International Journal on Bioinformatics & Biosciences (IJBB) Vol.1, No.

1, December 2011

BIOLOGICALLY INSPIRED DESIGN FRAMEWORK FOR

ROBOT IN DYNAMIC ENVIRONMENTS USING
FRAMSTICKS

Raja Mohamed S1 and Raviraj P2

1
Associate Professor, Department of CSE, SVECW, Bhimavaram, India
[email protected]
2
Professor, Department of CSE, KKIT, Coimbatore, India
[email protected]

ABSTRACT
Robot design complexity is increasing day by day especially in automated industries. In this paper we
propose biologically inspired design framework for robots in dynamic world on the basis of Co-Evolution,
Virtual Ecology, Life time learning which are derived from biological creatures. We have created a virtual
khepera robot in Framsticks and tested its operational credibility in terms hardware and software
components by applying the above suggested techniques. Monitoring complex and non complex behaviors
in different environments and obtaining the parameters that influence software and hardware design of the
robot that influence anticipated and unanticipated failures, control programs of robot generation are the
major concerns of our techniques.

KEYWORDS
Biology, Khepera, Framsticks, Framework, Simulation.

1. INTRODUCTION TO VIRTUAL WORLD

The realistic physical modelling of characters in games and virtual worlds is becoming a viable
alternative to more traditional animation techniques. Physical modelling of characters and
environments can enhance realism and allow users to interact with the virtual world much more
freely. By modelling the forces and torques acting on bodies within a virtual environment,
detecting and responding to collisions between bodies, respecting constraints arising from, for
example, joints in articulated multi-body objects, etc., the system behaves in a believable manner
in all situations, and the user can therefore be allowed to interact with it much more freely. The
growing popularity of this approach is demonstrated by the appearance over the last couple of
years of a number of off-the-shelf physics engines aimed at games programmers [1]. While these
engines handle the modelling of inanimate bodies, programmers are still left with the task of
writing controllers for motile objects within the environment (e.g. cars, human characters,
monsters, etc.). Writing a controller for a physically modelled character is a question of
calculating the appropriate forces to apply to each body part at each point of time to achieve the
desired movement (e.g. a realistic walking gait for a human character). Artificial life techniques
can be useful in automating this task [2] [3]. For example, artificial evolution can generate
27
 International Journal on Bioinformatics & Biosciences (IJBB) Vol.1, No.1, December 2011
suitable controllers for simple behaviours, given only a high level description of that behaviour in
terms of a fitness function. In this paper, the state of the art in evolving controllers, and also in
evolving the characters’ body shapes, is described. It is then suggested that current approaches
will not be able to scale up to more complicated behaviours.

2. BACKGROUND
The design of robotic systems is particularly challenging due to the breadth of engineering
expertise required. In general, a robot may be represented schematically as shown in Figure 1a. In
brief, the Control box refers to any off-board tele-operator, whether it is a computer or a human.
Robots can be created without this component or the following Communication component,
which represents whatever means the robot, has to pass and receive information. Presiding over
the control of the robot is High-End which takes in objectives from the outside as well as internal
objectives, compares those to the information provided by the sensors, and then issues objectives
to “Low-End. Examples of high-level objectives are navigational imperatives, or data collection
routines, or manipulation directives. The needs of these objectives are translated by high-level
control into specific needs for individual actuators, primarily related to motion (position, speed,
etc…). Low-level control represents the hardware and software directly responsible for producing
the excitation signal to the system actuators. Again, information from the sensors is compared to
the desired goal, and the appropriate stimulus is fed to the Control Signal. The actuators, in turn,
act on the Mechanics, which is an enclosure name for the system, whose dynamics and statics the
controllers seek to modify. In general, the Mechanics refers to the mechanical components of a
robot. However, it could just as easily refer to a chemical solution, a magnetic field, or any other
of a myriad of other physical systems.

Figure 1. (a) Schematics of General Robot design

Simulation is a technique for exploring interesting regions of this immense landscape of robot
design. It is a platform for generating suitable and interesting forms and behaviours, not limited
by the preconceptions of a human designer's imagination. It can only get off the ground, if the
initial randomly generated robots with a non-zero score on the fitness function. For example,
consider an attempt to evolve a behaviour whereby a robot needed to process complex visual data
about the movement of another robot standing in front of it, and use this to decide whether that
object is a friend or foe. This enterprise would clearly have little chance of success if the
evolutionary process was starting from scratch.

28
 International Journal on Bioinformatics & Biosciences (IJBB) Vol.1, No.1, December 2011

3. BIOLOGICALLY INSPIRED TECHNIQUES

Researchers in the field of evolutionary robotics have considered various methods to overcome
the difficulties like selecting a non-zero fitness function, mutation etc. These generally involve
either the incremental acquisition of increasingly complicated tasks [3], or the decomposition of
the task into a sequence of easier tasks together with some way of combining them. A problem
with many of these approaches is that the decomposition of the task into easier and/or incremental
steps is something of an art in itself and there are no general guidelines to suggest the most
appropriate way to do this sensible task. Decomposition from the designer’s point of view may
not be the best route by which one can evolve a complex behaviour. Despite the problems
described above with evolving single robot to perform complex tasks, there are a number of
alternative approaches that have shown some signs of success. Five such popular methods
inspired from biological creatures are given below:

3.1. Co-Evolution
Co-evolution is a very promising technique for developing complex behaviours, especially when
there is a competition between two or more robots. The idea is that rather than evolving a robot
against a fixed fitness function, two robots are used instead, with one evolving against the other.
For example, Hillis evolved efficient number sorting algorithms by co-evolving a population of
candidate algorithms against a population of numbers to be sorted [5]. As the sorting algorithms
got improved, so are the population of numbers to be sorted, evolved to present tougher
challenges to the algorithms.

3.2.Virtual Ecologies
Here we can concurrently simulate no of robots of same type having same fitness function against
virtual world parameters like survival time, ability to complete the task and so on based on
natural selection. In order to identify complex and non complex behaviours some steps have been
proposed [7]. Now single physical simulation performed concurrently in parallel mode to test
against other robots.

3.3. Lifetime Learning

Virtual robot controller gets improved by evolution, but they do not actually adapt whilst an
individual robot is being simulated. Some recent results from evolutionary robotics suggest that
combining an evolutionary algorithm with the ability of the controllers’ to adapt or learn over an
individual robot’s lifetime can lead to improved robustness and complexity of behaviours
compared to evolution by itself [4]. Giving an individual robot the ability to adapt and learn
during its lifetime effectively smoothes the search space over which evolution is happening,
thereby helping the process to progress. It is reasonable to assume that adding these sorts of
abilities to our artificial robot will improve its ability to evolve complex behaviours just as it has
done in evolutionary robotics.

3.4. Behavioral Primitives

The evolution of complex behaviours can in general be evolved by task decomposition. Rather
than trying to evolve complex behaviours, another approach is to evolve a collection of primitive
behaviours, and then use other, non-evolutionary techniques for combining these primitive into
more complicated sequences. The task of programming a robot using this approach is like
commanding it perform an action in the virtual world.

29
 International Journal on Bioinformatics & Biosciences (IJBB) Vol.1, No.1, December 2011
3.5. User Guided Evolution
Another alternative to supplying a fixed fitness function to the genetic algorithm is to present the
user with a variety of robots from the evolving population at various intervals, and allow them to
select their favourite prototypes to be used as the foundation of the next generation. The user may
select it under any criteria, and can therefore guide the path of evolution according to their own
preferences without having to formally instruct the individual robots.

4. MOVING ROBOT DESIGN

Simulation and optimization of digital robot were attempted by several researchers. In this case
we used Framsticks software to simulate and evolve digital robots.

4.1 Framsticks

Framsticks is 3D simulation software for agents and controllers. It allows using user-defined
experiment setups, fitness functions, and neurons (sensor network) and is suitable for testing
various research hypotheses with fast 3D simulation and evolutionary optimization. The physical
structure of Framsticks agents is made of parts (material points) and joints which have touch,
move, and rotation sensors along with the end points of its structure. The control system is made
of neurons (including sensors and actuators) and their connections. Framsticks supports multiple
genetic representations and operators and it ranges from simple and direct descriptions of agents
to the ones encoding developmental process in genotypes. Further possibilities are like
performing simulation of several real life creatures and storing in a library to use later.

4.2 Autonomous Moving Robot

Figure 1. (b) K-Team’s Khepera Robot (c) Framsticks Khepera Robot

Khepera robot having 2 wheels (left, right) and 10 sensors in all directions to pick the data from
the nearby places and to decide the direction of navigation without hitting obstacles or reaching a
target in the virtual world. Sensor data will be passed to the brain or NN where decision will be
made to select the direction based on threshold values of each neuron. Khepera robot is having a
neural network associated with it to make decisions based upon real time parameters that are
picked up from the environment. Here each node in represents a neuron (i.e.) two neurons for
each wheel and a neuron for each sensor as well. Other parameters like ambience, surface type,
obstacle sensors all to have its effect over the decision which would be taken by the robot before
making a movement as shown in Figure 2 and Figure 3.

30
 International Journal on Bioinformatics & Biosciences (IJBB) Vol.1, No.1, December 2011

Figure 2. Neural network of Khepera Robot

4.3 Dynamic Simulation

The Robot is given by a set of rigid bodies called links and their kinematic structure. The links
are described by their physical attributes like inertia properties, mass, centre of mass and their
geometry in a polygonal representation. The robot's kinematic structure is defined by connecting
links with joints. Additional parameters are global gravity and maximum and minimum joint
forces and angles. Further forces result from the collision of the robot links with the floor.
Collision between robot links is not taken into account in the actual implementation.

Now we are going to design an autonomous moving robot by first creating the same in
framsticks’ virtual world. After that we will apply several experimental parameters to test the
stability of robot in different environments available in framsticks (e.g. flat land, bumpy land) as
per the techniques that are proposed earlier and both artificial and real robots are shown in Figure
1b and 1c. Khepera robot’s kinematics structure is defined using genome editor exclusively
available for framsticks and here we use f1 format with the following genomic sequence as it
describes the robot in terms of small segments in framsticks.

(rrX(lX(llSSSEEX[T:1],),lmXMMMMEEX[|1:2,-1:-3]rrSEEX[T:-
0.407](SSIISSLlEEX,,SSIISSLlEEX),))

After creating this structure we have simulated the above the digital robot in the framsticks virtual
world with several environmental conditions with different set of parametric values. Simulation
was conducted in a system with configuration of Intel Pentium Dual Core E2140 @ 2.8 GHz and
2 GB RAM in framsticks 3.0.

4.3.1 Fitness Function

In order to analyze the behavioural parameters of Khepera robot we need to define a fitness
function that can monitor the properties like time taken to perform an action, no of rotations
performed by each wheels. Even parallelized function could also be used to save computation
time.
F (t) = {Σ Si-Th => (Lw, Rw)=>r } (1)

Figure 3. Khepera’s neural net connection pattern in Framsticks

31
 International Journal on Bioinformatics & Biosciences (IJBB) Vol.1, No.1, December 2011
In the above fitness function Si represents signal from all sensor to be compared against Th
(threshold value) and movement made by left (Lw) and right (Rw) sided wheels in one unit time
or to reach a target from the current location where as r denotes no of rotations performed to
reach target. This was repeatedly applied for n no of iterations during simulation.

5. EXPERIMENTS AND RESULTS

Khepera robot’s fitness value for all types of actions will be calculated from the distance between
the robot wheels and the surface by setting default unit value 1. If this condition is not satisfied or
violated then a penalty will be applied to over all fitness of the robot. In subsequent tests the
movement of the wheels based on no of rotations from one location to the other in a stipulated
time limit was tested. Then the above sequence was repeated with fast movement by reducing the
time limit and again testing the fitness functional parameters of the robot.

Simulation was divided in to discrete steps translating the process from starting time to finishing
time Ts - Tf. + C where C is coarseness of simulation over a range of 0-1. Mean while other
simulation parameters such as pressure applied to the physical parts of the robot was also
monitored with respective sensor networks to predict the overall performance of the robot i.e.
force and torque. During simulation we have monitored several parameters that play major
concerns in designing the robot like position, velocity, and acceleration and so on.

The aim of our experiment is to realize the advantages in using biologically inspired design
techniques in building robots in terms of hardware and software components. By simulating the
robots and then monitoring several key parameters we can build robots with limited resources
using the influential data. Watching out for anticipated failures in aging of components we could
avoid catastrophic problems. Further this was extended to unanticipated failure cases also

5.1 Khepera Robot in a flat surface

To analyze the behaviour of digital robot in the virtual world we have selected a flat land surface
where there are no bumps and disturbances. During this test we have started with the fitness value
as 0 and after 30 minutes of simulation to reach the target in a flat land with and without obstacles
we achieved the fitness function value of 0.7 and the average fitness function for reaching the
target by relocating the robot in other corners of the virtual world was improving to 0.9. Further if
we increase the no of simulation steps above 1,000 certainly we can reach the maximum fitness
level. Failure rate for sensors and joints were in the region of 2 out of 25 and out of 25 for with
and without obstacles respectively.

5.2 Khepera Robot in a bumpy surface

In bumpy surface we have performed movement test for the Khepera robot. Since there is a
penalty for each move if it does not satisfy the distance between wheels and surface. Here we had
some diversified results. First without obstacles we got a fitness value of 0.4 and by increasing
them slowly over 2000 steps in the same time limit of 30 minutes it has reached the maximum
fitness value of 0.6. Performance of the robot suddenly changes due to the influence of change in
environmental parameters.

Note: - Test results shown in the table I is for target reaching test of about 25 times in two
surfaces and they are only average values under a simulation time of maximum 30 minutes.

Figure 4 shows the effect of sensors in joints and angles of the robot with respect to the user
defined fitness function including penalty value for keeping low distance between surface and

32
 International Journal on Bioinformatics & Biosciences (IJBB) Vol.1, No.1, December 2011
robot’s bottom section to avoid physical component failures in the virtual world A)flat B)Bumpy
C)Combined. Using the above data obtained by from simulation using Framsticks. Simulation
data will help selecting the right sort of hardware components like motors used in joints, sensors,
control programs and even walking strategies also. Several parameters like acceleration, force,
and walking strategies can be used for physical robot design which reduces considerable time.
Further this could be extended to other applications like surgical robots, industrial robots, land
exploring robots also.

Table 1. Comparison of Virtual Robot’s Performance in different environments

Flat Surface Bumpy Surface Combined

Name of
S.
the
No With With
Parameter Without With Without Without
Obstac Obstac
obstacle Obstacle obstacle obstacle
le le
Fitness
1 0 to 0.9 0 to 0.7 0 to 0.6 0 to 0.4 0.6 0.3
value
No of
Rotations
2 13 27 38 42 27 33
by Left
wheel
No of
Rotations
3 14 23 36 44 39 42
by Right
wheel
Sensor
4 Performanc 80% 65% 58% 49% 58% 49%
e

Figure 4. Fitness graph for different types of surfaces.

5.3 Anticipated and Unanticipated Failures

To detect such failures we have to pass through controller and simulator evolution. The
exploration phase (i.e. controller) evolves a controller for the physical robot using a robot
simulator. The estimation phase (i.e. simulator) evolves a robot simulator, given sensor data
generated by the physical robot using the controller evolved in the previous phase. Initially, the
exploration phase is run, given an approximate simulation of the robot and its environment. Once
terminating estimation phase, the best evolved controller can be dumped in to the physical robot.
The robot then behaves, and the resulting sensor data is then supplied, along with the evolved
controller, to the estimation phase. The estimation phase evolves the simulator so that the
simulated robot, given the previously evolved controller, produces the same sensor data as the
real robot. This helps in selecting appropriate hardware parts (sensors, wheels, etc) and software

33
 International Journal on Bioinformatics & Biosciences (IJBB) Vol.1, No.1, December 2011
components (evolved controller programs) which will certainly avoid catastrophic failures and to
have a fool proof robot design by reducing considerable amount of time.
F(Ri) = [Σ Fp (i) – Fs ( i ) ] / C (2)

Fitness function for simulation phase to detect failure of both type is describe as above where Ri
is the type of robot and Fp for physical robot and Fs for simulator one. C represents no of
controllers evolved in reference with equation (1). The above process was repeated for n of times.

Table 2 Failure Categorization

Cases Type of Failure

1 One of the motor weakens

2 Left wheel damage
3 Right wheel damage
4 Body damage
Neurons that control 2 wheels
5
respectively
6 Any one of the Sensor fails
7 Joints fail
8 Hidden neuron fails
9 Nothing fail

Figure 5. Anticipated / Unanticipated Failure distribution.

Failures categorized in both anticipated and unanticipated manners when we perform designing
and testing. This data helps us in coming up with a robust structure and program design to make
the robot survive in extremely critical conditions that one would be either anticipating or even a
worst unanticipated form. Here neural network failure happens very rarely but whereas hidden
layer may fail at any time. Failure of joints is highly unpredictable. Bodily damage is very rare
and happens only in unexpected cases. Sensors may fail due to the environmental or operational
34
 International Journal on Bioinformatics & Biosciences (IJBB) Vol.1, No.1, December 2011
conditions. It is obvious that the robot’s parts can be tested according to the methods suggested
and can be effectively designed.

6. CONCLUSIONS
With the help of proposed simulation techniques we were able to capture the best design for
Khepera robot and to identify complex behaviours of the same with different fitness levels. It also
helps capturing the finest details / movements of the robot in a set of artificial environments.
Methods suggested in this paper vindicate that effective design of robot by evolving them in an
artificial environment provides sufficient information to choose minimal hardware and software
parts. This improves the standard in design as well as testing robots in a successful way especially
in games and automation industries i.e. forward and backward compatibility for robot design like
simulation to reality and vice versa testing can be done. It also avoids the risk involved in
complex robot design and provides plenty of minute information to pick individual parameters
and by concentrating on low level data items when we have to deal with anticipated and
unanticipated problems involved in design and testing phases with considerably less failure rate.
In future this can be extended to design and test complex automation robots by predicting their
failure rates to avoid wasting resources and man-hours.

ACKNOWLEDGEMENTS
We are really thankful to our family members and colleagues for their patience and constant
support.

REFERENCES
[1] Arnold, D. (1997). Evolution of Legged Locomotion. MSc thesis, School of Computing Science,
Simon Fraser University.

[2] Bongard, J.C. & Paul, C. (2000). Investigating MorphologicalSymmetryandLocomotive Efficiency

Using Virtual Embodied Evolution. Meyer, J.-A. et al. (eds.) From Animals to Animats: The Sixth
International Conference on the Simulation of Adaptive Behavior: MIT Press.

[3] Dorigo, M. & Colombetti, M. (1997). Robot Shaping: An Experiment in Behavior Engineering. MIT
Press.

[4] Floreano, D. & Urzelai, J. (1999). Evolution of Neural Controllers with Adaptive Synapses and
Compact Genetic Encoding. Floreano, D. et al. (eds.) Advances in Artificial Life: Proceedings of the
Fifth European Conference on Artificial Life, pp.183-194: Springer Verlag.

[5] Hillis, W. D. (1991). Co-evolving parasites improves simulated evolution as an optimization

technique. Langton, C. G. et al. (eds.), Artificial Life II, pp.313-384: Addison-Wesley.

[6] Komosinski, M. & Ulatowski, S. (1999). Framsticks: Towards a Simulation of a Nature-Like World,
Creatures and Evolution. Floreano, D. et al. (eds.) Advances in Artificial Life: Proceedings of the
Fifth European Conference on Artificial Life, pp.261-265: Springer Verlag.

[7] Van de Panne, M. & Fiume, E. (1993). Sensor- ActuatorNetworks.ComputerGraphics (SIGGRAPH

93 Proceedings), pp.335-342: ACM SIGGRAPH.

35