Distributed control and navigation system for quadrotor 
UAVs in GPS-denied environments 
Konstantin Yakovlev, Vsevolod Khithov, Maxim Loginov, Alexander Petrov 
Institute for Systems Analysis of Russian Academy of Sciences, Moscow, Russia 
yakovlev@isa.ru 
 Soloviev Rybinsk State Aviation Technical University, Rybinsk, Russia 
vskhitkov@gmail.com 
 Soloviev Rybinsk State Aviation Technical University, Rybinsk, Russia 
lunarstrainut@gmail.com 
NPP SATEK plus, Rybinsk, Russia 
gmdidro@gmail.com 
Abstract. The problem of developing distributed control and navigation system 
for quadrotor UAVs operating in GPS-denied environments is addressed in the 
paper. Cooperative navigation, marker detection and mapping task solved by a 
team of multiple unmanned aerial vehicles is chosen as demo example. Devel-
oped intelligent control system complies with on 4D\RCS reference model and 
its implementation is based on ROS framework.  Custom implementation of 
EKF-based map building algorithm is used to solve marker detection and map 
building task. 
Keywords: intelligent control system, distributed architecture, 4D/RCS, visual 
navigation, marker detection, SLAM, map building, ROS, AR.Drone 
1 
Introduction 
Modern intelligent systems are characterized by increasing autonomy and distribu-
tion. From one side autonomy increases in a sense of behavior independence from 
operator and from other side, in case of multiagent system, its autonomy could be 
characterized by degree of automation in agents collaboration. Modern systems be-
come distributed both in terms of ensuring a collective interaction of several intelli-
gent agents (the distribution of knowledge and information between agents), and in 
terms of a distributed software technologies.  
We present distributed control and navigation system for unmanned aerial vehicles 
(UAV), e.g. quadrotors, in GPS-denied environments. We demonstrate architecture 
and implementation of our system on map building task solved by the collaboration of 
UAVs. The task is following – N quadrotors (we use AR.Drones in our experiments) 
are operating in a GPS-denied environment, e.g. a room, and are remotely controlled 
(via wi-fi link) by software control system on ground station. There are several mark-
ers, that placed somewhere in the room (we use cubes with QR-codes on their sides) 
and AR.Drones should firstly detect markers then compute markers’ position relative 
to drone’s start position (individual map) and after that integrate and elaborate shared 
map. 
In this paper first we discuss conceptual framework for layered representation of 
intelligent system. Next we will focus on marker detection, drone and marker position 
estimation and map building algorithms. Next we describe control system software 
implementation based on ROS framework [1]. In conclusion we present experimental 
evaluation results and future direction of our work. 
1.1 
Related works 
 
Previous research in distributed robot control system includes: [2] which propose 
ALLIANCE fault-tolerant cooperative control architecture. ALLIANCE does not 
require any use of negotiation among robots, but rather relies upon broadcast messag-
es from robots to announce their current activities. [3] poses solution for multiple 
robotic vehicles motion planning in terms of classical control theory (work is devoted 
to prediction of input/output reachability, structural observability, and controllability 
of the multiagent system). [4] describe CAMPOUT architecture for distributed control 
algorithms within NASA planetary surface rover systems. All sequencing in 
CAMPOUT is done through Finite State Machine for deterministic control. In [5] 
main features of cooperative grasping and transport control system for multiple quad-
rotors are illuminated, the main attention is paid to control signal generation for trans-
portation of some payload simultaneous by multiple drones. 
Article [6] proposes using adopted Bayesian approach for location estimation of 
unknown target in known, but complex environment by multiple robots. The target 
detection and localization is done in radio medium. Control strategy applied in [6] to 
solve the task is based on reduction in the uncertainty of the target estimate principle.  
Previous research in cooperative map building algorithms and systems includes: 
[7], which considers the problem of cooperative navigation and mapping by a hetero-
geneous team of multiple autonomous underwater vehicles. In [7] authors address a 
form of cooperative Simultaneous Localization and Mapping (SLAM) in which only 
one vehicle is responsible for maintaining estimates of the map and poses for each 
robot. By combining inter-vehicle measurements with observations of the environ-
ment made by each vehicle, the result is a better knowledge of the poses of each robot 
in the group. In [8] describes a multi-robot map merging algorithm with the Fast-
SLAM method and its evaluation on experimental map with landmarks. The article 
[9] proposes a novel extension to incremental smoothing and mapping (iSAM) algo-
rithm, that facilitates multirobot mapping based on multiple pose graphs[10]. The 
article [11] describes a framework for cooperative 3D mapping of unstructured envi-
ronments, which utilizing AR.Drone as UAV and combination of markerless and 
landmark SLAM algorithms. 
To sum up above said, there exists numerous related works which are devoted to 
different aspects of distributed control system and cooperative map building exist 
nowadays. But in contrast with our work part of them describe 2D case, part of them 
lack of conceptual framework and architecture model, and others don’t use ROS 
framework (or at least don’t mention that in text). Yet another difference from most of 
the related work is usage of the low cost AR.Drone UAV which is turn poses some 
restriction on quality of income sensor data. 
2 
Multilevel architecture of distributed control and navigation 
system for quadrotors  
Different approaches to design the architecture of the UAV control system exist 
nowadays. Some approaches suggest the usage of flat (one-level) architecture consist-
ing of separate interacting modules tailored to solve different tasks. An example of 
such architecture is described in [12] and the identified tasks (and corresponding 
modules) are: mission planning, collaboration, contingency management, situational 
awareness, communications management, air vehicle management. Other approaches 
(more numerous) suggest splitting the functionality not only between different mod-
ules (abstracting the functionality itself) but also between different levels or layers 
(abstracting the level of “deliberativeness”) each of which consists of bundle of mod-
ules tied together. Examples of the most known multi-leveled architectures for intelli-
gent agents operating in real world are 3T [13], ATLANTIS [14], Aura [15] and oth-
ers. Typically in robotics (as well as in UAV design) control system consists of 3 
levels: deliberative (or strategic) level (top), reactive level (bottom) and intermediate 
(tactical) level. Top level contains modules that deal with computational expensive 
(AI) reasoning and planning. Intermediate level deals with analysis of spatial data and 
navigation tasks (e.g. SLAM, path-planning and others). Low level controls sensors 
and actuators (often called subsystems) of the UAV [16]. We follow that approach 
and decompose our control system into 3 above mentioned levels. 
In our experiments we use Parrot AR.Drone platform [17] and do not override ex-
isting modules of reactive control implemented by Parrot and described in [18]. Each 
AR.Drone has it’s own built-in flight controller which automates the execution of 
basic flight maneuvers: take off, land, fly forward/backward, strafe left/right, turn 
clockwise/counterclockwise. Commands to execute these maneuvers are sent via the 
open data exchange protocol from the ground station. Also we do not implement stra-
tegic level of the control system as only one high-level task, e.g. “build-map”, is ad-
dressed.  As a result we mainly concentrate on the design of tactical layer capable of 
solving navigation tasks (e.g. identifying features and building a map of them). To do 
so we follow 4D/RCS reference model [19]. The main idea that lies in the basis of this 
model is abstract functional decomposition, e.g. each level of the architecture is com-
posed of nodes which abstract different instances of controllable subjects and each 
node is composed of 4 identical functional processes (which can be viewed as bundles 
of modules): behavior generation, world modeling, sensory processing, value judg-
ment (see figure 1).  
 
Fig. 1. 4D/RCS Node (from [19]). 
In our case 4D/RCS processes are interpreted as following: 
SP – receiving (via wi-fi link) and transcoding videostreams from two AR.Drone’s 
cameras (forward looking and downward looking), receiving (via wi-fi link) data 
regarding UAV state and orientation (AT-packets contacting measurements of the 
internal navigation system). 
VJ – AR.Drone pose estimation, recognition and identification of visual markers, 
computing the distances between UAV and the markers, computing position of the 
markers relative to local and global coordinate systems. 
WM – constructing local map, integrating multiple local maps into the global 
(shared) one. 
BG – generating rules for choosing next flight destination. 
Each 4D/RCS node is in charge of controlling a single UAV. Functional processes 
are implemented as a bunch of Robotic operating system [20] modules (see detailed 
description in section 4) executed in parallel on ground station (we use standard 
Linux-run laptop). Global knowledge database is maintained by the system, e.g. each 
AR.Drone builds it’s own local map which is stored on the ground station, at the same 
time all the gained maps are processed and a global map consistent with the meas-
urements of all UAVs is constructed and an access to it is granted to all drones. De-
tails of pose estimation, markers identification, map building procedures are described 
further as well as the details of software implementation of the system. 
3 
Map building algorithm 
Implemented map building algorithm constructs a map of markers’ locations as a 
simplified abstraction of more general map building task, that, however, can be used 
for practical applications where visual marker based navigation is applicable. A map 
is constructed simultaneously by several drones observing different subsets of mark-
ers placed in corresponding parts of common environment and merging them into a 
global map. This is done in a following manner… 
Each UAV performs a SLAM algorithm based on both inertial and visual data 
fused with Extended Kalman Filtering (EKF) while being in flight. During that, 
markers are observed and identified uniquely (they all are supposed to be unique) and 
their relative poses are calculated by each drone using vision. Any time a drone ob-
serves a marker, it checks it’s existence on the global map which storage is shared 
appropriately. Locations of markers that appear to be already known are used to cor-
rect current UAV position state. New markers are added to the global map and their 
positions are fused with the next few observations to reduce inaccuracy of visual pose 
measuring. As subsets of markers visible by different UAV’s may not overlap for a 
long time, these markers are stored linked to independent coordinate frames until 
some marker or a group is observed by two or more drones, then coordinate systems 
are merged, and translation of markers’ poses from eliminated coordinate frame to 
conserved one is performed. As markers’ total count is assumed to be relatively small, 
bundle adjustment from a set of keyposes like in [21] is used to unbias the map. 
We utilize ArUco [22] marker system and library capable of detection, recognition 
and 6-DOF pose estimation of up to 1024 different bar-codes, distinctive reliably with 
the help of modified Hamming code. With the camera calibrated and known marker 
parameters it becomes easy to build the map in meter units. 
We use Extended Kalman Filtering for drone pose estimation with state vector  ht = 
(xt, yt, zt, αt, βt, γt)T, where elements correspond to 3 spatial coordinates and 3 Euler 
angles in global frame (totally global or one of the independent global ones) corre-
spondingly. Odometry observation and prediction models are similar to that described 
in [23]. Pose observation by means of vision relies on relative pose estimation of all 
visible markers with known positions at some moment. Given observed marker pose, 
described as translation vector Tm and rotation matrix Rm in camera coordinate sys-
tem, we first calculate approximate measurement noise covariance matrix (assuming 
measured values along 3 axes independent for simplicity). Then we use observed 
relative marker coordinates to calculate UAV’s global pose observation; the covari-
ance is combined with the covariance assigned to the marker on the map. The EKF 
state is then updated using obtained pose observation and resulting covariance. The 
EKF state is updated with the information obtained from each of observed already 
known markers. 
Merging coordinate systems of drones when they observe the same markers for the 
first time is done in the following manner. Firstly RT matrix is calculation as 3D af-
fine estimation based on matching markers, then this matrix is used to transform 
markers poses from one drone coordinate system to other drone coordinate system. 
After that if a new matching markers have found, RT matrix is refined and markers 
poses are transformed again. 
4 
Software implementation of distributed control and 
navigation system for quadrotors based on ROS 
Software implementation of described system is based on ROS framework and 
4D/RCS scheme. Structure of software is shown on figure 2. 
 
 
Fig. 2. – Structure of distributed control system software implementation based on ROS 
ROS packages, nodes and also external libraries are shown on figure 2. There are 
two main ROS nodes: “Navpts” node and “Ground station” node. Several instances of 
Navpts node are executed simultaneously each of which controls a single AR.Drone. 
Ground station node implements map building and sharing algorithm and only one 
instance of that node is executed at a time. If we describe Navpts node following 
4D/RCS we could say that OpenCV library relates to sensory processing, Aruco and 
EKF packages to value judgment, marker searching control package to behavior gen-
eration. 4D/RCS process of world modeling process relates to distributed between 
Navpts and Ground station nodes map representation. 
Ground station node consists of map builder module, which implement described 
algorithms, and node manager, that start and remap necessary topics for Navpts 
nodes. 
For map visualization we use Rviz package of ROS. 
5 
Evaluation and conclusions 
We evaluate our system implementation in laboratory environment. The result is 
shown in form of merged map on figure 3 (right). 
 
 
Fig. 3. Marker detection on image (left) and merged map in Rviz (right) 
On figure 3 (left) experimental setup and detection of one marker is shown. On 
figure 3 (right) shared map is shown and trajectories of drones and position of com-
mon markers are indicated.  
As a conclusion we could point out, that our implementation of map building isn't 
optimal solution for such task compared to related works and we use it like an illustra-
tion of chosen architecture and software implementation of distributed control and 
navigation system for quadrotors in GPS-denied environments. We plan to extend our 
system in the following direction: integrate with more powerful SLAM algorithm, 
which will be applicable in environment with unknown landmarks, and modify be-
havior generation process by introduction cooperative planning to struggle with situa-
tion in which one of drones couldn’t detect any markers and to increase efficiently of 
the system. 
Acknowledgments. This work was supported by the Ministry of Education and Sci-
ence of the Russian Federation (№ 14.577.21.0030 agreement for a grant on "Con-
ducting applied research for the development of intelligent technology and software 
systems, navigation and control of mobile technical equipment using machine vision 
techniques and high-performance distributed computing"). 
References.  
 
1. Robot Operation System, http://www.ros.org 
2. Parker, L.E. 1998b. Distributed control of multi-robot teams: Cooperative baton-passing 
task. In Proceedings of the 4th International Conference on Information Systems Analysis 
and Synthesis (ISAS' 98), Vol. 3, pp. 89–94. 
3. J. Feddema, C. Lewis, and D. Schoenwald,  "Decentralized control of cooperative robotic 
vehicles: Theory and application",  IEEE Trans. Robot. Automat.,  vol. 18,  pp.852 -864 
2002 
4. T. L. Huntsberger, A. Trebi-Ollennu, H. Aghazarian, P. S. Schenker, P. Pirjanian, and H. 
D. Nayar, "Distributed control of multi-robot systems engaged in tightly coupled tasks" 
Autonomous Robots, vol. 17, no. 1, pp. 79-92, 2004 
5. Michael, N., Fink, J., & Kumar, V. (2011) Cooperative grasping and transportation using 
multiple quadrotors. In International Symposium on Distributed Autonomous Robotic Sys-
tems (DARS), Lausanne, Switzerland   
6. B. Charrow, N. Michael, and V. Kumar. Cooperative multirobot estimation and control for 
radio source localization. In Proceedings of International Symposium on Experimental 
Robotics, 2012 
7. Walter, M., Leonard, J.: An experimental investigation of cooperative SLAM. In: 5th In-
ternational Symposium on Intelligent Autonomous Vehicles, Lisbon, 5–7 July 2004 
8. N. E. Ozkucur and H. L. Akin, "Cooperative multi-robot map merging using fast-slam," in 
Proc. RoboCup International Symposium 2009. Springer-Verlag, jun 2009, pp. 449-460 
9. Been Kim et al. “Multiple Relative Pose Graphs for Robust Cooperative Mapping.” IEEE 
International Conference on Robotics and Automation (ICRA), 2010. 3185–3192. © Cop-
yright 2010 IEEE 
10. M. Kaess, A. Ranganathan, and F. Dellaert, “iSAM: Incremental smoothing and mapping,” 
IEEE Trans. Robotics, vol. 24, pp. 1365–1378, Dec 2008. 
11. P. S. Nekkundi and S. Dulman, "A Framework for Cooperative 3D Mapping of Unstruc-
tured Environments", Master Thesis,  Delft University of Technology (2011) 
12. Jameson S., Franke J., Szczerba R., Stockdale S. Collaborative Autonomy for 
Manned/Unmanned Teams. AHS International Forum 61. Grapevine. TX. 2005. 
13. Bonasso, R.P., et al. Experiences with an Architecture for Intelligent, Reactive Agents. in 
International Joint Conference on Artificial Intelligence. 1995. 
14. Gat, E. Integrating planning and reacting in a heterogenous asynchronous architecture for 
controlling real-world mobile robots. in National Conference for Artificial Intelligence. 
1992 
15. Arkin, R. Motor schema based navigation for a mobile robot: An approach to program-
ming by behavior. in IEEE International Conference on Robotics and Automation. Pro-
ceedings. 1987. 
16. Freed, M., et al., An Architecture for Intelligent Management of Aerial Observation Mis-
sions. AIAA, 2005. 2005-6938. 
17. Ar.Drone 2.0, http://ardrone2.parrot.com/ 
18. Bristeau, P. J., Callou, F., Vissière, D., & Petit, N. The navigation and control technology 
inside the AR.Drone micro UAV. In 18th IFAC World Congress (Vol. 18, No. 1, pp. 
1477-1484). 2011. 
19. Albus, J., Huang, H. M., Messina, E., Murphy, K., Juberts, M., Lacaze, A., & Finkelstein, 
R. 4D/RCS Version 2.0: A reference model architecture for unmanned vehicle systems. 
National Institute of Standards and Technology, Gaithersburg, MD, NISTIR, 6912. 2002. 
20. Quigley, M., Conley, K., Gerkey, B., Faust, J., Foote, T., Leibs, J., ... & Ng, A. Y. ROS: an 
open-source Robot Operating System. In ICRA workshop on open source software (Vol. 3, 
No. 3.2). 2009. 
21. G. Klein and D. Murray, “Parallel tracking and mapping for small AR workspaces. In In-
ternational Symposium on Mixed and Augmented Reality”, 2007 
22. R. Munoz-Salinas, S. Garrido-Jurado, Aruco library, http://sourceforge.net/projects/aruco/ 
23. J. Engel, J. Sturm, D. Cremers. Camera-Based Navigation of a Low-Cost Quadrocopter. In 
Proc. of the International Conference on Intelligent Robot Systems (IROS), 2012