Oceans ‘06, Singapore(authors’ manuscript – do not distribute)

Effective Communication in Schools of Submersibles

Felix Schill

��

& Uwe R. Zimmer

�

�

Research School of Information Sciences and EngineeringAutonomous Underwater Robotics Research Group

The Australian National University, ACT 0200, Australia

�

National ICT Australia Ltd., Locked Bag 8001, Canberra ACT 2601, Australia

1

felix.schill@anu.edu.au | uwe.zimmer@ieee.org

Effective communication mechanisms in schools of sub-mersibles are a key requirement for their meaningful de-ployment. Furthermore a fully distributed communicationschema is preferable for reasons of reliability. The requiredcommunication form is usually many-to-many, especiallyomnicast [4] (or ‘gossiping’). All these constraints arehard to achieve at the same time in a low-bandwidth, shortrange communication setup. The theoretical findings in[4] and [5] are expanded and employed for the actual un-derwater schools communication in the Serafina project(e.g. [2], [3]). The results are reported here.

1. Introduction

Short range, long-wave radio and optical communi-cation links are deployed to build up and maintain adynamical communication schedule which allowsfor a highly efficient exchange of all locally sampledinformation to all members of the school.

Short range, long-wave radio and optical communi-cation links are deployed to form a multi-hop radionetwork in a school of submersibles. The nodes ofthis network build up and maintain a dynamicalcommunication schedule which allows for a highlyefficient exchange of all locally sampled informationto all members of the school. Wireless communica-tion under water, especially in seawater, is very limit-ed. Commonly available high-frequency radio linkssuch as 802.11b or BlueTooth encounter extremelyhigh attenuation, which makes them unusable. Threemethods of wireless communication are known forunder water setups: acoustic (sonar modems), opti-cal (using blue/green light, refer to [6]), and long-wave radio. In this paper we will focus on electro-magnetic communication channels only. Theexperimental section of this paper has technical de-

tails about the transmitter and receiver modulesused for the experiments. A common problem of allknown wireless channels under water is that theirbandwidth and range are very limited. Due to thebandwidth limitations, these links are usually onlyhalf-duplex, and have only one single channel. Theseconstraints favour Time Division Multiple Access(TDMA) protocols, which efficiently assign timeslots for sending to nodes without wasting valuablebandwidth.Being able to efficiently distribute and share locallyavailable data from all nodes with the entire networkis a problem rarely discussed in classical communi-cation theory. Many network applications requirededicated links between designated nodes, or have ahierarchical structure. However, in a school of identi-cal submersible robots, there is no hierarchy. Also, lo-cal sensing only delivers sparse information aboutthe environment, especially in an underwater setup.It is therefore crucial that submersibles share their lo-cal information with their direct neighbourhood (lo-cal control information, positions, etc.) and the wholeschool (measured gradients, maxima and minima ofsampled data, global control parameters, etc.). Themost suitable form of communication is “many-to-many”, or

omnicast

(refer to [4], [5]). In some contextsthis is also referred to as

gossiping

. This paper dis-cusses the design and implementation of an efficientTDMA scheduling algorithm, and its performanceregarding omnicast.For theoretical considerations the multi-hop radionetwork is modelled as a graph . Thismodel is very common in the literature. Vertices

of the graph represent communication nodes.Edges represent a communication link,meaning that can send messages to . For sim-plicity, we assume that links are symmetric –

. A node receives a mes-sage if and only if exactly one node in the directneighbourhood of ( ) sendsa message in some time interval. If two or more

1.

National ICT Australia is funded through the AustralianGovernment’s “Backing Australia’s Ability” initiative, in part through the Australian Research Council

G V E,( )=

v V∈v1 v2,( ) E∈

v1 v2

v1 v2,( ) E∈ v2 v1,( ) E∈⇔ vi

vk

vi vk V∈ vk vi,( ) E∈( ){ }

2 Section: Omnicast communication

nodes in the neighbourhood of send at the sametime, it is usually assumed in the literature that will not be able to distinguish the two messages fromnoise. This is called a

collision

in node . This paperfollows this common network model for theoreticalconsiderations, but will address the differences be-tween model and reality in the experimental section.

2. Omnicast communication

The required many-to-many communication formfor distributed sampling and control of schools ofsubmersibles can be formalized as:

Definition 1: (the omnicast problem)

Let be a graph describing a communication network with

nodes. In the start state, every node hasa set of information tokens, which contains exactlyone unique token of information. During the commu-nication phase, a node updates its set

, if and only if it successfully receives amessage from in time step , and

otherwise. The end state is reached,when all nodes have the full set with all tokens,

for all .

whereas the optimal solution is specified as:

Definition 2:

(

the optimal omnicast problem

) Find aschedule , for ,with being the set of sending nodes in time step ,such that solves the omnicast on the network graph ,and is minimal.

Previous publications discussed the theoretical sideof omnicast communication [5]. It could be shownthat a general upper bound for optimal solutions inmulti-hop wireless networks with nodes is .It has also been pointed out that optimal solutionsare not necessarily collision-free. An interestingpoint has been made, that fully connected networksactually perform worse with regard to omnicast thansome connected networks with a smaller in/out-de-gree. Fully connected networks require steps toachieve omnicast. There are networks for which a so-lution exists with less then steps.

3. Distributed Dynamical

Omnicast Routing

The theoretical analysis assumes full knowledge ofthe network topology. In the case of a real network,there is no entity which has this global knowledge.The information which is available to each node isonly the information contained in messages whichthey receive. In practice this means that nodes canbecome aware of their 2-hop neighbourhood.A detailed description and discussion of the DDORmethod can be found in [4]. Continuous real-timesimulations in a multi-tasking environment demon-

strated distributed schedules which were alwaysbetter than the known upper bound for optimal solu-tions ( ), and usually close to and smallerthan (persistently smaller than for heterogene-ous networks with more than 40 nodes).

Distributed measurements of Omnicast performance

These definitions consider the static case in which allmembers synchronize before a new set of informa-tion tokens becomes available. In a practical setupthis synchronization phase is omitted so that the in-dividual omnicast schedules actually overlap and ameasurement of a complete omnicast schedule (i.e.its ‘roundtrip’-time) can only be done in a distributedand averaging way.Furthermore each submersible sends its most recentset of data/measurements, so that the criterion of astatic, optimal omnicast is replaced by a measure-ment of the maximal communication delay betweennodes in a certain topological distance in the commu-nication network. As a formal basis this article specifies and discussesthe required practical performance criteria forschools of underwater vehicles. Implications fromearlier distributed scheduling works [1] and more re-cent schemes [4] are embedded and discussed in thecontext of practically achievable performances.

4. Experiments

After having shown the performance of the Distrib-uted Dynamical Omnicast Routing algorithm, exper-iments have to be carried out to test the validity ofthe underlying network model, and to verify the per-formance of DDOR in a real distributed environ-ment.

4-1. Longwave radio modules

Longwave radio modules were designed and builtfor the experiments. These modules operate on a car-rier frequency of 122.88kHz, and employ differential

vi

vi

vi

G V E,( )=

n V= u V∈Iu t0( )

Bu

v Iv t 1+( )Iv t( ) Iw t( )∪( )=

w V∈ tIv t 1+( ) Iv t( )= t f

Iu t f( ) Bv v V∈{ }= u V∈

SG T1 … Tt, ,( )= Ti V⊂ i 1 … t, ,=Ti i

SG Gt

n 2n 2–

n

n

2n 2– nn n

Thrusters Battery

Motor Controllers

Pressure Sensor

Main Processor

Gyroscopes

Acceleration Sensor

figure 1: One of the autonomous submersibles: Serafina

Section: Experiments 3

binary phase keying for digital modulation. The databit rate can be chosen to be 1024, 2048, 4096 or 8192bits per second. The modulation and generation ofthe carrier wave is done in software on a RISC micro-controller, using a number of hardware timers. Themicrocontroller runs at 9MHz and offers 32 kilobytesof flash memory and 2 kilobytes of RAM. The gener-ated modulated carrier is amplified and drives atuned resonance circuit, which acts as the transmit-ting antenna. The receiver uses a second antenna.The signal is pre-amplified, and decoded by a dedi-cated longwave decoder chip, which passes on thereceived data to the same microcontroller that imple-ments the transmitter. The complete modules areboth transmitter and receiver.The actual data communications is packet-based. Apacket can be sent in one time slot, and consists ofmultiple frames, which can be up to 16 bytes long.Contained in a packet are the sender identity, the log-ical clock value, the schedule currently used by thesender, and the user data.

The receiver rejects frames with invalid headers. Forbetter data integrity, the scheduling and sender in-formation are protected by check-sums (8 bit perframe). For performance reasons, the user data blockis not error-checked on the hardware layer. If needed,this can be done on higher protocol levels.

The current implementation of DDOR on the micro-controller allows a schedule length of 8 scheduleslots and up to 254 nodes. Packets can be up to 256bytes long. The duration of a time slot is calculatedaccording to the chosen packet length and bit rate.The overhead per packet for omnicast is 14 bytes for

sender information and the schedule, plus an addi-tional 3 bytes per frame used. The binary code is lessthan 8 kilobytes

While not transmitting, the module consumes lessthan 20mA current at 5V. The output amplifier oper-ates currently at 5V output voltage amplitude, con-suming 20mA. This corresponds to 100mW power.The transmitter output power can easily be increasedby changing the drive voltage amplitude to up to20V, but this requires currently an external 20V volt-age supply, which complicates testing. The next de-sign iteration will include an integrated, adjustableboost converter for dynamically changing the drivevoltage to suit the conditions.

4-2. Range measurements

Obviously range measurements always depend onmany parameters (output power, antenna efficiencyand tuning, noise) and the environment. The follow-ing measurements are only examples to provide de-sign guidelines to meet application requirements.The definition of range is the maximum distance be-tween transmitter and receiver that provides lessthan 10% frame dropout. It should be noted that forthis particular implementation, the increase in framedrops is very sharp – typically an increase in distanceof less than 0.3metres corresponds to a transitionfrom virtually error-free reception to total loss of re-ception. That means that within the given range, re-ception can be assumed to be quite reliable. The fol-lowing experiments were carried out with a 0.1meterdiameter circular transmitting antenna, and a 10mmdiameter 50mm length ferrite core receiver antenna.The drive voltage amplitude for the transmitter was5V. Refer to figure 3 for the results.

It can be seen that the effect of water is minimal. Evenstrongly conductive sea water has only a small im-pact on the range. The actual range might seemshort, but it lies well within our expectations. For thefinal application, the drive voltage will be increasedto 20V, which will almost quadruple the range of themodules. A communication range of more than10meters is sufficient for a school of miniature sub-mersibles, which are only 50cm long.

figure 2: Long wave radio transmitter inside Serafina

Medium Conditions Bit rate [bit/s] Range [mm]

Air Inside building4096 41501024 5400

Fresh water (chlorinated) Pool (2m deep) in 1.5m depth 4096 4000Salt water (Pacific) Coastal water (4m deep), in 3m depth 1024 3900

figure 3: Range measurement results in various conditions

4 Section: Experiments

4-3. Interference measurements

An important question is how transmitters interferewith each other, and how receivers experience mes-sage collisions. For this experiment, two almost iden-tical transmitters T1 and T2 were placed with vary-ing distance from each other in one meterincrements. Both transmitters have approximately5m range, and are continuously sending out framescontaining their own identifier. A receiver wasplaced on the straight line connecting the two trans-mitters. The receiver was then moved on this line todetermine the boundaries of the region were the re-ceiver reliably receives T1 or T2, or has no reliable re-ception. The results are shown in figure 4.It may be surprising to see that the zone in which col-lisions occur is actually quite narrow. The theoreticalnetwork model which is commonly used in the liter-ature would have predicted collisions for the entireregions where the ranges overlap. In reality, the re-ceiver can easily decode the respective stronger sig-nal of the closer transmitter. Only in the middle be-tween the transmitters, there is a narrow zone wherereception is distorted. This obviously means that theassumed theoretical network model is wrong in theassumption that a node only receives a message if ex-actly one of its neighbours is transmitting withinsome time interval. A better model would be that itreceives the message from the closest transmitterwhich is currently sending, if no equally close trans-mitter is sending at the same time. However, sincethere is no direct way for measuring these distances,nodes are generally unaware of the geometric ar-rangements of the network. Only the topological in-formation can be reliably determined, which leadsagain to the common graph model.The remaining question is now how this affects theperformance of omnicast scheduling algorithms.

This question is easy to answer for the case of alreadyestablished, collision-free schedules, such as globalsolutions or the schedules that DDOR will eventual-ly produce after a stabilization period. If a scheduleis collision free in the graph model, then it will still becollision free in reality. The performance is not affect-ed at all.The second case are established schedules containingcollisions in the graph model. These schedules willcontain less collisions in reality - only those collisionsremain, where a receiver is exposed to two equallydistant sending transmitters. With regard to omni-cast performance, this means that the performancewill not be worse, but might be better, since more in-formation that theoretically assumed is in fact ex-changed.It is much harder to answer the question how thisdifference between model and reality will affect thestabilization of a schedule in a distributed environ-ment. Nodes will receive messages which theywould not have received in the theoretical model. Inthe case of DDOR, this does not affect the calculationof schedules, which is based only on the informationwhich nodes lie in the 2-hop neighbourhood. The in-formation which can be gained by

not

receiving amessage which should have been received is not ex-ploited by DDOR. The calculated schedules willeventually be collision free. Therefore the differentconditions in reality will not affect the performance

not measured

receiving T2

Distance T2-T1 [m]

Rec

eive

r Po

siti

on R

-T1

[m]

no reception

receiving T1

figure 4: Reception zones and collision zone

figure 5: Long wave radio experiments in the Pacific Ocean

Section: DDOR in the physical world 5

of DDOR - however, the start-up phase in realitymight differ from the simulations. Since more mes-sages will be received, which are lost in the simula-tion, the stabilization is expected to be quicker.

5. DDOR in the physical world

The complete DDOR protocol has been implementedin four water proven, and battery powered radiotransceivers. All possible combinations of networktopology and joining/leaving sequences of individu-al nodes have been tested. A compact, collision-freeschedule has been created in all configurations andin all nodes.

More interesting are the configuration, and re-config-uration times which could be achieved. The joiningtime depends on the lengths of the establishedschedules in this neighbourhood. In the four nodesystem, this means that after a maximum of 15 sched-ule-slots each nodes can be joined into the network.Nodes are deleted after five missed schedule-slots.Even considering extremely low bit-rates, all thosereconfiguration procedures are completed after a fewseconds of real-time, which is perfectly acceptablefor dynamic underwater school-communications.The achieved schedule performance is identical withthe simulated measurements (as reported in [4]): al-ways less than , usually close to, or smallerthan .

6. Conclusions

A distributed scheduling algorithm suitable forschools of underwater vehicles has been presentedand implemented. It could successfully be shownthat the Dynamical Distributed Omnicast Routing(DDOR) algorithm performs well with regard to in-formation dispersal both in simulations and in reali-ty. The algorithm requires only a small amount ofmemory, computations and communication over-head, which makes it suitable to be implemented oncost-efficient, low-power, low-bandwidth longwaveradio modules. Experiments could show that digitallongwave radio is a suitable method for wirelesscommunication in sea water, and that the range is

only minimally affected compared to transmission inair. Further experiments regarding interference of twotransmitters showed that the commonly assumedmethod to model a wireless network using a graph isimprecise. The correct modelling of collisions re-quires geometrical information. However, the tradi-tional graph network model is still valuable for com-puting schedules without collisions - these scheduleswill remain collision free in reality as well.

References

[1] Ted Herman and Sébastien Tixeuil;

A DistributedTDMA Slot Assignment Algorithm for Wireless SensorNetworks

; In ALGOSENSORS, pages 45–58, 2004.[2] Shahab Kalantar, Uwe R. Zimmer;

Distributed ShapeControl of Homogeneous Swarms of Autonomous Under-water Vehicles

; to appear in Autonomous Robots (jour-nal) 2006

[3] Shahab Kalantar, Uwe R. Zimmer;

Control of OpenContour Formations of Autonomous Underwater Vehicles

;International Journal of Advanced Robotic Systems,Dec.'05

[4] Felix Schill, Uwe R. Zimmer;

Distributed DynamicalOmnicast Routing

; to appear in Intl. Journal of Com-plex Systems ‘06

[5] Felix Schill, Uwe R. Zimmer & Jochen Trumpf;

To-wards Optimal TDMA Scheduling for Robotic SwarmCommunication

; Proceedings of the TAROS (TowardsAutonomous Robotic Systems) intl. conference, Sep-tember 12-14, 2005, London

[6] Felix Schill, Uwe R. Zimmer & Jochen Trumpf;

VisibleSpectrum Optical Communication and Distance Sensingfor Underwater Applications

; Proceedings of the Aus-tralasian Conference on Robotics and Automation,December 6-8, 2004, Canberra, Australia

2n 2–n

figure 6: LED communication transmitter/controller