KAIRO 3: Moving over stairs & unknown obstacles with reconfigurable snake-like robots

Abstract

We present a planning approach for complex motions of reconfigurable snake-like robots to overcome challenging obstacles like stairs or large steps. The current robot configuration as well as different optimization criteria like distance, time or energy, are taken into account. The planning time remains unaffected by the amount of robot modules by combining the planning method with a follow-the-leader control approach. Implementation details on the developed method are presented and extensively evaluated with the reconfigurable snake-like inspection robot KAIRO 3 in three different scenarios. Our results show that the approach enables snake-like robots to overcome previously unknown obstacles according to their robot configuration.

Full text

©2015 IEEE. Personal use of this material is permitted. Permission from IEEE must be obtained for all other uses, in any current or future
media, including reprinting/republishing this material for advertising or promotional purposes, creating new collective works, for resale or
redistribution to servers or lists, or reuse of any copyrighted component of this work in other works. DOI: 10.1109/ECMR.2015.7324209
Mobile Robots (ECMR), 2015 European Conference on, vol., no., pp. 1-8, 2-4 Sept. 2015.
1
KAIRO 3: Moving Over Stairs & Unknown Obstacles
with Reconfigurable Snake-Like Robots
L. Pfotzer∗, M. Staehler∗, A. Hermann∗, A. Roennau∗and R. Dillmann∗
∗Department of Interactive Diagnosis and Servicesystems, FZI Research Center for Information Technology
Haid-und-Neu-Str. 10-14, 76131 Karlsruhe, Germany
Email: [pfotzer, staehler, hermann, roennau, dillmann]@fzi.de
Abstract—We present a planning approach for complex mo-
tions of reconfigurable snake-like robots to overcome challenging
obstacles like stairs or large steps. The current robot config-
uration as well as different optimization criteria like distance,
time or energy, are taken into account. The planning time
remains unaffected by the amount of robot modules by combining
the planning method with a follow-the-leader control approach.
Implementation details on the developed method are presented
and extensively evaluated with the reconfigurable snake-like
inspection robot KAIRO 3 in three different scenarios. Our
results show that the approach enables snake-like robots to
overcome previously unknown obstacles according to their robot
configuration.
I. INTRODUCTION
There exists a large variety of conventional planning meth-
ods, such as described by Latombe et al. [1], LaValle et al. [2],
Siciliano et al. [3] and Coenen et al. [4]. They focus on mobile
and wheeled robots and optimize plans to avoid obstacles
and prevent collisions. Bio-inspired inspection robots, which
are applicable in unstructured terrain, come with contrasting
requirements. Thus, for snake-like robots, biologically inspired
motion generation and planning methods are used to create
whole body motions. The simplest form is constituted by
fixed motion patterns, which are implemented in different
forms by Tsakiris et al. [5] and Bayraktaroglu et al. [6]. More
sophisticated approaches rely on central pattern generators
[7], [8] or so called backbone curves [9] which produce bio-
inspired motions. Hirose et al. [10] and Chirikjian et al. [11]
also considered motions of snake-like robots in unstructured
terrain with obstacles. In particular, there exist hybrid systems
like OmniThread [12], Wheeeler [13], KOHGA [14] and
KAIRO 3 [15], which are equipped with additional wheels.
Basic motion of wheeled snake-like robots can be achieved
with the follow-the-leader approach as shown by Scholl et al.
[16] and Yamada et al. [17]. A second method is called n-trailer
and illustrated by Granosik et al. [18], Murugendran et al. [19]
and Altafini et al. [20].
Most of the shown conventional navigation and motion
planning methods generate paths around obstacles to avoid
collisions. Because snake-like robots are able to move or climb
over obstacles, specialized motions and planning strategies are
required for such kinds of robots. Additionally, only very few
approaches are able to consider different robot configurations1
for navigation and motion planning, which is very important
in the case of reconfigurable robots.
1In this context, a robot configuration is denoted as the assembly and
connection of compatible hardware modules to form a robot system.
Fig. 1. The modular snake-like inspection robot KAIRO 3 [15]. On the left
figure KAIRO 3 raised its head e.g. to get an overview of the environment.
The right figure shows KAIRO 3’s ability to move through narrow places.
The major goal of our work is to increase and enhance
the autonomy and adaptability of snake-like inspection robots,
especially for search and rescue purposes in unstructured
terrain with obstacles. Thereby, our proposed approach aims to
reduce the complexity of control and to improve the navigation
and motion planning for modular reconfigurable robots with a
high degree of freedom (DoF). Precisely, this means planning
of complex robot motions in 6-D space to overcome obstacles
autonomously and considering the current robot configuration
in the planning process to optimize the locomotion due to time
and distance regarding the capabilities of the robot. Therefore,
we developed a three stepped motion planning process based
on the RRT∗algorithm and an accurate environment model.
This navigation system is then connected to the existing motion
control software of the reconfigurable, snake-like and wheeled
inspection robot KAIRO 3.
II. THE MODULAR SNA KE -L IK E ROB OT KAIRO 3
In 2013, the snake-like reconfigurable inspection robot
KAIRO 3 (shown in Fig. 1) has been developed based on the
KAIRO II platform. Due to the modular and flexible design,
KAIRO 3 is applicable to many different inspection tasks
in narrow places and environments which are dangerous for
humans. KAIRO 3 is biologically inspired by inchworms and
snakes, likewise it is equipped with wheels in every second
module resulting in a hybrid snake-like wheeled robot system.
Because it consists of two fundamentally different types of
modules, drive modules and joint modules, illustrated in Fig. 2,
KAIRO 3 can be classified as a chain-type heterogeneous
modular robot. The drive modules are connected with the joint
modules by mounting interfaces which enable changing the
configuration of KAIRO 3 easily. All actuators and control
units are integrated very compact, thus it is possible to add
several sensors, batteries, embedded systems or even transport
items inside the drive modules.978-1-4673-9163-4/15/$31.00 c
2015 IEEE

2
Fig. 2. A joint module of KAIRO 3 is shown on the left and a drive module
with two individual controlled wheels is depicted in the middle. The right
side shows a sketch of a joint module with three actively driven axes: alpha,
beta and gamma. The alpha and gamma axis are aligned along the drive
direction while the beta axes is tilted by 45◦.
A. Different Configurations
In principle, ignoring technical and computational con-
straints, KAIRO 3 may have an unbound number of modules.
In order to keep the system manageable, our robot consists
of six drive modules, which are interleaved with five joint
modules in the largest configuration shown in Fig. 1. With this
configuration KAIRO 3 has 27 DoF, a length of about 1.8 m
and a weight of about 47 kg including batteries. However,
KAIRO 3 is also able to pass through small openings and pipes
with a minimum diameter of 25 cm because of its slim snake-
like design. Due to the connection interfaces, the robot can
be separated into several different configurations of smaller
robots. Fig. 3 illustrates examples of two different multi-robot
configurations with different amounts of modules.
Fig. 3. The left configuration comprises three single robots with at least
three modules which is the minimal amount of modules. On the right image
two robots with five as well as six modules form another configuration.
B. Low-level System Architecture
The Adaptive Control architecture of KAIRO 3 is imple-
mented in the robotic framework MCA2 (Modular Controller
Architecture) [21]. This hierarchical control software is able to
adapt to the actual amount of hardware modules. Fig. 4 illus-
trates the connection between the Adaptive Control, consisting
of two layers, and the navigation planning system.
The decentralized Base Control is a mixture of hardware
and software located in every module of the robot. On top
the Base Control, the kinematics calculation of KAIRO 3 as
well as basic maneuvers like Normal Drive and Inspection are
realized within the Motion Control, which uses a virtual rail
[16]. The virtual rail is build by several way points, which
are interconnected by linear line segments. All joint modules
of KAIRO 3 follow this rail. Furthermore, two biologically
inspired locomotion methods have been implemented [22] to
allow KAIRO 3 movement in uneven terrain without using
wheels.
III. NAVIGATIO N SY ST EM A PP ROACH
Based on the PlexNav [23] navigation framework and the
Adaptive Control of KAIRO 3 [24], [15], we developed an
integrated navigation framework for reconfigurable snake-like
robots. This approach is able to generate and execute complex
trajectories regarding the robot kinematics and configuration.
Fig. 4 shows the overall structure which divides the planning
problem into several steps realizing individual processing when
required.
Fig. 4. System architecture of the integrated navigation and motion planning
approach.
A. Mission Control and User Interface
The Mission Control is the central part and controls the
planning system. It takes care of scheduling mission goals and
parameters, specified via the User Interface. A list of naviga-
tion goals is utilized to select the next target pose2as input for
the planning system. Furthermore, the Mission Control triggers
the perception and modeling of the environment. To increase
the environment coverage regarding different types of sensors,
our approach offers the possibility to include robot specific
mapping maneuvers as special planning goals. These mapping
goals can be triggered either regularly after a specific time
or distance but also on external events like detection of large
steps or gaps.
B. Environment Modeling and Localization
The PlexMap [23], which has been developed for the
PlexNav framework, is utilized for modeling the environment.
This specialized 2.5-D map is able to import point clouds in the
common PCL point cloud format. Thus, it supports data from a
large variety of different 3-D sensors like KaRoLa [25]. Due to
an integrated ICP based matching algorithm, new point clouds
are registered against the previously taken point clouds and
then converted to the existing map. Besides the height value,
also gradient and variance are stored for each map cell. Cell
sizes are dynamically adjusted depending on the spread and
amount of measurements. Additionally, the point clouds are
matched against each other to establish a relative localization.
This is used to correct the robot odometry and to realize an
accurate robot localization without the need for an external
tracking.
2A target pose consists of six dimensions (x,y,z,roll,pitch,yaw) and
describe the goal position and orientation of the first robot module.

2
C. Trajectory Planning
In general, path or motion planning methods generate
different kinds of trajectories which the robot is able to follow.
Several approaches like Rapidly-Exploring Random Trees [26]
and Randomized Kinodynamic Planning [27] are imaginable
to deal with the high number of DoF of snake-like robots
(remember KAIRO 3 has 27 DoF in joints plus six for the
pose of the base point).
The classic approach would be the usage of a mathematical
kinematics robot model (e.g. joint angles and pose) to constrain
the RRT∗planning. This direct optimization of the joint angles
would be the best solution to find an ideal and accurate path.
But with an increasing amount of robot modules also the
dimension of this model and the search space would enlarge
extremely. Also, with a full kinematics model it is difficult to
consider both, adapting the robot smoothly to the ground and
overcoming obstacles and gaps.
Due to these drawbacks we propose an approach which
uses a reduced planning space and multiple planning stages.
Thereby, the planning space is considered constant, also with
a changing amount of modules, by planning a full 6-D path
only for the first module of the robot. This is possible because
afollow-the-leader approach is used for Motion Control (also
described as virtual rail in Section II-B). The planning itself is
divided in the three steps Coarse Planning,Precise Planning
and Joint Angle Planning as stated in Fig. 4.
1) Coarse Planning: This step uses a RRT∗algorithm [28]
with the cost function shown in Section III-D1 to generate
a coarse path for the first robot module. Furthermore, the
RRT∗has been extended by a secondary nearest neighbor
space, introduced in Section III-D2, to improve planning with
nonholonomic constraints. The planner uses a large step width
and is allowed to advance in areas with sparse map information
(over-the-horizon planning).
2) Precise Planning: In a second step, sections of the gen-
erated path, which are located in well known areas of the map,
are refined. Therefore, the same extended RRT∗algorithm
is utilized, but with a smaller step width and modified cost
function parameters. Over-the-horizon planning is not allowed
here.
3) Joint Angle Planning: The third step smooth the path
and adapt the poses to the ground respectively the surface
of obstacles considering the robot kinematics and using the
ground contact points of the wheels. For each pose, the target
joint angles for the first joint module is determined in order to
rotate redundant joints in advance and to keep them in sync
with the neighbor modules. The desired joint angles can be
calculated with an inverse kinematics model and will be added
to each joint module 6-D pose of the path.
D. Extended RRT∗Algorithm
1) Cost function: To fit the characteristics of a wheeled
snake-like system, the cost function proposed for PlexNav [23]
was adapted and extended. The cost to reach state s2from state
s1is calculated by
costs1,s2=DI ST(s1, s2)1
N
N
X
q=0
C(q)
+ǫ(YAW(s1, s2) + 1)2−1
DI ST(s1, s2)·π
+ζ·OR IE NTATIONDIFF(s2, st)
The first term is adapted from PlexNav and calculates the
average cost for all intermediate states qbetween s1and s2
multiplied with the Euclidean distance of the two states. C(q)
is the maximum of the weighted costs for the
•map height difference between following modules3,
•roughness of the ground and
•cell size dependent on the number of scan points.
Inclination costs are not considered due to the robot’s ability
to climb vertical steps. Two new terms has been introduced to
handle the nonholonomic constraints of a wheeled snake-like
robot. One term calculates the yaw angle difference between
the two states divided by their Euclidean distance and normed
by π. Orienting new samples to the direction from their parent
poses, another term is needed, which takes the orientation
difference between s2and the target state stinto account. Both
terms are weighted by ǫand ζ.
2) Secondary Nearest Neighbor Space: The newly intro-
duced cost function terms caused a discard of many sampled
poses, because they do not satisfy the nonholonomic planning
constraints. To get rid of this problem, an extension of the
RRT∗algorithm is proposed to reduce the amount of dropped
poses by a factor of ten. For every pose a second one is
created in a secondary nearest neighbor structure as shown
in Fig. 5(a). Both poses share the same orientation, but the
secondary pose is translated forward by a parameterizable
amount in direction of their orientation (dashed lines). For
a new sampled pose (red point), the nearest neighbor search
is executed in the secondary space. But the new pose is
added and connected to the corresponding pose of the nearest
neighbor in the primary space as shown by the new black
point in Fig. 5(b). The orientation is set to the direction from
the nearest neighbor to itself and a corresponding pose for
the new sample is inserted into the secondary space. This
(a) (b) (c)
Fig. 5. Example of inserting a new pose into RRT∗with the Secondary
Nearest Neighbor Space extension. Poses with filled line are in the primary
space while poses with doted lines are in the secondary space.
process has been inverted to realize the reconnection step
3To consider the current robot configuration, the maximum height difference
is automatically adapted to the length of the robot (number of modules).

2
of the RRT∗algorithm. Nearest neighbors are searched in
the primary space using the correspondences for poses of
the secondary space (red dashed line in Fig. 5(b)). Fig. 5(c)
illustrates the successfully conducted reconnection (red line)
and the movement of the secondary nearest neighbor to fit
the new orientation. All collision checks take place in primary
space that represents the real world.
E. Motion Control and Base Control
The generated motion trajectories are executed by the
Motion Control and mapped to the robot kinematics. A velocity
controller as an alternating control, switching between two
proportional controllers for acceleration and deceleration, is
used to slow down the robot in passages of fast changing joint
angles. Hidden through an hardware abstraction layer, either
the real robot hardware or a simulation is controlled by the
Base Control.
IV. IMP LE ME NTATIO N OF
MOTION PLANNING AND CONTRO L
A. Coarse and Precise Path Planner
The previously introduced coarse and precise planning
steps both use the extended RRT∗algorithm with different
parameters. Ground contact points of the four wheels in the
first and second drive module are considered in the cost
function to evaluate the sampled robot pose on the given map.
Each node of the RRT∗tree represents a 6-D pose of the center
point of the first joint module of KAIRO 3. Connecting two
nodes is conducted by a simple line check whilst setting the
orientation of the child node to the direction obtained from its
parent. Thus, in the RRT∗reconnection step, only the motion to
the direct child state has to be checked. Based on the generated
RRT∗tree, a path to the next intermediate target is transferred
to the Joint Angle Planner.
B. Joint Angle Planner
The Joint Angle Planner uses the path of 6-D poses given
by the precise path planning and creates intermediate poses.
For every 6-D pose, the height is calculated by translating
the pose to the wheel contact points of the first drive module.
Height smoothing is performed using a filter chain that consists
of the following elements:
1) Moving median filter
2) Relative threshold filter
3) Linear interpolation of discrete steps
4) Moving mean filter
Because the virtual rail runs through the center of each joint
module, a forward translation is performed. This, results in an
overshoot on top of a step, which turns the robot smoothly
around the corner (Fig. 6). The same appears during descend-
ing from an obstacle which is captured by an increase of the
height in those poses. Once all height values are adjusted,
the desired joint angles alpha,beta and gamma of the first
joint module are calculated using the inverse robot kinematics.
Thereby, only the first module has to be considered, because
all other modules follow on the same path with the same joint
angles. The resulting path of 6-D poses is enriched with joint
angles for alpha,beta and gamma at each pose and then
given to the Motion Control.
C. Motion Control
1) Virtual Rail Extension: The virtual rail of the Adaptive
Control is used to execute short sections of the planned path.
Thus, the joint angles are written into rail base points which
are specified by 6-D poses. Joint angles for each joint module
can be calculated via linear interpolation of the joint module
center on the rail.
2) Drive Speed Controller and Wheel Velocities: To allow
the joints to follow the given trajectory a drive velocity
controller is used to regulate the speed of the robot. It is im-
plemented as an alternating control between two proportional
controllers. The wheel velocities are calculated by a special
differential drive function which depends on the neighboring
modules to reduce slippage in curves.
V. EVALUATION AND TESTS
Evaluation was conducted with KAIRO 3, introduced in
Section II. In the following, three different test scenarios are
explained and the results of our experiments are presented.
A. Scenario 1: Climbing Stairs
To evaluate the capabilities of our approach for au-
tonomously guiding a wheeled snake-like robot over stairs,
we used the test scenario shown in Fig. 7 on the top left.
KAIRO 3 was placed about 1 m in font of the stairs (see Fig. 7
top left) and a navigation goal (target pose) was manually
specified 4 m straight ahead of the robot. The following steps
were all processed and executed autonomously by KAIRO 3
itself. First of all, one 3-D point cloud has been taken with the
3-D laser KaRoLa at the starting position. Then, KAIRO 3
planned and conducted the mapping maneuver illustrated in
Fig. 7 (top right) and took a second 3-D scan. Both point
clouds were merged together into a PlexMap, which has been
used by the navigation planning to generate the joint angle
trajectory shown in Fig. 6. This joint angle trajectory was
Fig. 6. Scenario 1: Generated PlexMap with coarse, precise and joint angle
trajectories.
piecewise translated into the virtual rail by the Motion Control.
The remaining frames of Fig. 7 illustrate how KAIRO 3 moved
along the virtual rail to autonomously climb up the stairs and
stopped at the given target pose. In this scenario, KAIRO 3
was able to reach the target pose autonomously in four
of the five conducted test runs. Faults mainly arise due to
robot odometry drifts and not accurately matched point clouds
(overlap between point clouds was too small). Most of the
localization errors have been corrected after performing a new
3-D scan.

2
Fig. 7. Scenario 1: Sequence showing KAIRO 3 climbing up the stairs.
B. Scenario 2: Overcoming a Box
Besides the capability of moving up stairs, we also evalu-
ated the behavior of overcoming larger obstacles. A box with
42 cm height and 56 cm depth was utilized as obstacle in
the second scenario. Again KAIRO 3 was placed about 1 m
in front of the obstacle and the target pose was set behind
the box. Preventing the planning algorithm from generating a
path around the box, obstacles higher than the box have been
placed besides. The resulting coarse, precise and joint angle
trajectories are shown in Fig. 8. Fig. 9 illustrates the motion
Fig. 8. Scenario 2: Large box for testing the overcoming of obstacles
especially the descending into previously unknown areas.
of KAIRO 3 by an image sequence. Moving on top of the box
worked fine every time, but only one out of three test runs
could be finished successfully. Problems of descending from
the box mainly arose from two different causes:
1) Localization errors due to slipping of the wheels on
the plastic surface of the box.
2) Difficulties with capturing the floor behind the box
correctly: If the scan was triggered too early, the map
looked like a ramp instead of a sharp decline.
An entire planning process took in average 12 s on a 3.3 GHz
Intel i5 processor. This includes 5.5 s for coarse, 4.5 s for
precise and 2 s for joint angle planning.
Fig. 9. Scenario 2: Sequence showing KAIRO 3 moving over a box.
C. Scenario 3: Changing Robot Configurations
The third scenario aimed to evaluate the planning capabili-
ties for different robot configurations. With regard to KAIRO 3
this means varying the amount of modules as shown in Fig. 3.
Next to a 6 m long ramp a large step with a height of 54 cm,
as illustrated in Fig. 10 (top left), has been prepared. For
Fig. 10. Scenario 3: KAIRO 3 with eleven modules was able to autonomously
move up a 54 cm step.
each experiment KAIRO 3 was placed 1.5 m in front of the
step parallel to the ramp. The first configuration consisted of
eleven modules (six drive and five joint modules), the second
used only five modules (three drive and two joint modules).
As shown in the image sequence of Fig. 10, KAIRO 3 au-
tonomously captured the environment, planned a trajectory,
moved up the large step and reached the target pose.
With a configuration of only five modules the planning
system also successfully generated a trajectory to the target

2
Fig. 11. Scenario 3: Planning approach generated a trajectory for KAIRO 3
with five modules using the ramp to reach the goal. The generated trajectories
are visualized on the left and the RRT∗tree build from the coarse planning
is shown on the right.
pose as shown in Fig. 11 (left). The planned robot trajectory
conducted a U-turn and moved to the ramp for overcoming
the height difference to the target. Furthermore, Fig. 11 (right)
shows that the kinematics constraints of our nonholonomic
robot are considered correctly by the used RRT∗algorithm.
VI. CONCLUSIONS
In this work we introduced an adaptive variation of the
RRT∗algorithm, in combination with a joint angle planner for
reconfigurable snake-like robots. Our planning system has been
implemented with the three stages coarse, precise and joint
angle planning, to generate and optimize complex motions
using 6-D trajectories. The planning time remains unaffected
by an increased amount of modules due to a reduced planning
space: The planning of 6-D poses for the first module is
combined with the usage of the follow-the-leader approach.
Especially for path planning of nonholonomic robots, we
introduced an extension of the RRT∗algorithm to improve
the connectivity of pose samples for better planning results
and speed. The planning system also adapts to different robot
configurations and generate paths regarding the capabilities of
the robot. In combination with the Adaptive Control based on
the virtual rail, this results in a robust planning and control
system. The proposed methods has been evaluated with our
wheeled snake-like robot KAIRO 3 in three different scenarios.
Our results show that KAIRO 3 is able to autonomously move
up stairs and steps which are much higher than the robot itself.
There is still potential for further investigations and im-
provements. The evaluation of more sophisticated constraints
like considering reconfiguration costs has to be done in order
to ensure correct planning for multiple reconfigurable robots.
Although the system is managed in a hierarchical structure,
the planning speed could be reduced even more by optimizing
the algorithms. In order to prevent the robot from falling over,
the already existing stability calculation will be included into
the planning process and also be used for detection of critical
situations.
REF ER EN CE S
[1] J.-C. Latombe, “Robot Motion Planning”, in Springer Int. Series in
Engineering and Computer Science, Springer, 1991.
[2] S. M. LaValle, “Planning Algorithms”, Cambridge, U.K.: Cambridge
University Press, 2006, available at http://planning.cs.uiuc.edu/.
[3] B. Siciliano, L. Sciavicco et al., “Robotics: Modeling, Planning, and
Control”, in Advanced Textbooks in Control and Signal Processing,
Springer, 2009.
[4] S. A. M. Coenen, “Motion Planning for Mobile Robots - A Guide”.
[5] D. Tsakiris, M. Sfakiotakis et al., “Polychaete-like Undulatory Robotic
Locomotion”, in IEEE Int. Conf. on Robotics and Automation, 2005.
[6] Z. Bayraktaroglu, A. Kilicarslan et al., “Design and Control of Biolog-
ically Inspired Wheel-Less Snake-Like Robot”, in IEEE Int. Conf. on
Biomedical Robotics and Biomechatronics (BioRob), Feb 2006.
[7] J. Conradt and P. Varshavskaya, “Distributed Central Pattern Generator
Control for a Serpentine Robot”, in Int. Conf. on Artificial Neural
Networks (ICANN), 2003.
[8] A. M. Bloch, P. S. Krishnaprasad et al., “Nonholonomic Mechanical
Systems with Symmetry”, in ARCH. RATIONAL MECH. ANAL, 1996.
[9] G. Chirikjian and J. Burdick, “An Obstacle Avoidance Algorithm for
Hyper-Redundant Manipulators”, in IEEE Int. Conf. on Robotics and
Automation (ICRA), pp. 625–631, vol. 1, May 1990.
[10] S. Hirose and M. Mori, “Biologically Inspired Snake-like Robots”, in
IEEE Int. Conf. on Robotics and Biomimetics (ROBIO), Aug 2004.
[11] G. S. Chirikjian and J. W. Burdick, “The Kinematics of Hyper-
Redundant Robot Locomotion”, in IEEE Trans. on Robotics and Au-
tomation (ICRA), pp. 781–793, vol. 6, Dec 1995.
[12] J. B. Grzegorz Granosik and Malik G. Hansen, “The OmniTread
Serpentine Robot for Industrial Inspection and Surveillance”, in An Int.
Journal of Industrial Robot, vol. 32, 2005.
[13] G. Granosik, K. Mianowski and M. Pytasz, “Wheeeler – Hypermobile
Robot”, in Research and Education in Robotics – EUROBOT 2008.
Springer Berlin, Heidelberg, vol. 33, pp. 68–83, 2009.
[14] T. Kamegawa, T. Yamasaki and F. Matsuno, “Evaluation of Snake-Like
Rescue Robot ”KOHGA” for Usability of Remote Control”, in IEEE Int.
Workshop on Safety, Security and Rescue Robotics, 2005.
[15] L. Pfotzer, S. Ruehl et al., “Kairo 3: A Modular Reconfigurable Robot
for Search and Rescue Field Missions”, in IEEE Int. Conf. on Robotics
and Biomimetics (ROBIO), pp. 205–210, Dec. 2014.
[16] K.-U. Scholl, “Konzeption und Realisierung einer Steuerung f ¨
ur vielseg-
mentige, autonome Kanalroboter”, in Verl. f¨
ur Wissenschaft und Kultur
W iKu −V erl., Stein, Berlin, 2003.
[17] S. Yamada and H. Hirose, “Development of Practical 3-Dimensional
Active Cord Mechanism ACM-R4”, in Journal of Robotics and Mecha-
tronics, vol. 18, pp. 305-311, 2006.
[18] G. Granosik and M. Pytasz, “Control Methods for Wheeeler – The
Hypermobile Robot”, in Robot Motion and Control 2011, Springer
London, vol. 422, pp. 27–37, 2012.
[19] B. Murugendran, A. Transeth and S. Fjerdingen, “Modeling and Path-
Following for a Snake Robot with Active Wheels”, in IEEE/RSJ Int.
Conf. on Intelligent Robots and Systems (IROS), Oct 2009.
[20] C. Altafini, “Some Properties of the General n-Trailer”, in Int. Journal
of Control, vol. 74, no. 4, pp. 409–424, March 2001.
[21] K. Uhl and M. Ziegenmeyer, “MCA2 – An Extensible Modular Frame-
work for Robot Control Applications”, in Int. Conf. on Climbing and
Walking Robots. World Scientific, pp. 680–689, 2007.
[22] L. Pfotzer, S. Bohn et al., “Biologically-Inspired Locomotion with
the Multi-Segmented Inspection Robot Kairo-II”, in Proceedings of
CLAWAR2011, p. 181, Sept. 2011,
[23] J. Oberlander, S. Klemm et al., “A Multi-Resolution 3-D Environment
Model for Autonomous Planetary Exploration”, in IEEE Int. Conf. on
Automation Science and Engineering (CASE), Aug 2014, pp. 229–235.
[24] C. Birkenhofer, “Adaptive Steuerung eines mehrsegmentigen Inspek-
tionsroboters”, Ph.D. Dissertation, FZI Karlsruhe, 2010.
[25] L. Pfotzer, J. Oberlaender et al., “Development and Calibration of
KaRoLa, a Compact, High-Resolution 3D Laser Scanner”, in IEEE Int.
Sym. on Safety, Security and Rescue Robotics (SSRR), Oct 2014.
[26] S. M. LaValle, “Rapidly-Exploring Random Trees: A New Tool for
Path Planning”, Department of Computer Science, Iowa State University,
Tech. Rep., 1998.
[27] S. M. LaValle and J. J. Kuffner, “Randomized Kinodynamic Planning”,
in Int. Journal of Robotics Research, vol. 20, pp. 378–400, May 2001.
[28] A. H. Qureshi, S. Mumtaz et al., “Adaptive Potential Guided
Directional-RRT”, in IEEE Int. Conf. on Robotics and Biomimetics
(ROBIO), pp. 1887–1892, Dec 2013.