Applied Computer Systems doi: 10.

2478/acss-2013-0009
__________________________________________________________________________________________________2013/ 14

Use of Learning Methods to Improve Kinematic

Models
Kintija Priedniece, Agris Nikitenko, Aleksis Liekna, Guntis Kulikovskis (Riga Technical University, Faculty of Computer
Science and Information Technology, Institute of Applied Computer Systems).

Abstract – Kinematic model is the basic aspect in robot design to provide the solution of how to apply learning in order to
and motion planning. Kinematic models are idealized, however improve the kinematic model of a particular mobile robot.
there exist certain specific aspects of particular robot or
This paper presents the method to improve kinematic model
environment, so that during navigation, the robot can
significantly deviate from the planned trajectory. To increase the of a differential drive robotic platform with the help of
accuracy of motions, kinematic model can be improved and to supervised learning. In this way the accuracy of the existing
achieve that the artificial intelligence methods can be used. In differential drive can be improved and specified for a
case of fixed base robots different approaches are used to train particular task.
kinematics, at the same time, for the mobile base robots it proves The paper is structured as follows. Section II summarizes
to be a more complicated task. The reason is that a mobile robot the related work in learning kinematic by using artificial
can move unbound with respect to environment thus it is difficult
to control the platform without deviation from the target intelligence methods. Section III explains the mobile robot
position, which leads to inaccuracy in the position estimate. This kinematics by emphasizing the differential drive. Section IV
paper presents the method meant for improvement of the presents the research method in detail. Section V explains the
accuracy of motion of differential drive platform. Genetic choice of the learning method for a particular task.
programming is used to obtain the wheel velocity function, from Experiments and details related to them are described in
which the coefficient, which describes different factor influence Section VI. Finally, the results of the proposed method are
on motion, is obtained. As a result, the kinematic model of a
particular platform for a particular task is obtained. This method
discussed in Section VII.
is effective because the developed kinematic model is more
specific than the general one. II. RELATED WORK
Different artificial intelligence methods have been used to
Keywords – genetic programming, learning kinematic, mobile
improve fixed base manipulator kinematic models. Neural
platforms
networks have proved to be useful for learning the kinematics
I. INTRODUCTION of robot manipulators. Some authors used multi-layer
perceptron networks of self-organizing maps to train the
Robot kinematics deals with the analytical description of forward kinematics model of various simple systems [1],
displacement, speed, velocity, and acceleration without giving however, forward models of serial robots can easily be
regard to the force that causes the motion, thus being a computed by analytical method, therefore, in general, the
fundamental aspect of robot design, analysis, control, and evaluation is based on the accuracy of the model itself rather
simulation. Kinematic models are commonly used to control than on its control capabilities. Training inverse kinematic
motion and predict behavior of a robot. Improvement of these models is more difficult. New neural network architectures
models can reduce localization errors, ensure repeatability of have been introduced to overcome the discontinuity of the
motion, and help to avoid obstacles, etc. To improve these inverse kinematic function of typical robot arms with joint
models artificial intelligence methods can be applied. It deals constraints [2]. A new expert neural network based approach
with imprecise, variable, dynamic, or nonlinear decision has been developed, where each of the experts approximates
making. A branch of artificial intelligence research is machine the continuous part of inverse kinematics function [3]. There
learning, which focuses on automatic recognition of complex is the method, which solves some specific issues in robotic
patterns and decision-making based on sensor data. As a neuro-controller learning [4]: it avoids any neural network
result, the degree of autonomy can be increased. learning algorithm, which relies on the classical supervised
There are two types of kinematic models: fixed and mobile input-target learning scheme, it can converge beyond the
base robots. While the fixed based robot kinematics is more or locally optimal solutions; and can learn the neural network
less easy to build, obtaining mobile robot kinematic models is topology. Instead of finding a direct approximation of inverse
a much more challenging task. The main reason for this is the kinematics, it is also possible to model the joint probability
number of parameters that are known yet difficult to calculate, distribution of end-effector positions and joint angles [5].
in particular, in terms of practical implementation. Good Robot manipulator can be decomposed into two or more
examples of such parameters are: static and dynamic traction virtual robot arms to speed up the learning of the inverse
factors of a wheel on particular soil, exact position of the mass kinematics [6]. These solutions are not applicable to mobile
center, etc. Besides, there is no direct way to measure the robots, because mobile robots are not fixed. They move
robot’s position and it must be integrated over time, so it leads within the environment thus requiring to constantly estimate
to inaccuracies of the position estimates over time. Our aim is

73
Unauthenticated
Download Date | 11/4/16 9:24 AM
 Applied Computer Systems
2013/ 14__________________________________________________________________________________________________

their position over time. This means that there are no fixed
reference points to initiate kinematic calculations. (1)

III. MOBILE ROBOT KINEMATICS

When a mobile platform is analyzed, the kinematic model is
assumed to be idealized, some kinematic constraints are To solve inverse kinematic, there are two special cases of
assumed to exist [7]: movement is on a horizontal plane, control:
wheels are not deformable, there is no slipping, skidding or 1. vr=vl =>R=∞=> The robot moves in straight line
sliding and no friction for rotation around contact point, etc. It and θ remains the same;
is not possible to create an ideal system and the real situation
2. vr=-vl =>R=0 => The robot rotates in place about
always differs in some way from the predicted one.
ICC (any θ is reachable while (x,y) is unchanged).
Differential drive robot was selected to analyze mobile
This problem most often has no solution, in the sense that
platform because these platforms are ubiquitous in robotics.
the robot cannot reach an arbitrary pose by simply setting
These robots use a simple drive mechanism that consists of the
appropriate values for wheels speeds vl, vr and let the motors
two drive wheels mounted on a common axis. Moreover, each
run for a while. In real situations there are some differences in
wheel can be independently driven both in forward and
the model, therefore there is some deviation from the
reverse directions which is controlled by the speeds vl and vr
predicted motion.
set separately for the left and the right wheels. Position in the
global frame is determined by three parameters (x, y, θ),
IV. PROPOSED METHOD
where θ denotes the direction of movement. When left and
right wheel rotate with different speeds, the robot rotates For differential drive there are two control parameters of the
around the common point denoted Instantaneous Center of motion, yet there are various factors, which can affect
Curvature (ICC). By varying vr and vl, ICC moves and causes performance of a robot. Vehicles of this drive are very
different trajectories and motion curves of the robot. R in Fig. sensitive to slight changes in velocity in each of the wheels.
1 is the signed distance from the ICC to the midpoint between Small errors in relative velocities of the wheels can affect the
the wheels, l – the distance between the centers of the two robot trajectory significantly. They are also very sensitive to
wheels. roughness of the ground plane, and may need extra wheels for
There are three different motion cases that must be support. There is a number of difficult to estimate parameters
considered: that represent different imperfections of the manufacturing
1. If vl=vr, then forward linear motion in a straight line process resulting in slight deviations of the wheel speeds from
(R→∞); the set of expected values.
2. if –vl = vr, then R=0, and there is rotation about the Therefore, we introduce two parameters - coefficients kr, kl
midpoint of the wheel axis; (2), (3), which represents all possible deviations from the
3. if vl=0, then we have rotation about the left wheel. ideal, like disposition of mass center, reducer’s mechanical
In this case R =l/2. Same is true if vr=0. resistance, etc.
(2)

(3)

From the expressions (2) and (3) we define our problem

statement as finding coefficients kr and kl, which for above
mentioned reasons are specific to particular robotic platform.
Mathematically the task can be expressed and solved as rather
typical optimization task. Unfortunately the task have no
analytical solution because of the optimization criteria, which
in our case is a set of experimental data specific to particular
robotic platform. Therefore, we propose to use machine
learning that usually has good performance in solving the
problems involving experimental data. Machine learning
algorithms are organized into taxonomy, based on the desired
Fig.1. Differential drive
outcome of the algorithm. Common algorithm types include:
1. Supervised learning, where the algorithm generates a
Forward kinematic solution is typically used for automatic function that maps inputs to desired outputs;
control of a robot such that it follows the desired trajectory 2. Unsupervised learning, which models a set of inputs:
[8]. At certain given wheel speeds forward kinematic labeled examples are not available;
equations provide an updated pose. If the starting position (x, 3. Reinforcement learning, where the algorithm learns the
y, θ) is known, the new position (x’, y’, θ’) can be obtained by policy of how to act given an observation of the world.
using 3D rotation matrix:

74
Unauthenticated
Download Date | 11/4/16 9:24 AM
 Applied Computer Systems
__________________________________________________________________________________________________2013/ 14

When the desired task and the environment for mobile robot planning of motion to reach the desired goal for a particular
platform are known, it is appropriate to use supervised task.
learning, because it is possible to experimentally measure the
necessary parameters and respective desired output values. V. LEARNING METHOD
Most of the supervised learning algorithms analyze the It is very important to choose the appropriate learning
training data and produce an approximated function. The method, which is the best to apply for a particular task. The
approximated function should predict the correct output value choice was between neural networks and genetic
for any valid input object with the acceptable accuracy. In this programming, which might be used to solve regression
manner it leads to proposed method, which is shown in Fig.2. problem [9].
A. Artificial neural networks
An artificial neural network is a mathematical model or
computational model that is inspired by the structure and
functional aspects of biological neural networks. A neural
network is composed of a number of nodes, or units,
connected by links. Fig. 3 shows a typical unit. Each link has a
numeric weight associated with it wj,i. Weights are primary
means of long-term storage in neural networks, and learning
usually takes place by updating the weights.

Fig.2.Proposed method.

Our particular step implementation is discussed in the

following sections.
A. Data collection
Training data must contain the main parameters, which
describe a robot motion. In this case, control parameters and
the distance R from ICC to the midpoint between the wheels
specifies the character of motion. Distance R is important,
especially when linear motion needs to be achieved (R→∞). It
can be done by using the global coordinate system or doing it Fig.3. A unit.
manually, in this case use the drawing tool to get the necessary The computation is split into two components. The first is
parameters. Control parameters need to be randomly chosen. the linear component that is called the input function, ini, that
B. Function extraction using learning method computes the weighted sum of the unit input values. Second is
a nonlinear component called the activation function, g, that
Collected data (training examples) are used to obtain
transforms the weighted sum into the final value, which serves
regression function. Training examples are spitted into two
as the unit activation value, ai. Usually, all units in a network
subsets, yet there is no common way to do that, therefore it is
use the same activation function.
important to vary the length and examples of subsets. In
Each unit has a set of input links to other units, a current
supervised learning each example is a pair consisting of an
activation level, and a means of computing the activation level
input object and a desired output value. A supervised learning
at the next step in time, given its inputs and weights. The idea
algorithm analyzes the training data and produces an inferred
is that each unit does a local computation based on the inputs
function, which is called a regression function. The inferred
from its neighbors, but without the need for any global control
function should predict the correct output value for any valid
over the set of units as a whole. In practice, most neural
input object.
network implementations are in software and use synchronous
C. Integration of function in prediction equations for a particular control to update all the units in fixed sequence.
mobile platform To build a neural network to perform a certain task, one
A lot of parameters, which affect performance of a robot, must first decide how many units need to be used, what kind
are difficult to detect, thus variation from the idealized model of units are appropriate, and how the units will be connected
can be expressed in the form of a coefficient, which describes to form a network. One then initializes the weights of the
different factor influence. When the function of velocity is network, and trains the weights by using the learning
obtained, it can be integrated into the robot control system. algorithm applied to a set of training examples for the task.
The use of examples also implies that one must decide how to
D. Prediction or planning according to task encode the examples in terms of inputs and outputs of the
Obtained coefficients kr (2) and kl (3) are used to improve network.
position estimate in the global coordinate system or robot Advantages:
control system. In this way, it would improve prediction and

75
Unauthenticated
Download Date | 11/4/16 9:24 AM
 Applied Computer Systems
2013/ 14__________________________________________________________________________________________________

Adaptive learning: the ability to learn how to do tasks based programming operations, standard mathematical functions,
on the data given for training or initial experience [10]. Self- etc.
Organization: An artificial neural network can create its own
organization or representation of the information it receives
during the learning time. Real time operation: artificial neural
network computations may be carried out in parallel, and
special hardware devices are being designed and
manufactured, which takes advantage of this capability. Fault
tolerance via redundant information coding: partial destruction
of a network leads to the corresponding degradation of
performance. However, some network capabilities may be
retained even in case of a major network damage.
Disadvantages:
Minimizing over the fitting requires a great deal of
computational effort. Individual relations between the input Fig.4. A function represented as a tree structure.
variables and the output variables are not developed by Each individual computer program in population is
engineering judgment so that the model tends to become a measured in terms of how well it performs in the particular
black box or some input/output table without analytical basis. problem environment. This measure is called fitness measure.
The sample size has to be large. The nature of the fitness measure varies with the problem.
Typically, each computer program in the population is run
B. Genetic programming over a number of different fitness cases so that its fitness is
measured as a sum or an average over a variety of
Genetic programming is a powerful method for
representative different situations. These fitness cases
automatically generating computer programs via the process
sometimes represent a sampling of different values of an
of natural selection [11]. Genetic programming is a domain-
independent variable or a sampling of different initial
independent method that genetically breeds a population of
conditions of a system. For example, the fitness of an
computer programs to solve a problem. Specifically, genetic
individual computer program in the population may be
programming iteratively transforms a population of computer
measured in terms of the sum of the absolute value of the
programs into a new generation of programs by applying
differences between the output produced by the program and
analogs of naturally occurring genetic operations.
the correct answer to the problem.
Programs are expressed in genetic programming as syntax
Typically, the best individual that appeared in any
trees rather than as lines of code as shown in Fig. 4. The tree
generation of a run (i.e., the best-so-far individual) is designed
includes nodes and links. The nodes indicate the instructions
as the result produced by genetic programming. In symbolic
to execute (functions). The links indicate the arguments for
regression we have no a priori knowledge of the form of the
each instruction (terminals).
solution and we expect the genetic programming system to
In more advanced forms of genetic programming, programs
find both the form and the details of the solution equation.
can be composed of multiple components. In this case the
Advantages [10],[11]:
representation used in genetic programming is set of trees (one
Genetic programming does not impose any fixed length of
for each component) grouped together under a special node
the solution. Computer programs evolved by genetic
called root. These (sub) trees are branches.
programming have variable length, so that the algorithm can
The genetic operations include mutation, crossover (sexual
find the size necessary for a solution itself. The solution of
recombination), reproduction, gene duplication, and gene
genetic programming is the equation, which can be used for
deletion. Two most common operators are mutation and
different applications and it is human readable.
crossover. Mutation involves the replacement of an arbitrarily
Disadvantages:
chosen subprogram with a newly generated random
Genetic programs that can be constructed by the algorithm
subprogram. Crossover is applied on an individual by simply
are numerous. This is one of the reasons why people thought
switching one of its nodes with another node from another
that it would be impossible to find programs that would be
individual in the population. With a tree-based representation,
good solutions to a given problem.
replacing a node means replacing the whole branch. This adds
The process of finding a good solution can take a very long
greater effectiveness to the crossover operator. The
time.
expressions resulting from crossover are very much different
from their initial parents.
Most of researchers use neural networks to improve
Genetic programming starts with an initial population of
kinematic models for fixed base robots, yet genetic
randomly generated computer programs composed of the
programming seems to be more applicable for this application.
functions and terminals appropriate to the problem domain.
Its main advantage is that the result is a symbolic expression,
The functions may be standard arithmetic operations, standard
which can be integrated in the control system to improve
position estimate or control parameters. At the same time,

76
Unauthenticated
Download Date | 11/4/16 9:24 AM
 Applied Computer Systems
__________________________________________________________________________________________________2013/ 14

neural network model tends to be a black box therefore more be applied in various engineering problems of modeling and
complicated to integrate or analyze . optimization. The project contains a C# library with Genetic
Genetic programming does not require as much knowledge programming implementation algorithms and Windows Forms
about the problem and the possible solutions as neural application for graphical and visual presentation of the results.
networks. Besides, the functions which can be used for result GPdotNET also supports parallel processing for multicore
gathering can be selected manually. It can be more useful in processors based on ParallelFx library [12].
the situations when the required function is with some The preparatory steps to apply the technique for a problem
predictable characteristics. includes the selection of appropriate functions and terminals
out of which programs will be constructed, the definition of a
VI. EXPERIMENTS problem specific fitness function. For this model, 6 randomly
The goal of the experiments is to obtain specific kinematic generated constants in the interval from 0 to 10 were used. As
model, which would correspond to particular platform. First of one of wheel velocity was constant, the other one was used as
all, the training data needs to be collected and after that, the the terminal. Predicted character of the target function has to
function of the wheel velocities can be obtained. be analyzed to choose the functions, which are appropriate for
The experiments have been conducted in the indoor this task. It is predictable that nonlinearity is not cyclic,
environment. The platform used was a simple differential therefore periodical functions are not suitable for this
drive with controllable wheel velocities. The training data application. The functions used here were: addition,
contains the parameters from which the trajectory of the robot subtraction, multiplication, division, squaring, cube, square
depends: including wheel velocities and radius of ICC. For root, exponent, module, inverse proportion. Fitness value was
getting these parameters it is important to obtain the exact root mean square error. Visual representation of genetic
trajectory of the robot, because almost always it will be a programming function fitting training data was used to keep
curvature (Fig. 5). The drawing tool attached to the midpoint track of genetic programming evolution. Value difference
of the wheel axis was used for draw it on the motion surface. from real and predicted function was evaluated to find out
This provides experimental trajectory with high precision. which is the best of all the obtained results.

VII. RESULTS AND DISCUSSION

As a result, a specific kinematic model was obtained. Fig.6.
shows the graph demonstrating how the radius of ICC is
changing depending on the left wheel velocity. These results
clearly indicate that there are differences from idealized
model. In this situation the obtained coefficient is 2% of the
full wheel velocity. The largest absolute prediction errors
occur in situations when R→∞, however this does not affect
performance of the model.

Fig.5. Trajectory of differential drive.

The motion curve radius is obtained by expression (4),

where parameters from the obtained trajectory are used.

(4)

In these experiments one wheel velocity was constant. The

speed was controlled in one direction with the pulse width
modulation duty cycle (0 bit = 0%, 255 bits = 100%). To
ensure precision of measurements, for each of the applied
control parameters 4 experiments were made, and training data
contained only mean values of them. Fig.6. Graph of obtained function (vl- velocity of right wheel in %).
Experiments showed that there is a small deviation from the To evaluate precision of the obtained function, the error,
linear trajectory when control parameters were the same. which is the difference between predicted and real values, was
To obtain symbolic expression, which describes the relation analyzed statistically. As a result, 78.58% of all training set
of the parameters, GPdotNET was applied. It is a tree based examples errors fall within the interval [-5.76 cm; 6.2 cm].
genetic programming application for solving problems based Predicted values, which are calculated by using the obtained
on symbolic regression. This is an open source project and can equation, are widely dispersed around the mean value. This

77
Unauthenticated
Download Date | 11/4/16 9:24 AM
 Applied Computer Systems
2013/ 14__________________________________________________________________________________________________

happens because with the increase of the radius of curvature [5] B. Bosci, D. Nguen-Tuong “Learning Inverse Kinematics with
Structured Prediction”, Intelligent Robots and Systems (IROS), 2011, pp.
the error also increases. Precision of the test set examples is 698-703.
high. All examples are located in the interval [-5.16 cm; 3.58 [6] V. R. de Angelo and C. Torras “Learning Inverse Kinematics: Reduced
cm]. Sampling Through Decomposition into Virtual Robots”, IEEE
Transactions on Systems, Man, and Cybernetics, Part B(Cybernetics),
The obtained model is very precise, because these errors do Vol. 38, 2008, pp. 1571-1577.
not affect the choice of control parameters for obtaining linear [7] G. Dutek and M. Jenkin Computational Principles of Mobile Robotics,
trajectory when the choice is based on the radius of curvature. Cambridge: Cambridge University Press, 2000.
When pulse width modulation duty cycle rises up, the changes [8] S. Kucuk, Z. Bingul “Robot Kinematics: Forward and Inverse
Kinematics”, Industrial Robotics: Theory, Modeling and Control,
in the radius of curvature are very small (in mm), which S.Cubero, Germany: Pro Literatur Verlag, 2006, pp. 117-148.
means that even if there are some deviations, they are very [9] J. S. Russel, P. Norwig Artificial Intelligence: A modern approach,
small. Englewood Cliffs, United States of America: Prentice Hall, 1995.
[10] Y. Singh, P.K. Bhatia, O.Sangwan “A review of studies on machine
Robots are made to reach some goals and usually it is done learning techniques”, International Journal of Computer Science and
by using some plan based on some pre-generated kinematic Security, Vol.1 Issue: 1, 2007, pp. 70-84.
model of the robot,. Sometimes it is impossible to detect all [11] K.R Koza. Genetic Programming: On the Programming of Computers
by Means of Natural Selection, Cambridge, MA: MIT Press, 1992.
factors, which affect this model during its operation, therefore, [12] GPdotNET.CodePlex, Documentation, 2010 [Online]. Available:
it is difficult to find the way how to improve it. The proposed http://gpdotnet.codeplex.com/documentation.[Accessed: May 2, 2012]
method proved that mobile base kinematic models can be [13] A. Nikitenko, G.Kulikovskis “Eight wheel robotic platform and its
Fuzzy control system”, International Conference on automation, robotics
improved with the help of artificial intelligence methods, yet
and c control systems, Orlando, USA, 2010, pp. 16-23.
one needs to be clear about task specifications and
corresponding environment. The main advantage here is that it
shows this model not in the common context but in a specific
one. This provides the opportunities to use it in various
applications.
Linear motion for idealized differential drive model can be
obtained when wheel velocities are equal, thus it is the
simplest way to verify precision of the model. If there are
certain factors, which affect the motion, the platform
trajectory will not be a straight line, and the robot would
deviate from the target position. It is more advisable to use the Kintija Priedniece has received her Bc.sc.ing degree in computer control of
drawing tool to estimate deviations, because if the global electrical technologies in 2010. In 2012 she received a Mg.Sc.ing. degree in
intellectual robotic systems. Both degrees were acquired at Riga Technical
coordinate frame is used for this purpose, there may be University, Riga, Latvia.
situations when the trajectory is linear yet the robot deviates She is working as a research assistant at Riga Technical University. Her
from the target only because of the direction of motion of its research interests include robotic, artificial intelligence.
initial position.
In future we are interested in applying the learning methods
to improve kinematic model of a more difficult mobile
robotized platform, such as a 8-wheeled mobile platform [13].

ACKNOWLEDGEMENT
The research described in this paper is supported by funding
of the ERDF Project «Development of technology for
multiagent robotic intelligent system», Contract Nr.
2010/0258/2DP/2.1.1.1.0/10/APIA/VIAA/005.
Agris Nikitenko has received his Dr.sc.ing. in 2006 from Riga Technical
REFERENCES University as well as Bc.sc.ing. and Mg.sc.ing. focusing on artificial
[1] H. Sadjadian, H.D. Taghirad, “Numerical Methods for Computing the intelligence in his thesis.
Forward Kinematics of a Redundant Parallel Manipulator”, IEEE Currently he is docent in the Department of Systems Theory and Design of
Conference on Mechatronics and Robotics, Aachen, Germany,2004, Riga Technical University and vice dean of study affairs in the Faculty of
pp.557-562. Computer Science and Information Technology. His scientific interests cover
[2] E. Oyama, S. Tachi „Modular neural net system for inverse kinematics artificial intelligence and autonomy as well as their application in robotics.
learning”, Proceedings of the 2000 IEEE Int. Conf. on Robotics and He has received award of Verner fon Siemens for his doctoral thesis in
Automation, San Francisco CA, USA, 2000, pp 3239-3246. 2006. At present he is member of IEEE and ACM as well as represents
[3] E. Oyama, N.C. Young “Inverse Kinematics Learning by Modular Latvia in NATO RTO AVT panel.
Architecture Neural Networks with Performance Prediction Networks”,
IEEE International Conference on Robotics & Automation, 2001, pp.
1006-1012.
[4] J. Martin, J. Lope, M. Santos “A Method to Learn the Inverse
Kinematics of multi-link robots by evolving neuro-controllers”,
Neurocomputing, vol. 72 , Amsterdam, The Netherlands: Elsevier, 2009,
pp. 2806-2814.

78
Unauthenticated
Download Date | 11/4/16 9:24 AM
 Applied Computer Systems
__________________________________________________________________________________________________2013/ 14

Aleksis Liekna has received his Bc.sc.ing degree in 2008 and his Mg.sc.ing.
degree in 2010 from Riga Technical University. At the moment he is a PhD
student at Riga Technical University. His major field of study is computer
science.
He is working as a Research Assistant at Riga Technical University. His
research interests include artificial intelligence and multi-agent systems.
He was awarded by the Latvian Foundation for Education for his
bachelor's thesis “Development and Implementation of Reinforcement
Learning Model”.

Guntis Kulikovskis has received his Mg.sc.ing. in Riga Technical

University focusing on driving mechanisms of underwater vehicles. At the
moment he is a PhD student in the Faculty of Transport and Mechanical
Engineering.
Currently he is researcher in the Division of Autonomous Robotic Systems
of Riga Technical University. His main scientific focus areas are highly
mobile ground vehicles and underwater robotic systems.

79
Unauthenticated
Download Date | 11/4/16 9:24 AM