SELF-ORGANISED DISTRIBUTION OF TASKS INSIDE

A NETWORKED ROBOTIC SYSTEM

Sebastian Smolorz and Bernardo Wagner

Institute of Systems Engineering, Real Time Systems Group

Leibniz Universität Hannover, Appelstr. 9A, D-30167 Hannover, Germany

Keywords: Autonomic Computing, Organic Computing, Self-organisation, Self-X, Real-time, Optimisation, Genetic

algorithm, Task distribution, Networked robotic system.

Abstract: For several years, mobile multi-robot systems are a main focus in robotic research. The possibilities to carry

out demanding algorithms in a networked robotic system grow as computational resources used by those

robots become more and more powerful and inexpensive. One major question is how to distribute the

necessary algorithms and tasks to the available computers so that all requirements are met and furthermore,

an optimal distribution pattern is achieved. This paper presents a new approach for generating an optimal

task distribution pattern without human intervention based on the concepts of Autonomic Computing and

Computational Intelligence. Although still in its conceptual state, we describe why our approach seems to be

promising for this particular problem. Additionally, further challenges and research demand are outlined.

1 INTRODUCTION

Autonomous mobile robots used for research during

the past decade usually were systems with only one

computer taking care of executing all software

modules, computations and algorithms. In the field

of robotic research, several algorithms exist that

have high computational costs, like the ICP

algorithm used for 3D point mapping as stated by

Langerwisch and Wagner (2010). Those algorithms

and with it the quality of the robots’ application

would benefit directly from the availability of more

computation power, especially when real-time

constraints have to be considered. One viable way to

boost performance is to connect several computers

and parallelize the algorithmic load.

Advanced frameworks for robotic applications

are usually organised in modules with well-defined

responsibilities and interfaces (Wulf et al., 2003). By

using a messaging service to exchange data via a

real-time capable network like RTnet (Kiszka et al.,

2005), a distribution of interdependent tasks with

fulfilled real-time requirements can be achieved.

The proper distribution of tasks to computers is

not always apparent, especially when a large number

of tasks and computers exist. Certain requirements

have to be fulfilled in any case, like deadlines. But

possible distribution patterns could differ in quality,

like the used network bandwidth or the load of

computers. To find a satisfying distribution pattern

by hand can be tedious at least; it could become

impossible at a high number of tasks and computers.

This paper presents the proposal of a system

capable of finding an optimal distribution pattern for

a set of real-time tasks and computers automatically.

Based on the concepts of Autonomic Computing

(Kephart and Chess, 2003) and Organic Computing

(Schmeck, 2005), the proposed system distributes

the tasks in an optimal manner with regard to the

minimal usage of network bandwidth and the

algorithmic load balance. The system is adaptive and

dynamic, i.e. it does not only emit a valid

distribution pattern during the setup phase but

handles environmental changes while the networked

robotic system performs like loss of processing

capacity or added computing components.

Section 2 gives a short overview of related work

and existing approaches in this area. In Section 3 the

concept of the proposed system is introduced

together with the idea of the optimization approach.

Section 4 outlines the open questions and challenges

that lie on the way to a real robotic system with the

presented autonomic properties. In Section 5 the

conclusion of our idea of a self-organised networked

robotic system is given.

235

Smolorz S. and Wagner B..

SELF-ORGANISED DISTRIBUTION OF TASKS INSIDE A NETWORKED ROBOTIC SYSTEM.

DOI: 10.5220/0003572402350238

In Proceedings of the 8th International Conference on Informatics in Control, Automation and Robotics (ICINCO-2011), pages 235-238

ISBN: 978-989-8425-74-4

 2011 SCITEPRESS (Science and Technology Publications, Lda.)

2 STATE OF THE ART

The question of implanting a software component

with self-management capabilities into a real-time

robotic control system was posed by Steiner, Hagner

and Goltz (2007), especially for the control of

parallel kinematic machines. One goal the authors

wanted to achieve with the aid of this so called self-

manager was the distribution of control tasks to

several connected PCs in a way that higher cycle

frequencies of control components became possible.

The scheduling analysis tool SymTA/S (Henia et al.,

2005) was used to check whether potential

distribution patterns meet all real-time requirements

before actually applying them.

The drawback of this approach was the necessity

of creating SymTA/S models of every system before

the execution could start. In contrast, Stein, Hamann

and Ernst (2006) present a first step towards a

distributed online performance analysis and

optimization system. However, SymTA/S models

are not generated automatically during the execution

of a distributed system but supposed to be delivered

together with system or application changes.

The facility to distribute tasks to several

computing units by means of a self-manager leads to

the question how an amount of tasks is apportioned.

Steiner et al. (2008) state that an optimal mapping of

control tasks to control PCs minimizes the network

communication and balances the algorithmic load.

As there are M

possibilities to distribute N tasks to

M computing units a two-stage heuristic is proposed

to tackle this NP-hard problem. The first stage

arranges for the balancing of the algorithmic load

while the second step tries to minimize the bus

communication. This last step is prone to destroy

some of the balance reached in the first stage so it

could be seen as problematic in view of the whole

optimization problem.

Jakob et al. (2009) broached the issue of job

scheduling and resource allocation in the context of

grid computing which can be seen as related

problem. Users send jobs to a grid cluster and expect

them to be executed as quickly as possible but also

preferably cheap. The provider of the cluster, on the

other hand, is keen to operate the resources at full

capacity and for this reason generate more income.

The authors introduce an Evolutionary Algorithm

(EA) to optimise schedules which were planned by a

previous heuristic. Experiments show measurable

improvements over the heuristic plans, even

correcting illegal schedules because of budget

violation.

3 CONCEPT

Steiner et al. (2008) identify two criteria which act

as optimization targets for the distribution problem:

1. Minimizing network communication;

2. Balancing the algorithmic load.

These criteria are adopted in our concept since

they lead to reasonable task distribution patterns.

Minimizing communication over a slow bus

improves the response time of interconnected

modules as their data can be copied inside the fast

system memory. The more bandwidth remains free

the easier it gets to add more data transmissions

between the computers if it becomes necessary at a

later time during system operation. A small band-

width utilization lowers the risk of a malfunctioning

system in case of fragile network quality.

Balancing the algorithmic load is desirable if an

algorithm can be scheduled in parallel. A good

example is the Monte Carlo Localization used by

mobile robots to determine their global position

inside a given map (Wulf et al., 2005).

Choosing an appropriate optimization algorithm

is crucial for obtaining the optimal distribution

pattern for a given task-resource conglomerate in a

specific situation. In opposition to Steiner et al.

(2008) we propose a genetic algorithm (GA) to

tackle the distribution problem. GAs, or EAs

generally (Engelbrecht, 2007), have strengths when

it comes to multi-objective optimization problems.

Moreover, GAs are not limited to a certain problem

class, thus being applicable to our distribution

problem in principle. However, for a successful

optimization process the genetic problem

representation plays an essential role which will be

handled in the following subsection in more detail.

3.1 Optimization Algorithm

The chromosome of the problem representation must

be suitable to conduct the usual genetic operations

(selection, crossover, mutation) and permit an easy

interpretation of the genotype, in our case the

mapping of tasks (T

) to computing resources

(CR

). We propose a chromosome containing as

many genes as tasks are intended to be distributed.

The value of one gene, the allele, determines the

computing resources the task is executed on.

Figure 1 shows an exemplary initial population

of three individuals. Six tasks (T

-T

) are to be

distributed to three computing resources (CR

-CR

For example, the second individual of the initial

population assigns T

and T

to CR

, T

and T

ICINCO 2011 - 8th International Conference on Informatics in Control, Automation and Robotics

236

and T

to CR

. Figure 1 also shows a

standard GA process with individual 1 and 3 being

selected and crossed after gene 4 of individual 1. A

mutation is missing in this example but not excluded

in general. More about the fitness evaluation follows

in Subsection 3.4.

Figure 1: Genetic Algorithm.

3.2 Initial Task Distribution

Before putting a networked robotic system into

operation the decision about a correct and optimal

distribution pattern has to be made, thus it is

necessary to let the evolution process run in

advance. This setup mode is noncritical and lets the

robotic system perform different task distribution

patterns to test the validity and quality of the pattern

with respect to bus communication and algorithmic

load distribution. Figure 2 depicts the setup mode

and the subsequent operating mode:

Figure 2: Setup and operating mode.

The gathering of relevant system data measured

during the live runs of different distribution patterns

can be used to construct a model of the whole

networked system. This self-model can be used later

to perform further evolution steps without

interfering with the operating system.

Once the setup mode is completed, i.e. a correct

distribution pattern is found and the evolution

process has reached an individual good enough for

putting the robotic system into operation self-

configuration (in terms of Autonomic Computing)

was successful and the system enters the operation

mode. This mode is critical since now it is not

feasible to examine further distribution patterns.

Nevertheless, it is desirable to continue evolution

because there might exist a better distribution

pattern. This self-optimization can be carried out

with the help of the self-model in parallel to the

normal tasks of the system.

Once a better individual than that forming the

current distribution pattern is found a

reconfiguration of the task arrangement can be

conducted when the system enters a safe system

state. The construction and utilisation of a self-

model can be seen as a form of self-awareness.

3.3 Error Status

Due to external conditions computing resources

could become unavailable, e.g. in case of

malfunctioning hardware, during the runtime of the

system. In such a situation the tasks being executed

on the now missing resource have to be redistributed

to the remaining resources. The system has to enter

an error status, accompanied by an emergency halt

of all critical components. Since an unknown

distribution pattern with less alleles has to be found

a new evolution run with a transition to the setup

mode is necessary. Figure 3 illustrates this self-

healing procedure.

Figure 3: Error Status (Self-Healing).

3.4 Fitness Evaluation

During the evolution process individuals with illegal

distribution patterns could arise, e.g. with deadline

violations. They could be detected during the setup

phase and sorted out. The fitness of correct

distribution patterns has to be evaluated according to

the objectives specified above. Since minimizing the

bus communication is contrary to balancing the

algorithmic load it has to be analysed how those

SELF-ORGANISED DISTRIBUTION OF TASKS INSIDE A NETWORKED ROBOTIC SYSTEM

237

conflictive objectives can be both optimized. We

have to check which multi-objective optimization

approach is applicable for our problem. One general

approach was described by Beume et al. (2008).

4 FUTURE WORK

The described concept still has to be proven and

tested. Preparations for experiments are under way,

an adequate software framework for mobile real-

time robotic systems with a plethora of modules and

sophisticated algorithms is available (Web, 2011). A

good candidate for gathering first experimental

results is the Monte Carlo Localization (MCL) as

described in Section 3. In addition to the need for

parallel computation resources during global

localization, the position tracking of the MCL needs

high iterative computing power. Therefore we have

to consider that situations exist where balancing the

algorithmic load is not an appropriate evolution

objective but rather a minor loaded computing

resource to carry out a high number of algorithmic

iterations.

The autonomic self-manager which is

responsible for finding distribution patterns by

means of the presented Genetic Algorithm as well as

constructing the self-model and online-optimization

has to be implemented and integrated into the

robotic framework RACK (Web, 2011). An open

issue is the way the online-optimization will be

realized. We have to evaluate whether the approach

of Stein et al. (2006) (see also Section 2) is viable

for us or whether the scheduling analysis algorithms

SymTA/S is based on (Henia et al., 2005) will be

directly integrated into our self-manager.

5 CONCLUSIONS

In this paper, we presented the idea of a system

capable of autonomously distributing tasks to

available computing resources. The system will be

able to generate correct and optimal distribution

patterns in order to set up the system. During this

procedure a self-model of the system is obtained and

used while the system operates to further optimize

the distribution pattern currently in use. We

described how the system performs self-healing in

case of failing components. The need for more

research in the field of the proposed optimization

approach as well as future work for getting

experimental results were outlined as well. The

system is well suited for mobile autonomous

networked robots which are employed for various

duties in complex environments where localization

algorithms are used as well as alternate requirements

cause differing calculation times of the involved

algorithms.

REFERENCES

Beume, N., Naujoks, B. and Rudolph, G. (2008). SMS-

EMOA – Effektive evolutionäre Mehrzieloptimierung.

at – Automatisierungstechnik, 7/2008, pp. 357-364.

Engelbrecht, A. P. (2007). Computational Intelligence: An

Introduction (2

ed.). Wiley.

Henia, R., Hamann, A., Jersak, M., Racu, R., Richter, K.,

Ernst, R. (2005). System level performance analysis –

the SymTA/S approach. In IEEE Proc. Computers and

Digital Techniques, Vol. 152, 2, pp. 148-166.

Jakob, W., Hahnenkamp, B., Quinte, A., Stucky, K.-U.

and Süß, W. (2009). Schnelles Scheduling mit Hilfe

eines hybriden Evolutionären Algorithmus. at –

Automatisierungstechnik, 3/2009, pp. 106-114.

Kephart, J. O. and Chess, D. M. (2003). The Vision of

Autonomic Computing. IEEE Computer, 36(1) :41-50.

Kiszka, J., Wagner, B., Zhang, Y. and Broenink, J. (2005).

RTnet – A flexible hard real-time networking frame-

work. 10

IEEE International Conference on Emer-

ging Technologies and Factory Automation, Catania.

Langerwisch, M., Wagner, B. (2010). Registration of in-

door 3D range images using virtual 2D scans. 7

International Conference on Informatics in Control,

Automation and Robotics, Funchal, Madeira, Portugal.

Schmeck, H. (2005). Organic Computing – A new vision

for distributed embedded systems. Proc. of the 8

IEEE International Symposium on Object-Oriented

Real-Time Distributed Computing, pages 201-203.

Stein, S., Hamann, A. and Ernst, R. (2006). Real-time

property verification in Organic Computing Systems.

In Proceedings of the 2

International Symposium on

Leveraging Applications of Formal Methods,

Verification and Validation, Paphos, Cyprus.

Steiner, J., Hagner, M., Goltz, U. (2007). Runtime analysis

and adaption of a hard real-time robotic control

system. Journal of Computers, Vol. 2 (10), pp. 18-27.

Steiner, J., Amado, A., Goltz, U., Hagner, M. and Huhn,

M. (2008). Engineering self-management into a robot

control system. Proc. of 3

International Colloquium

of the Collaborative Research Center 562.

Web (2011). The Robotics Application Construction Kit –

RACK. http://developer.berlios.de/projects/rack/.

Wulf, O., Kiszka, J., Wagner, B. (2003). A compact soft-

ware framework for distributed real-time computing.

In 5

Real-Time Linux Workshop, Valencia, Spain.

Wulf, O., Khalaf-Allah, M. and Wagner, B. (2005). Using

3D data for Monte Carlo Localization in complex

indoor environments. In European Conference on

Mobile Robots (ECMR), Ancona, Italy.

ICINCO 2011 - 8th International Conference on Informatics in Control, Automation and Robotics

238