonInformaticsinControl,AutomationandRobotics

306