Link Quality Estimator for a Mobile Robot

Narek Pezeshkian, Joseph D. Neff and Abraham Hart

SPAWAR Systems Center, Pacific, CA 92152, San Diego, U.S.A.

narek.pezeshkian@navy.mil

Keywords: Link Quality, Video Quality, Estimator, Metric, Mobile Robot.

Abstract: Maintaining link connectivity between a mobile robot and its control station in a non-line-of-sight

environment is challenging. One solution is to use intermediate relay radios that the robot can carry and

deploy when and where needed to maintain the link. However, the precise placement locations for the

relays are not known ahead of time. Therefore, the deployment decision must be formulated online and the

relays deployed before the link with the control station breaks. A link-quality estimator is developed based

on video throughput and received signal strength indicator data. The estimator takes into account human

perception of video quality that is obtained via subjective testing by an operator. The data is used to train

the link-quality estimator, which issues an alert that can be used as a trigger for an automatic relay

deployment mechanism or to advise the operator to manually deploy relays before the link between the

robot and control station fails.

1 INTRODUCTION

Tactical mobile robots have been increasingly used

by the military over the past several years. This is

especially true for Explosive Ordnance Disposal

(EOD) teams that use robots to investigate and

neutralize Improvised Explosive Devices. These

robots are remotely controlled from the operator

control unit (OCU) using digital radios. The high

operating frequency of these radios requires a line-

of-sight (LOS) to the OCU, which is difficult to

maintain in urban environments. The link between

the robot and OCU can fail, usually rather quickly,

when operating beyond LOS due to multipath

interference and signal fading.

Controlling a robot via a tethered connection,

typically fiber-optic, eliminates the LOS problem

but introduces new ones. Tethered connections can

snag and break, limiting mobility. Advanced radio

systems that utilize sophisticated modulation

techniques and take advantage of MIMO antenna

technology thrive in multi-path environments and

can overcome the LOS limitations to a degree.

However, obstacles that severely block and attenuate

the signal can still be problematic.

The use of relays, on the other hand, adds an

unprecedented degree of freedom to where robots

can operate. Relays can entirely overcome severe

obstacle blockages so long as a LOS can be

maintained with adjacent radios in a chain of relays.

Determining the placement location of such relays is

critical. The focus of this paper is the formulation of

a link-quality (LQ) estimator, the output of which is

used either by the robot (automatically) or the

operator (command sent from the OCU) to release a

relay before the link breaks. Section 2 provides a

brief background of various relay systems designed

for tactical robots. Section 3 discusses the LQ

estimator design. Simulation results are outlined in

section 4, and section 5 concludes the paper.

2 BACKGROUND

The solution to address the LOS requirement

between a tactical robot and its OCU began in 2002

under the Autonomous Mobile Communications

Relay (AMCR) project (Nguyen et al., 2003). The

goal of the AMCR system was to provide extended

range and non-line-of-sight (NLOS) operational

capability for tactical robots. This was

accomplished through the use of dedicated mobile

relay robots (or mobile nodes) that followed the lead

robot in a convoy formation and automatically

stopped when needed to maintain the link. The

radios on-board the lead robot, mobile nodes, and

OCU formed a mesh network that allowed the

operator to teleoperate the lead robot based on video

Pezeshkian N., D. Neff J. and Hart A..

Link Quality Estimator for a Mobile Robot.

DOI: 10.5220/0004038400870094

In Proceedings of the 9th International Conference on Informatics in Control, Automation and Robotics (ICINCO-2012), pages 87-94

ISBN: 978-989-8565-22-8

 2012 SCITEPRESS (Science and Technology Publications, Lda.)

relayed to the OCU.

The mobile nodes must be set up in a specific

order in such a convoy since each mobile node is

programmed to follow the robot in front of it. In

addition, each mobile node is programmed to

monitor the received signal strength indicator (RSSI)

data of the node immediately behind it. For

example, the last mobile node in the convoy

monitors the RSSI of the OCU. The RSSI data,

which is used as a measure of the link quality, is

compared to a predetermined threshold, below

which the mobile node stops to maintain the link.

The AMCR solution proved to be very successful

and the commercial-off-the-shelf (COTS) 802.11b

radios and processor boards were extremely cost

effective. However, the AMCR system was a

research project and never designed for field use,

since the mobile nodes were expensive and

logistically impractical.

A more realistic solution was developed under

the Automatically Deployed Communication Relays

(ADCR) project (Pezeshkian et al., 2007). The

ADCR system shown in Figure 1 consists of a

Deployer and several Relay “Bricks”. The Deployer

carries the Relay Bricks and mounts onto a small

ground robot.

Figure 1: ADCR Deployer mounted on an iRobot PackBot

with one deployed and five stowed Relay Bricks.

The Deployer and the Relay Bricks each have the

same radio hardware and RSSI-based link-quality

estimator that is used by the AMCR system.

However, the only mobile node is the Deployer,

therefore, the link monitoring and the decision to

eject a Relay Brick is formulated by the Deployer

radio. Once a Relay Brick is ejected it self-rights and

extends the antenna. As the operator controls the

robot along its path more Relay Bricks are ejected as

needed to maintain the link.

The success of ADCR led to several licensing

agreements with commercial developers.

Subsequent projects led to additional developments

that improved upon the system. For example, the

redesigned Deployer of the Automatic Payload

Deployment System (APDS) (Pezeshkian et al.,

2010) allows a robot to carry and deploy not only

Relay Bricks but a wide range of other types of

payloads, such as leave-behind sensors and

containers. The container payloads can be used to

carry food, ammunition, medical kits, and anything

else that fits within. The Relay Bricks were also

redesigned to contain faster radios and an improved

antenna lift mechanism as shown in Figure 2.

Figure 2: APDS Deployer mounted on an iRobot PackBot.

Various payload types are shown around the robot.

The interest that was generated by numerous

publications and successful demonstrations of the

APDS and ADCR systems led the Naval EOD

Technology Division (NAVEODTECHDIV) to fund

the development of a robust radio repeater solution

for use by currently-fielded robotic vehicles. It was

necessary to deliver this solution quickly; therefore,

a simplified, albeit robust system based on APDS

technology was developed under the Manually

Deployed Communication Relays (MDCR) project.

The MDCR system omits the RSSI monitoring and

automatic-deployment capability of APDS, and

instead relies entirely on remote commands sent

from the OCU to deploy the Relays as the operator

sees fit. Although simple in design, the MDCR

system has been successfully field tested with plans

to mass produce additional units.

Although RSSI-based link monitoring has been

successful in the ADCR and APDS systems, it is not

an ideal solution, as will be explained in section 3.

Therefore, the goal of the LQ estimator outlined in

this paper is to provide a superior estimation method

that will assist the MDCR operator in placing relays,

and to also provide a trigger to automatically eject

relays for future ADCR systems.

ICINCO2012-9thInternationalConferenceonInformaticsinControl,AutomationandRobotics

3 QUANTIFYING LINK

In the MDCR system the Relays are deployed based

on operator command. Two factors play a role in

the Relay deployment decision-making process of

the operator: 1) prior knowledge of LOS loss – the

operator knows that controlling the robot around a

large obstruction will cause a loss of LOS so a Relay

is deployed before proceeding, and 2) video

degradation – as the distance between the robot and

OCU increases, even under LOS conditions, the

operator deploys a Relay when video quality

degrades.

Although these factors can be effective for

deploying Relays, in order to maintain the link

between the robot and the OCU, the operator for the

most part is guessing as to where to place the Relays

based on experience and intuition about the RF

environment. If the relaying system could provide

an indicator based on some sort of LQ estimator that

can warn of a failing link, however, the operator

would be in a much better position to optimize Relay

placement. This is important since the number of

Relays carried by a robot is limited and maximizing

the distance between the Relays translates into

maximizing the stand-off distance of the robot.

Furthermore, the LQ estimator can be used by a

relaying system (e.g., ADCR) to provide automatic

Relay deployment capability, effectively alleviating

the operator from the deployment task.

It is also important to keep in mind that the link

under consideration is between the robot and the

next-hop neighbor of the routing path leading back

to the OCU. This is, in fact, the only dynamic link

given that the only mobile node is the robot and all

other nodes (OCU and previously deployed Relays)

are static.

3.1 Link Quality

In this section a background on recent work on link

quality is given, followed by sections that describe

the proposed LQ metrics used by the LQ estimator.

3.1.1 LQ Background

A plethora of research on LQ estimation can be

found in the literature. Many schemes combine

multiple variables available from the physical and

link layers to form a more comprehensive and robust

LQ metric. Rondinone, Ansari, Riihijärvi, and

Mähönen (2008) propose multiplying the Packet

Reception Rate (PRR) of a link by the corresponding

mean RSSI value to obtain a new LQ indicator that

can be used by a network to select an optimal

routing path. Srinivasan, Kazandjieva, Jain, and

Levis (2008) combine PRR and channel burstiness

to estimate TCP throughput. Liu and Cerpa (2011)

combine RSSI, PRR, signal-to-noise ratio (SNR)

and the Link Quality Indicator (LQI) provided by

the CC2420 radio chip to provide a probability of

successfully delivering the next packet.

Yet combining variables is not the only

approach. Farkas, Hossmann, Ruf, and Plattner

(2006) propose using pattern matching to predict the

future behaviour of a link. Each node keeps a time

series record of the SNR with each of its links and

uses pattern matching to find the best match in an

attempt to estimate the future behaviour of the SNR.

Qin, He, and Voigt (2011) develop a new LQ

estimator, called the Spectrum Factor (SF), which is

derived from frequency-domain data.

3.1.2 LQ Data

An LQ estimator can be used by a routing protocol

in a mesh network to select optimal routing paths

(Liu et al., 2010 and Liu and Cerpa, 2011). The goal

of the LQ estimator for the MDCR system is

somewhat different: Develop an LQ estimator that is

suitable in predicting link failure such that a Relay

can be deployed before the link breaks.

The LQ estimators discussed in the previous

section are unsuitable for use given the stated goal.

Rondinone et al. (2008) suggest multiplying the

PRR of a link by the corresponding mean RSSI

value to help in selecting routing paths. Since there

is only one link under consideration (between robot

and next-hop neighbor along the routing path

leading to the OCU), this multiplication provides no

new information. Srinivasan et al. (2008) attempt to

estimate TCP throughput, which is unnecessary

since the video data of the robot uses UDP packets

and the throughput is readily available. Liu et al.

(2011) make use of SNR and LQI data that is

unavailable in the 802.11 radios used in the MDCR

system. Farkas et al. (2006) use pattern matching to

predict future behaviour of a link. This requires

some level of repetitive pattern to be present in the

collected data, which is highly unlikely given the

random movements of a teleoperated robot. Finally,

Qin et al. (2011) estimate LQ in the frequency

domain, which requires raw RF data that is not

easily obtainable from the MDCR radios.

The data selected for the development of the

proposed LQ estimator is UDP throughput (packets-

per-second) and RSSI, which are readily available

and ease integration of the estimator into the existing

mesh network software of the MDCR system. The

LinkQualityEstimatorforaMobileRobot

throughput data is also a direct indicator of video

quality – one of the key factors in the deployment

decision-making process of the operator. Video

quality, however, is subjective. A slightly choppy

video may be acceptable to one operator and

unacceptable to another. To quantify video quality,

an experiment was devised where an operator

controlled the robot along a predetermined path and

when the video quality, as judged by the operator,

began to degrade, the operator marked that point in

time. The marking method is simply a key press on

a test laptop that collects throughput and RSSI data

along with operator key presses, all synchronized in

time. There were two different key presses involved

in this experiment: The #2 key was pressed when

video quality began to degrade and the #3 key was

pressed when the link was completely lost. These

two moments in time are t

(failing) and t

(lost),

respectively. The link-failure period (t

) is simply

– t

. A sample of collected data and key presses is

shown in Figure 3. A simple moving average (MA)

process is applied to all data to smooth out

variations.

Figure 3: Example of video throughput (blue) and RSSI

data (green) received at the OCU from a PackBot using

the MDCR system. Solid line is the average (μ) of past

five samples of underlying (dotted) data. Left and right

black lines represent t

and t

, respectively.

Many such trial runs were performed under two

different environments, one more prone to multipath

than the other. In all test trials, clear trends are

observed in the throughput data during t

summarized as follows: 1) The throughput begins to

roll off sometimes gradually and sometimes

relatively sharp, and 2) the throughput variance

increases. The RSSI data, as expected, drops

gradually overtime as the robot moves away from

the OCU. Before t

, however, the throughput data

does not show any clear trend. The test trials show

that t

varies between 10 to 20 seconds, which

provides ample time to issue an alert. These trends

are exploited in the design of the LQ estimator.

3.1.3 RSSI as Early Warning

RSSI data has been proposed as a good link-quality

metric by Srinivasan and Levis (2006) but the

limitations of this statement must be understood. It

has been shown by Vlavianos, Law, Broustis,

Krishnamurthy, and Faloutsos (2008) that RSSI data

is measured at the lowest rate and cannot

characterize the LQ at high transmission rates.

Furthermore, RSSI is only measured from the packet

preamble; therefore, if an interfering signal happens

to prevent proper reception of the preamble, the

RSSI will simply not be recorded. If the interfering

signal happens to corrupt the packet after the

preamble has been received, then the RSSI will be

recorded as if there is no interferer. Hence, RSSI

data is unchanged even in the presence of an

interferer. The work of Judd, Wang, and Steenkiste

(2008) further supports this assessment.

Broadband noise, however, is a concern. If the

overall noise floor is raised due to external

broadband sources of noise, the overall SNR of

received packets will decrease. This means that

RSSI data can only be measured down to the raised

noise floor since packets received below the noise

level will be corrupted. Looking at Figure 3 it may

seem reasonable to threshold the RSSI at about 10,

below which the throughput data enters the region of

degraded video quality, t

. This approach may

work in the absence of broadband noise, but that

constraint cannot be guaranteed when operating in a

variety of environments.

The goal of the proposed LQ estimator is to

predict link failures so that a Relay can be deployed

before the link breaks. Preferably, some early

warning should be given to the operator by the

relaying system, followed by an imminent failure

alert so that the operator can deploy a Relay before

the link breaks. Interfering signals are not a major

concern given the operating environment, where the

overwhelming reason for link failure is due to signal

fading and loss of LOS. Broadband noise, however,

can exist. Given the limitations of RSSI, it is then

reasonable to use it only as a conservative early

warning system. Figure 4 shows the mean RSSI

value exactly at time t

, for all test trials. There are

clear variations but the overall range is low. A

conservative early warning of link failure can be

issued, for example, if the mean RSSI drops below a

threshold of 20. Selecting a high threshold leaves

quite a bit of margin should the noise floor increase

due to broadband noise.

100 120 140 160 180 200 220 240 260 280

Time (sec)

Throughput (packets per second)

100 120 140 160 180 200 220 240 260 280

RSSI

100 120 140 160 180 200 220 240 260 280

RSSI

ICINCO2012-9thInternationalConferenceonInformaticsinControl,AutomationandRobotics

Figure 4: Mean RSSI value at time t

for all test trials.

3.1.4 Link Quality Metrics

An accurate estimation of imminent link failure is

required to alert the operator of complete loss of

connectivity or trigger the deployment of a Relay

from an automated deployment system. Since RSSI

data does not accurately reflect the ability of a link

to successfully deliver packets as discussed in

section 3.1.2, throughput data is used instead.

The first trend of the throughput data is increased

variance σ during t

. At the same time, the mean μ

drops due to the second trend, the roll off. Since the

mean is high and variance low prior to t

and vice

versa during t

, it is reasonable then to use the ratio

of the two as a metric. This is inspired from the

Ricean K-factor (Greenstein, Michelson, and Erceg,

1999), which is used as a measure of signal fading.

The ratio here is given as κ = μ/σ and is the first LQ

metric, LQM

The second trend is the roll off. This is measured

by first taking N samples of throughput data then

calculating its intercept (x

) and slope (x

) using

linear regression. The assumption is that the N-

sample-long data is a straight line. Using a sliding

window, x

and x

are updated for each new sample.

The vector x = [x

] is the second LQ metric,

LQM

The trade-off between the false-alarm rate and

the miss rate is dependent on N. Low false-alarm

and low miss rates are desired. By setting N too

high, the data will be too smooth and the LQ

estimator slow to respond. This has the effect of

reducing the false-alarm rate due to reduced noise,

but increases the miss rate due to reduced response

time. In effect, the link is lost before the LQ

estimator has a chance to issue an alert. On the

other hand, setting N too low causes the data to be

too noisy, increasing the false-alarm rate, but

reducing the miss rate due to increased response

time. Since the cost of failing to issue an alert (a

miss) is much greater than alerting too soon (a false

alarm), the selection is biased towards reducing the

miss-rate by choosing N = 5.

3.1.5 Classifier

The keystrokes of the operator during the test trials

essentially label the collected data that are used to

train the LQ estimator. Half of the collected data is

used as training data and the other half as test data.

A labelling problem can be solved by classifiers.

Supervised training is used by two classifiers, one

for LQM

and the other for LQM

. Each classifier

finds the optimal decision boundary between two

different sets of labelled data: those marked before t

(signal OK) and those marked during t

(signal

failing). The hypothesis function for LQM

is given

by z

(θ

) = θ

+ θ

κ and for LQM

the hypothesis

function is z

(θ

) = θ

+ θ

. Both are

modelled as linear functions, which is a reasonable

assumption when looking at the data clusters in

Figures 5 and 6. The optimal parameter vector θ is

found by the classifier, which defines the decision

boundary that has values z(θ) ≥ 0 on one side and

z(θ) < 0 on the other.

The plot of the labelled κ values for all test trials

is shown in Figure 5. The plot of labelled x

and x

values for all test trials is shown in Figure 6. The

green circles represent values that take place before

and the red asterisks are data that take place

during t

. Using logistic regression, an optimal

decision boundary is generated, shown as the blue

line. All green circles above the line are hits (link

OK) and those below the line are false alarms (link

failing when in fact it is not). All red asterisks

below the line are hits (link failing) and those above

the line are misses (link failing but no alert issued).

It is clear from both figures that there is overlap

between the labelled data. Given the high cost of

misses, the decision boundary is biased so as to

reduce the number of misses.

Figure 6 supports the roll-off trend of the

throughput data. Looking at Figure 3, the flat part of

the throughput data roughly corresponds to 40

packets-per-second and since it is flat its slope is

about zero. This correlates to the green cluster seen

in Figure 6. As the throughput begins to fail during

, the packet rate drops, which corresponding to the

reduced x

(intercept) values. At the same time the

slope increases in the negative direction. This

corresponds to the red cluster in Figure 6. The

positive x

values are due to the variance of the

throughput data during t

that can cause the slope to

2 4 6 8 10 12 14 16 18 20 22

t tri

Mean RSSI value at t

LinkQualityEstimatorforaMobileRobot

Figure 5: Plot of labelled κ and decision boundary. Its log

is taken to improve computation of the boundary.

Figure 6: Plot of x

vs. x

. Both variables have been

scaled to reduce their range for improved computation of

the decision boundary.

go positive momentarily. Notice, however, very few

occurrences of positive slope and high intercept

values take place. The occurrences of high intercept

and high negative slope can be explained by sharp

roll-offs, where the throughput value is still

somewhat high but the slope is steep.

3.1.6 Link-Quality Estimator

The goal of the LQ estimator is to provide an early

warning of link failure (based on RSSI data) and a

more accurate imminent link-failure alert (based on

LQ metrics calculated from throughput data). These

metrics are somewhat noisy due to the selection of N

chosen to increase responsiveness (reduced miss

rate), and hence, sensitivity (increased false-alarm

rate). Each metric alone is not sufficient to provide

an accurate estimation, therefore they are combined.

The manner in which they are combined is

essentially an AND operation between the

hypothesis functions. This implies that both

hypothesis functions z

(θ

) and z

(θ

) must agree

that the link is failing, which occurs when both

(θ

) and z

(θ

) are less than zero. Furthermore, the

LQ estimator does not issue an alert unless both

(θ

) and z

(θ

) are less than zero for three

consecutive samples in a row. This eliminates

momentary glitches where both hypothesis functions

are below zero. Finally, the LQ estimator does not

start calculating the imminent link-failure alert until

a warning is issued when the mean RSSI data falls

below a conservative threshold. A simplified flow

chart for the LQ estimator is shown in Figure 7.

Figure 7: Simplified flow chart for the LQ estimator.

The flow chart does not show the additional steps

taken to deactivate the warning and alert indicators.

For example, instead of using a single threshold,

hysteresis can be added to the mean RSSI data

where falling below the lower threshold (e.g., robot

moving away from OCU) causes a warning to be

issued, which is removed when the mean RSSI

moves above the upper threshold (e.g., robot moving

back towards OCU). In a somewhat similar manner

the link-failure alert indicator can be removed. For

example, an issued alert can be removed if both

hypothesis functions agree that the signal is good,

say for five consecutive samples.

Figure 8 shows a sample of a test trial. The plot

shows that a warning is issued when the mean RSSI

falls below 21 (hysteresis enabled). Once the

warning has been issued, the LQ estimator begins

calculating the LQ metrics and testing the hypothesis

functions z

(θ

) and z

(θ

). A value of less than zero

indicates a hit, which is shown on the plot as a red

box for LQM

and a red diamond for LQM

. A link-

failure is indicated as a red ‘+’ sign when both

functions are less than zero at the same time

(LQM

AND

). An occurrence of this takes place at time

190 but no alert is issued. The LQ estimator issues

an alert when it observes three consecutive link-

0 50 100 150 200 250 300

-1

-0.5

0.5

1.5

2.5

log

(

)

< t

≤

< t

Decision boundary

-0.1 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9

-4

-3

-2

-1

Intercept (x

)

Slope (x

)

< t

≤

< t

Decision boundary

ICINCO2012-9thInternationalConferenceonInformaticsinControl,AutomationandRobotics

failure hits at time t

= 202. This occurs just after t

= 200, with plenty of time still left before the link is

completely lost at time t

= 213.

Figure 8: LQ estimator warning and alert. A warning is

issued based on RSSI and an alert based on throughput.

4 SIMULATION RESULTS

The LQ estimator will occasionally issue an alert

prior to t

due to the overlap in the training data as

shown in Figures 5 and 6. This is a desirable effect

because the alert is issued just before video

degradation begins. An alert issued after t

is also

acceptable so long as the alert does not take place

too close to t

, which may not provide enough time

to deploy a Relay before the link breaks. Therefore,

the accuracy of the LQ estimator is defined as the

percentage of alerts issued within a specified

window of time t

centered on t

for all test data.

The window t

is defined as t

± Δt. The value Δt is

equal to βt

where 0 < β ≤ 1. This ensures that Δt is

no greater than t

. The selection of β is somewhat

arbitrary. The smaller it is, the closer the alert issue-

time t

must be to t

before the alert is counted as an

accurate hit. Table 1 shows the accuracy result for

different values of β.

Table 1: LQ estimator accuracy.

Hit % False Alarm % Miss %

1/2 73 9 18

2/3 82 9 9

3/4 91 0 9

1 100 0 0

Table 1 shows that with β = 1 all alerts are issued

within t

± t

, and 73% of alerts are issued within t

± t

/2 with β = 1/2.

5 CONCLUSIONS

A link-quality (LQ) estimator is developed to

provide an accurate means of estimating an

imminent link failure, which is required to assist the

operator of a tactical mobile robot in deploying a

Relay before the link breaks. Since the robot carries

a limited number of Relays, increasing the distance

between deployment locations will increase the

operational range of the robot. The same LQ

estimator can also be used on an automatic Relay

deployment mechanism (such as the ADCR system)

as a trigger to eject a Relay.

The LQ estimator is based on LQ metrics

calculated from labelled throughput data. The

throughput data is labelled during test trials by the

robot operator, who marks the data when the video

quality begins to degrade and finally lost altogether.

This process is repeated for several trial runs in two

different operating environments. The labelled data

is used to train the LQ estimator, which is then

applied to test data that is not used in the training

session. The LQ estimator issues two alerts: 1) a

warning alert to the operator based on RSSI data,

which serves as a conservative estimate of a link

beginning to fail, and 2) a much more accurate link-

failure alert based on throughput data when an

imminent link failure is detected. The results from

the test data show that the LQ estimator achieves

high accuracy in issuing an alert before the link is

completely lost.

REFERENCES

Farkas, K., Hossmann, T., Ruf, L., Plattner, B., 2006.

Pattern Matching Based Link Quality Prediction in

Wireless Mobile Ad Hoc Networks. MSWiM’06.

Greenstein, L. J., Michelson, D. G., Erceg, V., 1999.

Moment-Method Estimation of the Ricean K-Factor.

IEEE Communications Letter, 175-176.

Judd, G., Wang, X., Steenkiste, P. 2008. Efficient

Channel-Aware Rate Adaptation in Dynamic

Environments. MobiSys’08.

Liu, T., Cerpa, A.E., 2011. Foresee (4C): Wireless Link

Prediction using Link Features. 10th International

Conference on Information Processing in Sensor

Networks, 294-305.

Liu, L., Fan, Y., Shu, J., Yu, K., 2010. A Link Quality

Prediction Mechanism for WSNs Based on Time

Series Model. Ubiquitous Intelligence & Computing

and 7th International Conference on Autonomic &

Trusted Computing, 175-179.

Nguyen, H. G., Pezeshkian, N., Raymond, M., Gupta, A.,

Spector J. M., 2003. Autonomous Communication

100 120 140 160 180 200 220

Throughput (packets per second)

Time (sec)

100 120 140 160 180 200 220

RSSI

Warning

Alert

Throughput

RSSI

LQM

AND

LinkQualityEstimatorforaMobileRobot

Relays for Tactical Robots. 11th International

Conference on Advanced Robotics, 35-40.

Pezeshkian, N., Nguyen, H. G., Burmeister, A., 2007.

Unmanned Ground Vehicle Radio Relay Deployment

System for Non-Line-of-Sight Operations. 13th

International Conference on Robotics & Applications.

Pezeshkian, N., Nguyen, H. G., Burmeister, A., Holz, K.,

Hart, A., 2010. A Modular Design Approach for the

Automatic Payload Deployment System, Association

for Unmanned Vehicle Systems International.

Qin, Y., He, Z., Voigt, T., 2011. Towards Accurate and

Agile Link Quality Estimation in Wireless Sensor

Networks. Ad Hoc Networking Workshop, 2011 the

10th IFIP Annual Mediterranean, 179-185.

Rondinone, M., Ansari, J., Riihijärvi, J., Mähönen, P.,

2008. Designing a Reliable and Stable Link Quality

Metric for Wireless Seonsor Networks. REALWSN’08.

Srinivasan, K., Kazandjieva, M.A., Jain, M., Levis, P.

2008. PRR Is Not Enough.

Srinivasan, K., Levis, P., 2006. RSSI Is Under

Appreciated. Third Workshop on Embedded

Networked Sensors.

Vlavianos, A., Law, L.K., Broustis, I., Krishnamurthy, S.V.,

Faloutsos, M., 2008. Assessing Link Quality in IEEE

802.11 Wireless Networks: Which is the Right Metric?

IEEE 19th International Symposium on Personal,

Indoor and Mobile Radio Communications, 1-6.

ICINCO2012-9thInternationalConferenceonInformaticsinControl,AutomationandRobotics