VOL. X, NO. X, XXXXXXXXARPN Journal of Engineering and Applied

Sciences

ISSN 1819-6608

1

www.arpnjourn als.co m

EVOLUTIONARY ROBOTICS + 3D PRINTING = RAPID & LOW-COSTDEPLOYMENT OF AUTONOMOUS MOBILE ROBOTS

Jason Teo1, Shun-Hoe Lim2, Wei-ShunChee3, Kim-On Chin4

1,4 Faculty of Computing & Informatics, Kota Kinabalu Campus, UniversitiMalaysia Sabah, Jalan UMS , 88400 Kota Kinabalu, Sabah, Malaysia.

1 E-Mail: jtwteo@ums.edu.my 2,3 Faculty of Engineering, Universiti Malaysia Sabah, Jalan UMS, 88400

Kota Kinabalu, Sabah, Malaysia.

ABSTRACT

The design, programming and deployment of autonomous mobile robots is a highlycomplex, time-consuming and expensive endeavor. In this research, we propose anapproach which combines evolutionary robotics with 3D printing as an approach forrapid and cheaper method for the fabrication of autonomous mobile robots. We havepurposefully chosen the domains of continuum robots and hybrid articulated-wheeled robots as the proving grounds for our approach as these two areas ofautonomous robotics have been proven to be among the most complex to design andprogram as well as being highly cost-intensive to fabricate and deploy in thereal world. Capitalizing on the automated design and optimization phases ofevolutionary robotics and harnessing the rapid and relatively low cost of 3Dprinting, our tests show that the time required and cost involved to design,fabricate and successfully deploy evolved and 3D printed continuum robots as wellas hybrid articulated-wheeled mobile robots can indeed be observably be reduced.Analysis shows that the transference from simulated to real-world robots isindeed feasible and readily achievable with functioning mobile robots withautonomous behaviors that display a good level of fidelity.

Key words: Evolutionary Robotics * 3D Printing * Continuum Robots * Hybrid Robots *Articulated Robots * Wheeled Robots.

INTRODUCTION

Evolutionary robotics is currentlya highly favored method for theautomated design of robot bodies as wellas their controllers for autonomousmobile robots (Nolfi & Floreano, 2004;Doncieux, Bredeche, & Mouret, 2009).However, the vast majority of theseapproaches achieve transference fromsimulation reality through conventional

fabrication processes that are bothlabor-intensive, costly and time-consuming (Teo & Abbass, 2003; Chin, Teo& Saudi, 2008; Walker, Garrett & Wilson,2003). This problem is furthercompounded when dealing with moresophisticated autonomous robots thathave many degrees of freedoms andmultiple modes of locomotion, such asthose found in continuum robots for the

VOL. X, NO. X, XXXXXXXXARPN Journal of Engineering and Applied

Sciences

ISSN 1819-6608

2

former and hybrid articulated-wheeledrobots in the latter cases respectively.

3D printing, also commonly knownas rapid prototyping, has been gainingtremendous interest the over the lastdecade and in particular the last twoyears with dramatically reducing costsof the printer hardware itself that hasseen entry-level units decrease inprices from hundreds of thousands ofdollars to only a few thousand dollarscurrently (Savitz, 2012). This hasentrenched 3D printing in the mainstreamwithin the reach of many everyday users.3D printing has also seen tremendousgrowth in the areas of application inrecent years where not only common usesof early prototyping of end-userproducts are made but 3D printing offood items with whole designs that fullycustomizable in a true 3-dimensionalmodeling manner (Mosendz, 2014) as well3D printing of food for space travelcurrently being explored by NASA (2013)to as far as the realms of bio-engineering where living, human tissueare being 3D printed by using 3D bio-printers (Griggs, 2014).

The use of rapid prototyping forevolutionary robotics was first exploredby Lipson & Pollack (2000). However, theevolved robots were not fully autonomousas they had no sensing capabilities andrelied solely on fixed motor actionsderived from simulation. Although 3Dprinting is now commonly used inrobotics, to the best of our knowledge,there have been no studies conductedthat have yet attempted to useevolutionary robotics for thefabrication and deployment of complexautonomous, fully-sensing mobile robotssuch as continuum and hybrid robots.Mobile robots have been showing a greatsuccess in the real worldimplementation. For the first time,robots were assisting in an actual urbansearch and rescue mission of the WorldTrade Center tragedy on 11 September2001. The rescue team was assisted bysearch and rescue robots that had

succeeded to discover and subsequentlyled to the rescue of more than 10 of thetragedy’s victims (Angela, 2002). Thesuccessful involvement of mobile robotsin real life rescue missions as such hasgarnered much attention fromresearchers.

In recent years, autonomous mobilerobots including both continuum robotsand hybrid robots have been designed forvarious functionalities and purposes.For example, hybrid robots designed forstairs climbing purposes, performingjumping behavior, in-situ robots posturereconfiguration and adaptation to uneventerrains, among many others. In general,continuum and hybrid robots can carryout their missions better in rough anduneven terrains compared to traditionalwheeled or legged mobile robots. Inparticular, hybrid mobile robots utilizethe advantages of both wheeled andlegged mechanisms while compensating thedownside of each other.

There are many successful examplesof hybrid mobile robots which are builtand designed for wide range ofoperations. A group of researchers froma few universities in Japan haddeveloped a hybrid wheeled-leggedplatform through a retracting mechanisminspired by the armadillo (Tadakuma etal., 2009). The idea of a retractablewheeled-legged module is that thespecially-designed wheels can betransformed into a legs-like mechanism.The PAW robot, proposed by McGillUniversity, is a four-legged robot withwheels equipped at the end of each leg(James, Inna & Michael, 2006). PAW isthe first to combine wheeled modelocomotion with dynamically stablelegged locomotion. University Lübeck inGermany developed WheeHy which iscapable of doing in-situ reconfigurationof its posture (Bojan et al., 2010). Oneof the key features of WheeHy is thatthe robot can perform adaptation of itsposture during its traversal over uneventerrain. National Taiwan Universityproposed a Quattroped platform with

VOL. X, NO. X, XXXXXXXXARPN Journal of Engineering and Applied

Sciences

ISSN 1819-6608

3

hybrid legged-wheeled locomotion. Theproposed system utilizes atransformation method where themorphology of its wheels can be directlytransformed into legs (Shen et al.,2009).

Gregor used of contexts andcontext blocks in GP to co-evolve thecontrol and morphology of robots(Gregor, Spalek & Capak, 2012). Whileco-evolution of robot morphology andbehavior has provided an alternative inevolutionary robotics design, there arealso researches carried out onhybridizing different evolutionaryapproaches for optimization to co-evolvecontrollers and robot bodies. Howard hadshown that the idea of hybridizinggenetic algorithm with geneticprogramming is capable of discoveringthe relationship between the sets ofdata and perform more efficientlycompared to classical evolutionarymethods (Howard & Angelo, 1995). Lee hadintegrated the idea in evolutionaryrobotics where a hybrid geneticprogramming and genetic algorithmapproach is used for co-evolvingcontrollers and robot bodies toaccomplish tasks (Lee, Hallam & Lund,1996).

The main goal of this study is toinvestigate the applicability of 3Dprinting for evolutionary robotics.Specifically, we intend to show that thehighly complex and cost-intensivedesign, fabrication and deployment of(i) continuum robots and (ii) hybridarticulated-wheeled robots are feasibleby combining evolutionary designoptimization algorithms with 3Dprinting. Simulation will be carried outfor the automated design optimizationphases, after which the final, evolveddesigns will be transferred to the realworld through 3D printing. In thispaper, the scope of our study coversonly the evolutionary optimizationphases of the proposed approach whichwere conducted using and includes real-

world transference to the physical, 3Dprinted robots.

MATERIALS AND METHODS

We will present this study in twoparts. The first part encompasses ourresults achieved thus far in theevolution of 3D-printed hybridarticulated-wheeled robots; and in thesecond part, evolution of 3D-printedcontinuum robots.

3.1 Methodology for Evolving HybridArticulated-Wheeled Robots

The evolution process takes placeduring the simulation in Webots. Webotsis a development environment used tomodel, program and simulate mobilerobots. With the high fidelity physical-based robot simulation, Webots have beenwidely used by researchersinternationally. Webots is integratedwith powerful application programminginterfaces (APIs) where user can programrobots controller using 200 APIfunctions available in 6 differentlanguages such as C, C++, Java, Python,Matlab, and URBI. Most importantly,Webots file format (.wbt) can be openedand modified with regular text editorswhere the content is human readable.Thus, users can modify or generateWebots file with their own tools.

In the current stage of our on-going study, this paper aims for solvinga single objective problem which is tominimize the size of the robot body andreducing the number of collisionsbetween the robot and the obstacleduring the climbing motion. Theevolutionary algorithm is presentednext.

The evolution starts withinitiating Webots robot devices such asservos, sensors and other components,then the robot is simulated to performclimbing motion. The controller of theclimbing motion is manually designed. Ineach generation, the robot is only given

VOL. X, NO. X, XXXXXXXXARPN Journal of Engineering and Applied

Sciences

ISSN 1819-6608

4

25 seconds to perform the climbingmotion. When the time runs out, amessage will be sent to supervisorcontroller to obtain the positioncoordinate of the robot in order tocalculate the fitness score as shownbelow:

Fitness = (b + l) x (1 + d x 10) + c / 50(1)

whereb = robot body size;l = robot leg length;d = robot final distance from targetdestination;c = number of collision occurs duringsimulation.

Our fitness function is designedas in the way that the lesser is thefitness score, the fitter the robot is.Thus, the first part of the fitnessfunction is a direct addition of therobot body size with the robot leglength where our objective is to obtaina smaller robot. Besides that, theability of the robot to climb overobstacle is also an important criterionin determining the fitness of a robot.Thus, the distance of the robot finalposition from the target destination ismultiplied with the first part of theequation to give merit to the fitnessscore if the robot succeeded to climbover obstacle and reach its destination.The last part of the equation is topenalize the fitness score of the robotwith the number of collision occursduring the climbing obstacle motion. Thetotal number of collisions occurs duringthe simulation is divided by fifty as toscale down the penalty score where thenumber of fifty is the maximum number ofcollision it may occur in one run.

After the fitness score isobtained, it will be compared to thefittest robot fitness score. If thecurrent fitness score is better than thefittest robot fitness score, the currentrobot will be kept as the fittest robot.Then the fittest robot will undergomutation and produce a new robot for

next generation. After replacing theWebots file with the new robot, amessage will be sent to supervisorcontroller to revert the simulation. Thesimulation will be repeated until themaximum number of generation is reached.

3.2 Methodology for Evolving ContinuumRobots

Again using the Webots simulator,the continuum robot, also known as amodular robot, is constructed through acombination of blocks with differencesizes. The size of each block is createdaccording to the segment lengthspecified in tree-based structure unit.A touch sensor is installed in themiddle of each block. If either side ofthe block is being touched, a HIGHoutput is produced by the touch sensor,or else the sensor output remains LOW.Distinct from multiple-branching modularrobots, the snake-like modular robot isequipped with an infra-red sensor infront of the first segment. Such aninfra-red sensor will give a LOW outputif an object is sensed, else it remainHIGH. In order to avoid segmentsoverlapping in simulation, bounding boxfeatures are implemented. A bounding boxis attached to every block with the samedimension which makes the virtual blocksegment become a solid object.

For the simulation environment, anopen area environment is used for multi-branching modular robots evolutionaryprocess whereas a narrow straight pathis designed for snake-like modular robotevolutionary process. In the artificial evolutionary process,each population consists of tenindividuals. Genetic Programming andDifferential Evolution were performed onthe population, each individual is runin the same environment one followinganother. After a particular period ofrunning time, each individual modularrobot will be evaluated using thefitness function modeled. The fitnessfunction for the snake-like modularrobot is the total distance travelled by

VOL. X, NO. X, XXXXXXXXARPN Journal of Engineering and Applied

Sciences

ISSN 1819-6608

5

the modular robot in meters. On theother hand, the fitness function of themulti-branching modular robot is shownbelow:

(2)

(3)

RESULTS AND ANALYSIS

4.1 Results for Evolved HybridArticulated-Wheeled Robots

Six legged articulated-wheeledhybrid mobile robots with variousparameters of morphology are obtainedfrom the mutation operation during theevolution process. Numerous robots withdifferent combinations of wheels, legand body sizes have been tested in thesimulation on its ability for climbingobstacles. In the evolution stage,smaller robots with the ability to climbover obstacles without collision willhave the best fitness score. Thus,robots with large body sizes or robotswhich are too small to perform theclimbing motion are discarded from theevolutionary-optimization process.

Before the simulated of artificialevolution is conducted, a six leggedarticulated-wheeled hybrid mobile robothad been manually designed for climbingobstacle purposes. The robot was used asthe initial generation of the evolution.Table 1 shows the comparison of theinitial robot and the fittest robotproduced from the evolution. The handdesigned robot is shown in Fig. 1.

TABLE 1COMPARISON OF THE FIRST GENERATION ROBOT AND THE LAST

GENERATION ROBOT

Object Initialrobot

Fittestrobot

Wheels radius (mm)

15 22

Robot leg length(mm)

50 45

Robot body (mm) 200 150Fitness score 0.403 0.195Number of collisions

8 0

From the experiment, a fitter sixlegged articulated-wheeled hybrid mobilerobot is obtained which having fitnessscore of 0.195 comparing to initialhuman design robot 0.403. The fitterrobot evolved from the evolutionsimulation is smaller in size comparedto the initial human-designed robotwhile having zero collision withobstacles during the climbing motion.The fittest robot evolved from theevolution simulation process is shown inFig. 2.

Fig. 1. Hand-designed six legged-wheeled hybrid mobile robot presented inWebots.

VOL. X, NO. X, XXXXXXXXARPN Journal of Engineering and Applied

Sciences

ISSN 1819-6608

6

Fig. 2. The fittest six legged-wheeledhybrid mobile robot evolved from theevolution simulation process.

Table 2 shows the parameters ofeach of the fittest robots evolved infive different runs of evolutionarysimulation. Results show that everyevolved fittest robot has better fitnessscore than the initial human-designedrobot. Besides that, the results alsoshow that the fitness scores of thefittest robot from different runs arequite similar to each other. This showsthat all the fittest robots that wereobtained from the evolution are near tothe optimum parameter values which aresmall enough but yet have a good abilityin performing the climbing motion.

TABLE 2PARAMETER OF THE FITTEST ROBOT IN FIVE RUNS

Object RunOne

RunTwo

RunThree

RunFour

RunFive

Wheels radius (mm)

21 20 20 19 19

Robot leglength (mm)

45 47 47 50 48

Robot body (mm)

150 150 150 150 150

Fitness score

0.195 0.197 0.197

0.2 0.198

Number ofcollisions

0 0 0 0 0

As been seen in Table 2, theevolved robot with 0.195 fitness scorehave legs with lengths of 45mm is theoptimal length of the leg for a robot toperform climbing motion in the currentsimulation setup. Robots with legs whichare shorter than 45mm are too short thatit cannot reach the top of obstaclewhile robots with legs which are longerthan 45mm are credited with poorer

fitness score which can be seen in Table2. Fig. 3 shows the fitness score of thefittest robot over the evolutionaryoptimization process for five runs.Results show that the fittest solutioncan be obtained even in the early stagesof the evolutionary optimizationprocess.

Fig. 3. Fitness score of the fittestrobot over 500 generations for fiveruns.

4.2 Results for Evolved Continuum Robots

After each generation of theevolutionary process, the system willautomatically compare the performancescore of each individual according tothe number of segments possessed by themodular robot individual. The systemwill store the robot individual’sinformation under two conditions: (a) ifthe fitness scores of the currentindividual is higher than the previousstored individual with the same numberof segments (b) if the individual withthe particular total number of segmentsdoes not yet exist throughout theevolutionary process. The fitness scoresof multi-branching modular robots atdifferent generations are shown in Table3. Meanwhile the distance travelled (inmeters) by the snake-like modular robotsare shown in Table 4.

VOL. X, NO. X, XXXXXXXXARPN Journal of Engineering and Applied

Sciences

ISSN 1819-6608

7

TABLE 3Multi-branching Modular Robot Fitness

ScoreSegmentNumbers

5th

Generation

10th

Generation

20th

Generation

40th

Generation

3 3.3983 3.3983 3.39834 38.0616 38.0616 38.0616 38.06165 -

18.8675-

18.8675-

18.86756 145.278

8231.422

1249.819

3396.780

97 -

441.3252

109.8865

109.8865

9 -239.383

4

114.0557

114.0557

10 318.0996

318.0996

318.0996

318.0996

11 -12.1256

103.4066

103.4066

12 -109.488

8

-66.1687

-66.1686

-66.1686

13 -4.0802 449.2480

449.2480

15 -57.6467

195.1755

16 190.3900

190.3900

17 175.3790

175.3790

Table 3 and Table 4 show that themodular robots are able to improve theirperformances from generation togeneration through the evolutionaryoptimization process. Besides that, itcan also be observed that newindividuals with different structuresand total number of segments are able tobe created eventually through theevolutionary optimization process. Fromboth Table 3 and 4, it can be seen thatthere are some individuals created butthe fitness score did not changefollowing that. This is because suchindividuals were created only once

during the evolutionary process. Howeverthe created individuals did not performbetter compared to its parent. As aresult, it was discarded from thepopulation and no further evolutionaryprocess was performed on it to improveupon its behavior. Yet, the informationof such an individual will still bestored by the system as there is noindividual with that particular numberof segments throughout the artificialevolutionary process. Fig. 4 below showsthe evolved continuum robots with multi-branching and snake-like morphologies.

TABLE 4SNAKE-LIKE MODULAR ROBOT TRAVELLED DISTANCE

(METERS)SegmentNumbers

5th

Generation

10th

Generation

15th

Generation

20th

Generation

4 0.4809 0.4809 0.4809 0.48095 1.7197 2.2288 2.23238 2.4157 2.9985 2.9985 3.148510 2.0859 3.0344 3.0344 3.133812 1.6554 2.0822 2.0822 2.082215 2.2947 2.294716 1.6566 1.6566 1.6566 1.656617 2.0799 2.0799 2.079918 2.0025 2.002520 2.0908 2.4214

(a)

VOL. X, NO. X, XXXXXXXXARPN Journal of Engineering and Applied

Sciences

ISSN 1819-6608

8

(b)

Fig. 4. (a) Simulation for multi-branching modular robot in open area.(b) Simulation for snake-like modularrobot within a specified path.

By comparing Table 3 and 4, it isnoted that very often, a significantstep jump in the fitness score of themulti-branching modular robot happens ascompared to the case of the snake-likemodular robot. This is because the bodystructure orientation of the multi-branching modular robot is able to bealtered during the evolutionary processas compared to the snake-like robot,where the attachment of every segment isbeing constrained to connect in astraight line.

As a result, the multi-branchingindividual will have a totally differentbehavior if there is change on theorientation of its body structure. Sincea totally different behavior is beingevolved, the performance score can beachieved by the individual will alsovary significantly. Although the snake-like modular robot structure’sorientation is fixed, the segment lengthof the robot is still evolvable, whichalso contributes to the robot’s movingbehavior. The evolved segment lengthsfor the snake-like modular robot witheight segments are shown in Table 5.

TABLE 5Snake-like Modular Robot Segment Length

(8 Segments)SegmentSequenc

e

5th

Generation

10th

Generation

15th

Generation

20th

Generation

1 0.1100 0.1100 0.1100 0.11002 0.1200 0.1200 0.1200 0.12003 0.1300 0.1120 0.1120 0.11834 0.1400 0.1047 0.1047 0.10135 0.1700 0.1251 0.1251 0.11706 0.1800 0.1800 0.1800 0.18007 0.1840 0.1316 0.1316 0.10008 0.1940 0.1772 0.1772 0.1874

4.3 Resulting 3D Printed Robots

The results of transferringevolved robots from simulation to thereal-world robots are achieved through3D printing based on the fuseddeposition modeling methodology. Fig. 5below shows the fabricated and deployedevolved hybrid articulated-wheeledmobile robot and Fig. 6 which followsshows the fabricated and deployedevolved continuum robot.

VOL. X, NO. X, XXXXXXXXARPN Journal of Engineering and Applied

Sciences

ISSN 1819-6608

9

Fig. 5. Evolved and 3D printed hybridarticulated-wheeled mobile robot.

Fig. 6. Evolved and 3D printedcontinuum robot.

The robots were verified to befully-functioning in the real world,mimicking the expected autonomouslocomotion behavior exhibited andachieved during the evolutionary

optimization phases of the experiment.Hence, the real-world transference fromsimulation has shown that the proposedapproach of hybridizing 3D printing withevolutionary robotics is indeed feasibleand achievable. Compared to conventionalfabrication processes, the time and costinvolved were much lower. Elementarycomparisons for both the hybridarticulated-wheeled robot as and thecontinuum robot are given below in Table6 and Table 7 respectively.

TABLE 6TIME AND COST COMPARISON FOR DEPLOYMENT OFHYBRID ARTICULATED-WHEELED MOBILE ROBOTS

TABLE 7TIME AND COST COMPARISON FOR DEPLOYMENT OF

CONTINUUM ROBOTS

VOL. X, NO. X, XXXXXXXXARPN Journal of Engineering and Applied

Sciences

ISSN 1819-6608

10

CONCLUSION

We have proposed a novel approachto the hybridization of evolutionaryrobotics to 3D printing for the rapidand cost-effective design, fabricationand deployment of both continuum andhybrid articulated-wheeled mobile robotsthat are fully autonomous with sensingcapabilities. The evolved robotmorphologies from the high-fidelitysimulations show that efficient designscan be automatically evolved using ourapproach. Furthermore, these evolveddesigns in simulation were successfullytransferred to real-world robots via 3Dprinting that was less labor-intensive,faster and cheaper.

In future work, we will beincorporating multi-objectiveevolutionary optimization into ourapproach to enable multiple objectivesto be achieved during the evolutionaryoptimization phase, such as minimizationof the robot’s weight in addition to thecurrent maximization of its primarylocomotion behavior.

ACKNOWLEDGEMENT

This research work was funded underScienceFund project ref: SCF0085-ICT-2012 granted by the Ministry of Science,Technology & Innovation, Malaysia.

REFERENCES

S. Nolfi and D. Floreano, EvolutionaryRobotics, MIT Press, 2004.

S. Doncieux, , N. Bredeche, J.-B. Mouret(eds.), New Horizons in EvolutionaryRobotics, Springer, 2009.

J. Teo & H.A. Abbass, “Elucidating thebenefits of a self-adaptive Pareto EMOapproach for evolving legged locomotionin artificial creatures”, 2003 Congresson Evolutionary Computation, pp. 755-762, 2003.

K.O. Chin, J. Teo, A. Saudi, “Multi-objective artificial evolution of RF-localization behavior and neuralstructures in mobile robots”, 2008 WorldCongress on Computational Intelligence,pp. 350-356, 2008.

J. Walker, S. Garrett, and M. Wilson,“Evolving Controllers for Real Robots: ASurvey of the Literature”, AdaptiveBehavior, 11:179-203, 2003.

E. Savitz, “Manufacturing The Future: 10Trends To Come In 3D Printing” (online),www.forbes.com/sites/ciocentral/2012/12/07/manufacturing-the-future-10-trends-to-come-in-3d-printing/, 2012.

P. Mosendz, “3D-Printed Food ActuallyLooks (and Tastes) Pretty Delicious”(online),www.thewire.com/technology/2014/ 05/3d-printed-food-actually-looks-and-tastes-pretty-delicious/ 371863/, 2014.

NASA, “3D Printing: Food in Space”(online),www.nasa.gov/directorates/spacetech/home/feature_3d_food. html, 2013.

B. Griggs, “The next frontier in 3-Dprinting: Human organs” (online),edition.cnn.com/2014/04/03/tech/innovation/3-d-printing-human-organs/,2014.

H. Lipson and J.B. Pollack, “Automaticdesign and manufacture of roboticlifeforms,” Nature, 406:974-978, 2000.

D. Angela, “Urban Search and RescueRobots: From Tragedy to Technology,”IEEE, 2002.

VOL. X, NO. X, XXXXXXXXARPN Journal of Engineering and Applied

Sciences

ISSN 1819-6608

11

K. Tadakuma, R. Tadakuma., A. Maruyama,E. Rohmer, K. Nagatani, K. Yoshida, A.Ming, S. Makoto, M. Higashimori, and M.Kaneko, “Armadillo-Inspired Wheel-LegRetractable Module,” in IEEEInternational Conference on Robotics andBiomimetics, pp. 610-615, 2009.

J. Bojan, H. Martin, K. Michael, and M.Erik, “Design of a hybrid wheeled-leggedrobot – WheeHy,” IEEE, 2010.

S.Y. Shen, C.H. Li, C.C. Cheng, J.C. Lu,S.F. Wang, P.C. Lin, “Design of a Leg-Wheel Hybrid Mobile Platform,” in IEEEInternational Conference on IntelligentRobots and Systems, 2009.

S.A. James, S. Inna, and T. Michael,“PAW: a Hybrid Wheeled-Leg Robot,” inIEEE International Conference onRobotics and Automation, pp. 4043-4038,2006.

M. Gregor, J. Spalek and J. Capak, “Useof context blocks in genetic programmingfor evolution of robot morphology,” IEEEELEKTRO Conference, pp. 286-291, 2012.

L.M. Howard and D.J. Angelo, “The GA-P:a genetic algorithm and geneticprogramming hybrid,” IEEE EvolutionaryProgramming, 1995.

W.P Lee, J.Hallam, and H.H. Lund, “Ahubrid GP/GA approach for co-evolvingcontrollers and robot bodies to achievefitness-specified tasks,” IEEEEvolutionary Computation, 1996.