20th IMEKO TC4 International Symposium and

18th International Workshop on ADC Modelling and Testing

Research on Electric and Electronic Measurement for the Economic Upturn
Benevento, Italy, September 15-17, 2014

Autonomous navigation of mobile robots: from

basic sensing to problem solving
Paolo Tripicchio1 , Massimo Satler1 , Carlo Alberto Avizzano2 , Massimo Bergamasco2
1
Gustavo Stefanini advanced robotics research center, Scuola Superiore Sant’Anna, La Spezia
(Italy), [email protected],[email protected]
2
PERCeptual RObotics laboratory, Scuola Superiore Sant’Anna, Pisa (Italy),
[email protected],[email protected]
Abstract – Autonomous navigation is a complex task
that requires both sensing capabilities to react to sud-
den environmental changes or map the environment
and reasoning to schedule the next action to perform.
Starting from basic sensing technology used in the ma-
jority of mobile robotic systems, the introduction of
sensor fusion techniques allows to obtain useful infor-
mation to solve the localization, mapping and naviga-
tion problems. Applications of these methods to achieve
specific robot capabilities will be presented starting
from object detection and recognition, passing to scene
classification and ending with an industrial related ap-
plication: the visual inspection of industrial facilities by
means of a flying vehicle.

I. INTRODUCTION Fig. 1. Robot navigating in outdoor environment recogniz-

In the last years the field of mobile robotics is transition- ing driving lane and moving people.
ing from user-teleoperated systems to completely or semi-
autonomous systems. An autonomous robot is a robot that
performs behaviors or tasks with a high degree of auton- quires several proprioceptive and exteroceptive sensors.
omy, which is particularly desirable in cases like environ- Proprioception, or sensing one’s own internal status,
ments exploration, in activities that are potentially danger- is required for robots to work autonomously near people
ous for humans, or again in applications where the robot and in harsh environments. Common proprioceptive sen-
performs better or in a shorter period of time. sors include thermal, optical, electrical and haptic sens-
To achieve autonomous navigation and intelligent be- ing. Exteroception is sensing things about the environ-
havior in robots, it is necessary to solve specific prob- ment. Autonomous robots must have a range of environ-
lems: starting from the well-known mobile robotics ques- mental sensors to perform their task while keeping an ad-
tions ”Where am I?” and ”How can I reach a specific loca- equate level of knowledge and safety. Common extero-
tion?”. Such questions are related to the localization, map- ceptive sensors include the electromagnetic, sound, touch,
ping, relocation and path planning problems that require chemical, temperature, range, pressure, and altitude sen-
reactive behaviors of the robots like collision avoidance, sors. All these sensors cannot work indipendently but their
location reaching, following a target and so on and robot measures should be fused to obtain good estimation of en-
capabilities like moving objects detection, people recogni- vironmental and internal variables.
tion, place classification, etc. The paper presents possible approaches to solve the au-
In general a fully autonomous robot should be able to tonomous navigation problems and the specific solutions
gain information about the surrounding environment, work to obtain the required robotic capabilities. Section ii. will
for an extended period of time without human intervention introduce a short state of the art on localization and map-
and possibly avoid situations that are harmful to people, ping techniques introducing both indoor and outdoor nav-
property, or itself. An autonomous robot should also be igation. In section iii. a possible approach to obtain au-
able to learn or gain new knowledge to adapt to surround- tonomous navigation will be introduced. Section iv. will
ings changes. To fulfill all these capabilities, the robot re- present specific inplementations to achieve the required
sensing capabilities. The last section will present the final

ISBN-14: 978-92-990073-2-7 1
considerations and conclusions. III. POSSIBLE APPROACH
In many applications, the robot should be able to move
II. STATE OF THE ART independently in a semi-structured environment such as an
To obtain autonomous navigation, the robot requires to office, store or house. In most cases it is useful that the
solve the localization problem and to have at least the ca- robot moves through a series of waypoints or to a specific
pability to navigate point-to-point. Starting in the 1970s point of the environment without knowing the obstacles
with wire-guidance and progressed in the early 2000s to present in the environment. To achieve this goal we pro-
beacon-based triangulation, current state of the art robots pose a hybrid software architecture that consists of a reac-
autonomously navigate based on sensing natural features. tive and a deliberative part. The reactive part controls the
At first, autonomous navigation was based on planar sen- local navigation responding to obstacles that are found in
sors, such as laser range-finders, that can only sense at one the environment and to the commands of the deliberative
level. Nowadays, most advanced systems fuse informa- part that knows all the waypoints and navigation controls
tion gained from various sensors for both localization (po- globally.
sition) and navigation purposes. Sensor fusion is the key The deliberative architectures have historically been the
technique to obtain precise positioning and good mapping first to be used for the control of autonomous robots. They
capabilities as well as exact scene understanding. Many are based on a functional approach, which expects to ac-
techniques are used in literature for the position estimation. quire an ideally complete knowledge of the environment in
Basically they fuse the odometric information, the iner- order to build an accurate model on which to plan a series
tial measurements and the range sensor readings by means of actions that lead to the target. The goal of our robot is
of a state observer or Kalman filter. Algorithms based to reach a point on the map, while the actions that we have
on scan matching [18] employ all the data coming from to plan are a sequence of movements of the robot. The part
range scanners to generate a local map and then measure of the system that generates the trajectory starting from the
the motion occurred between consecutive scanned maps internal representation of the environment is called plan-
computing the minimum error coordinate transformation. ner. The algorithms used to implement the planner gen-
SLAM (Self Localization And Mapping) algorithms based erally explore a tree of trajectories, excluding those who
on grid mapping [8] make uses of a small number of parti- violate the physical constraints (e.g. obstacles), mechani-
cles which contain each a map represented as an occupancy cal (e.g., radii of curvature too tight) or performance (e.g.,
matrix. These maps represent free and occupied space on maximum time, distance covered). The trajectory gener-
a probabilistic base. Feature based SLAM algorithms [7] ated can also be optimum with respect to a cost function,
work on high level information (i.e. features) that are few which is generally the minimum time or minimum covered
in number with respect to sensors acquisitions and thus the space. If the environment in which the robot moves is dy-
computational load is reduced. The map representation namic, the path must be recalculated whenever a change
is reduced as well. More recently, visual odometry algo- occurs, which happens at each cycle in the case of moving
rithms [17] are based on feature extraction and on the as- robot in dynamic unknown environments. The computing
sociation of these features between successive frames. The power needed to explore a tree of possible trajectories is
identification of such features is performed with vision al- very large. The computers will be big and therefore ex-
gorithms like SURF [5] or SIFT [6] that identifies points pensive in terms of energy consumed. Generally an entire
or points clusters with an associated descriptor. Analyz- computer is dedicated only to trajectory planning [9], [13].
ing the motion of these features it is possible to estimate To reduce the dimension of the solution two planners can
the motion of the camera and at the same time of the robot be employed, one long and one short-term: the first deals
rigid body. with creating a series of waypoints up to the final goal,
Moving to outdoor exploration other issues arises. In the second generates the path to reach the next waypoint.
fact outdoor autonomy is more difficult for ground ve- When changes occur in the world only the second planner
hicles, due to the presence of three-dimensional uneven must be invoked [12]. The main advantage of using a plan-
terrains, disparities in surface density and different and ner with respect to any other architecture is the security
changing wheater conditions. In the contrary outdoor au- of a good result, a reactive architecture instead can indeed
tonomy is most easily achieved in the air, since obstacles vanish completely in balanced situations such as when a
are rare. In the case of outdoor navigation, the use of sen- target is behind to a dead end [11]. The main disadvantage
sor like GPS allows to obtain a good positional accuracy is the high computing power required that increases with
with ease during free space movements. For precise posi- the raise of the details of a map. The risk of relying on just
tioning however sensor fusion techniques have still to be a planner however is to let the robot out of control in the
used. case of collisions with new barriers along the old planned
route [14]. At the end of the 80s it was proposed a new
paradigm based on the concept of behavior. The consider-

2
Fig. 2. Automated Moving Object Detection and Tracking Algorithm(a). Example of detecting people that move in the
scene(b).

ations that led to this new kind of architecture were the in- poses some solution to acquire moving object and people
ability to model a complex world and the slow response of detection capability useful for collision avoidance, place
the robot to changes in the environment. A reactive robot recognition[16] and object recognition[10] that are use-
has only laws that define how to react to certain inputs of ful for scene understanding and reasoning and autonomous
the sensors, without the use of abstract representations or navigation capabilities in unstructured environments[15].
memory of past events and states. The intelligence of a re-
active robot is created by many behavior, coordinated with A. Moving objects and People detection
each other by some competitive or collaborative function. For object and people detection we use the software ar-
The disadvantages of this type of architecture is the lack chitecture shown in Figure 2 a.
of planning, which prevents an optimized navigation, and
In the MOG (Mixture Of K Gaussians) Background
a map, without which it is possible to get into dead ends or
Subtraction module the pixel color values are modeled as
to wander endlessly without ever achieve the goal.
a Mixture Of K Gaussians (µi , σi , ωi ) [2]. Multiple adap-
Choosing an hybrid solution allows to take advantage of tive Gaussians are necessary to cope with acquisition noise
the good characteristics of both the methods. and lighting changes. Pixel values that do not fit the back-
For the reactive part there are various approaches,a pos- ground distributions (Mahalanobis distance) are consid-
sible solution can be of using nonlinear control based on at- ered foreground. The model copes with multimodal back-
tractors and repulsors implemented with differential equa- ground distributions. The RGA Background Subtraction
tions, following the theory of dynamical systems[19]. with Selective Update module is based on the approach
The deliberative part instead will use modern machine proposed in [3] where the background at each (i; j) pixel
learning and AI techniques to perform reasoning on the location is modelled indipendently. The Running Average
action to accomplish and plan tasks esecution. technique is used for performing a more accurate Back-
Both the parts require specific knowledge acquired from ground Subtraction (BS). The Confidence Measurement
the environment and the robot internal state variables combines the results from BS and Color Clustering (CC)
[4]. It reduces false segmentation and provides more accu-
IV. SENSING COMPONENTS rate object boundary. The Object Tracking tracks objects
in successive frames to detect trajectories performing these
As discussed above, to perform autonomous navigation
tasks: finding correspondence, motion estimation, correc-
in an indoor or outdoor environment, some sensing com-
tive feedback and occlusion detection. The tracking is ro-
ponents are necessary in order to be able to execute spe-
bust by integrating feature vector template matching and
cific behaviors of the robot or to construct a correct rep-
Kalman filtering. An example is reported in Figure 2 b.
resentation of the environment map. This section pro-

3
 C. Object recognition
A similar approach has been used to solve the problem
of object recognition. The approach uses a stacked gen-
eralization classification method to perform a multi-layer
object recognition fusing heterogeneous spatial and color
data acquired with an RGB-D camera. The method as-
sumes a specific a-priori knowledge of the objects that may
be present in a scene. This tool exploits the accuracy of the
Fig. 3. Combination of classifiers in a two stage stack.
SVM introducing a multilevel architecture with a stacked
generalization approach. The main idea here consists in the
introduction of a second stage of generalization to deduce
B. Visual place recognition
the biases of the original generalizers. The second layer
We introduce a machine learning architecture to address input is the output of the first layer of generalizers, and its
the problem of scene or visual place classification in or- output is the generalization guess. The raw sensor data are
der to localize the robot and in the same time focusing on fed to a three stages pipeline, consisting on preprocessing,
generalization of the classification technique. Such archi- object detection and classification steps. The output is the
tecture provides the ability of understanding the topologi- identifier of every detected object. This technique has a
cal relations existing in an indoor environment. In particu- low computational cost and is suitable for on-line applica-
lar, this method associates semantic terms (i.e., "Corridor", tions, such as robotic manipulation or automated logistic
"Kitchen") to each specific place, giving an "intuitive" idea systems[10]. An example of multiple object recognition in
of the position of the robot. To accomplish this goal, four presence of occlusions is reported in Figure 5.
different methodologies have been integrated. All these
methodologies are based on the same multi-level machine
learning architecture (Figure 3). The developed architec-
ture is made of a first level of "weak" classifiers based on
”visual” features extracted from stereo images couple, and
a second level of classifiers performing fusion and regres-
sion of the first level outputs. Four different configurations
of the regressor stage have been compared in order to ob-
tain the best results to problem solving. These configura-
tions employ different methods that we call: "Committee
of Experts", "Stacked Regression with Support Vector Ma-
chines (SVM) stage", "Stacked Regression with Artificial
Neural Network stage", "Weighted Linear Combination".
The experimental tests show that "Stacked Regression with
Support Vector Machines stage" result to be the best ap- Fig. 5. Detection and classification of multiple objects with
proach providing a high generalization capability [16]. occlusions.

D. Atonomous navigation
Fusing techniques have also been employed in a flying
robot to optimally merge all information gained from the
on-board sensors in order to perform simultaneous local-
ization and mapping (SLAM). The flying robot is a Micro
Aerial Vehicle capable of navigating in outdoor, indoor and
confined environments. The aim of the work was to replace
the human operator during the inspection of failures or pro-
cedures correctness inside dangerous areas where human
presence should be avoided for long or short time peri-
ods. To be able to accomplish this specific task the system
presents a semi-autonomous behavior. The human opera-
tor instruct the system about the trajectory to be performed
Fig. 4. Visual place classification fusing different classifi- and the system is able to process all the navigation data on-
cation methods. board so that even in presence of comunication losses the
system remains stable and safe.

4
 Given the limited amount of computational power of [4] Zhu, Y.:“Moving Object Detection and Segmenta-
the embedded computer, an optimized SLAM algorithm tion Based on Background Subtraction Image Over-
that allows correct navigation of the robot has been imple- Segmentation". J. Software, 6 (2011)
mented. The designed system has been oriented towards [5] Herbert Bay, Andreas Ess, Tinne Tuytelaars, Luc Van
the specific context of visual inspection of industrial area Gool, "SURF: Speeded Up Robust Features", Com-
and it has been tested firstly in a laboratory test-bed [15] puter Vision and Image Understanding (CVIU), Vol.
and later in a real power plant, see Fig. 6. 110, No. 3, pp. 346-359, 2008
[6] Lowe, David G., ”Object recognition from local scale-
invariant features”, Proceedings of the International
Conference on Computer Vision, vol. 2, 1999, pp.
1150-1157
[7] Young H. Choi, Tae K. Lee, Se Y. Oh, ”A line feature
based SLAM with low grade range sensors using ge-
ometric constraints and active exploration for mobile
robot”. Auton. Robots, Vol. 24 (January 2008), pp. 13-
27
[8] Grisetti, G.; Stachniss, C.; Burgard, W., ”Im-
proved Techniques for Grid Mapping With Rao-
Blackwellized Particle Filters,” Robotics, IEEE Trans-
actions on , vol.23, no.1, pp.34,46, Feb. 2007
[9] Monteiro, S., Bicho, E.: ”A dynamical systems ap-
proach to behavior-based formation control". IEEE
ICRA, pp. 2606-2611 (2002)
[10] Peppoloni, L., Satler, M., Luchetti, E., Avizzano,
Fig. 6. Flying robot inspectioning an industrial boiler.
C.A., Tripicchio, P.:" Stacked Generalization for scene
analysis and object recognition". IEEE International
Conference on Intelligent Engineering Systems (2014)
V. CONCLUSIONS
[11] Saffiotti, A., Konolige, K., Ruspini, E.H.:” A multi-
Autonomous navigation requires complex robotic abili- valued logic approach to integrating planning and con-
ties. The robot should be able to react to changes in the en- trol”. Artificial Intelligence, pp. 481-526 (1995)
vironments still reaching the planned goal.This paper pre-
[12] Ajzen, I.: "From Intentions to Actions: A Theory of
sented a possible hybrid approach that makes use of both
Planned Behavior. Action-control: From cognition to
a reactive and a deliberative part for reasoning. To be able
behavior", pp. 11-39. Springer, Heidelberg (1985)
to react to sudden changes in the environment, specific vi-
sion and machine learning algorithms has been presented [13] Barraquand, J., Langlois, B., Latombe, J.C.: "Numer-
for the detection and recognition, in first instance, of ob- ical potential field techniques for robot path planning".
jects and people moving in the environment, and in second IEEE Transactions on Systems, Man, and Cybernetics,
instance, to classify and recognize the environment itself pp. 224-241 (1992)
(and take advantage of the robot knowledge). The pre- [14] Hagras, HA: "A hierarchical type-2 fuzzy logic con-
sented sensor fusion techniques are also applied to obtain trol architecture for autonomous mobile robots". IEEE
simultaneous localization and mapping in embedded com- Transactions on Fuzzy Systems, pp. 524-539 (2004)
puting hardware and allow small robots to safely operate [15] Satler, M.; Unetti, M.; Giordani, N.; Avizzano, C.A.;
in harsh environments. Tripicchio, P., ”Towards an autonomous flying robot
for inspections in open and constrained spaces”, Multi-
REFERENCES Conference on Systems, Signals & Devices (SSD),
[1] S.Mallat, “A wavelet tour of Signal Processing. The 2014 11th International , vol., no., pp.1,6, 11-14 Feb.
Sparse Way”, 3rd edition, Academic Press, New York, 2014
USA, 2009. [16] Walter Lucetti, Emanuel Luchetti, ” Combination of
[2] Stauffer, C., Grimson,“W.E.L.: Adaptive Background Classifiers fot Indoor Room Recognition”, In proceed-
Mixture Models for Real-Time Tracking”. IEEE ing of: CLEF 2010
CVPR (1999) [17] Scaramuzza, D.; Siegwart, R., "Appearance-Guided
[3] Wren, C., Azarbayejani, A., et al.:“Pfinder: real-time Monocular Omnidirectional Visual Odometry for Out-
tracking of the human body". IEEE Transaction on door Ground Vehicles," Robotics, IEEE Transactions
Pattern Analysis and Machine Intelligence (1997) on , vol.24, no.5, pp.1015,1026, Oct. 2008

5
[18] A. Censi, "An ICP variant using a point-to-line met- strot, M. de Melo, A. Schebesch, M. Simon, M. Spren-
ric" Proceedings of the IEEE International Conference gel, O. Tenchio.”Using Hierarchical Dynamical Sys-
on Robotics and Automation (ICRA), 2008 tems to Control Reactive Behavior”. In Lecture Notes
[19] S. Behnke, B. Frotschl, R. Rojas, P. Ackers, W. Lind- in Computer Science Volume 1856, 2000, pp 186-195