Published in Proc. of Int. Conf. on Intelligent Robots and Systems (IROS) 2004
Core Technologies for Service Robotics
Niklas Karlsson, Mario E. Munich, Luis Goncalves,
Jim Ostrowski, Enrico Di Bernardo, and Paolo Pirjanian
Evolution Robotics, Inc.
Pasadena, California, USA
Email: {niklas,mario,luis,jim,enrico,paolo}@evolution.com
Abstract— Service robotics products are becoming a reality.
This paper describes three core technologies that enable the
next generation of service robots. They are low-cost, make
use of low-cost hardware, and prepare for a short time-tomarket for product development. The first technology is an
object recognition system, which can be used by the robot
to interact with the environment. The second technology is
a vision-based navigation system (vSLAMTM ), which simultaneously can build a map and localize the robot in the map.
Finally, the third technology is a flexible and rich software
platform (ERSPTM ) that assists developers in rapid design
and prototyping of robotics applications.
I. I NTRODUCTION
Products such as Sony’s entertainment robot AiboTM ,
Electrolux’s robotic vacuum cleaner TrilobiteTM , iRobot’s
RoombaTM , and prototypes from numerous robotic research
laboratories across the world forecast a rapid development
of service robots over the next several years. In fact, the
momentum created by these early products combined with
recent advances in computing and sensor technology could
lead to the creation of a major industry for service robots.
The applications of service robotics range from existing
products, where robotic technologies are incorporated to
enhance and improve the product functionality, to new
products, where robotic technologies enable completely
new functionalities. One example of the first case is vacuum cleaners, where the enhancement is reflected in a never
before seen degree of autonomy. An example of the second
case is companionship robots. Other examples can be found
in interactive entertainment and edutainment.
A necessary element to make the industry for service
robots grow is to develop core robotic technologies that
deliver flexibility and autonomy with low-cost hardware
components. At Evolution Robotics (ER) we develop such
technologies, and they are available as individual components and at low cost for the consumer market. This
allows the robotics community to focus on the value
added applications development rather than solving the
core robotic problems, which on the other hand enables
the industry to proliferate.
A major challenge in providing core technologies for
service robots is that the technologies must be low-cost,
yet provide rich and flexible features. In particular, the
technologies should enable a short time-to-market. At the
same time, they must provide the robots “intelligent”
features, meaning that they should give the robots the
ability to interact with the environment and to operate
somewhat autonomously.
For example, a robotic vacuum cleaner which retails at
199USD can only cost about 60USD in bill of materials.
The 60USD must contain all the hardware components,
assembly costs, and typically packaging, manuals, and
shipping. This constraint greatly limits the choices of the
sensors, actuators, and amount of computation that will be
available for the autonomous capabilities of the product.
Due to this cost constraint, currently most robotic products
which retail at a price acceptable to the consumers provide very limited autonomous functionality. For example,
robotic vacuum cleaning products such as Roomba and
the like perform random navigation for floor coverage and
tactile sensing for collision detection and avoidance.
However, random navigation for floor cleaning is not
efficient since it cannot guarantee complete coverage from
wall to wall and from room to room at a reasonable
power consumption. Random coverage cannot guarantee
that the robot will be able to return to a home location,
e.g. a docking station for self-charging. Localization can
enable a robot to perform complete coverage at a closeto-optimal power budget as well as allow it to navigate
from room to room to perform its cleaning task and
finally return back to its docking station for recharging.
Combining cost constraints with requirements on product
durability, reliability, power consumption, and expectations
of autonomous capabilities creates a major challenge in
creating successful products.
This paper is organized as follows. Section II highlights
some of the requirements that are placed on the technologies that are used for service robotics. Each requirement
is thereafter addressed by proposing appropriate core technologies. Section III describes ERs object recognition system. This object recognition system is one component in a
novel navigation system, which is described in Section IV.
A competitive set of core technologies for service robotics
also requires a rich and flexible software architecture.
Section V describes such an architecture. A few examples
of product concepts, which are enabled by the proposed
core technologies, are described in Section VI. Finally,
Section VII establish some concluding remarks.
II. R EQUIREMENTS
The service robotics market imposes a complex set of
requirements on potential products: it is a consumer market, so robots should be inexpensive, the market is highly
competitive, so products should have a short time-tomarket. At the same time, robots need to provide services
in an “intelligent” way, so they should be autonomous,
adaptive, and able to interact with the environment. This
paper describes a set of solutions for these requirements
that have been developed at Evolution Robotics, Inc.
The price sensitivity of the market creates the need
for the development of low-cost technologies. This is
achieved first of all by basing the technologies primarily
on hardware components that are relatively inexpensive, for
example, cameras and IR sensors. Indeed, a web camera
is two orders of magnitude less expensive than, e.g., a
SICKTM laser range sensor.
In order for a robot to qualify as a service robot, it
must also possess high level of autonomy. It must be
able to operate within its environment with a minimum
amount of user interaction. This is achieved by providing
a robust navigation system containing mapping, as well as,
localization capabilities.
Furthermore, to serve its purpose, a service robot must
have the ability to interact with and make inference from its
environment. A strong object recognition system satisfies
both these requirements. By extracting relevant information
from vision data, the service robot can make conscious
and intelligent decisions based on interaction with the
environment and people in the environment.
To be competitive, the core technologies for service
robotics must also offer a short time-to-market, in that
the time from a conceptual idea to a final product must
be minimized. This translates into a requirement on the
software platform used for development. The software
platform should provide an architecture that is rich and
flexible. The architecture should be designed in a way that
allows the developer to reconfigure the hardware without
rewriting more than a very small amount of code.
Some of the above requirements also imply that the
core technologies must enable adaptivity of the system.
For example, the navigation system must be able to deal
with a changing environment.
III. ER V ISION
The ER Vision object recognition system offered by
Evolution Robotics is versatile and works robustly with
low-cost cameras. It addresses the greatest challenge for
object recognition systems, which is the development of
algorithms that reliably and efficiently perform recognition
in realistic settings with inexpensive hardware and limited
computing power.
The object recognition system can be used in applications such as navigation, manipulation, human-robot
interaction, and security. It can be used in mobile robots
to support navigation, localization, mapping, and visual
servoing. It can also be used in machine vision applications
for object identification and hand-eye coordination. Other
applications include entertainment and education. The object recognition system provides a critical building block
in applications such as reading children’s books aloud, or
automatically identifying and retrieving information about
a painting or sculpture in a museum.
The object recognition system specializes on recognizing
planar, textured objects. However, three-dimensional objects composed of planar, textured structures, or composed
of slightly curved components, are also reliably recognized
by the system. Three-dimensional deformable objects such
as a human face cannot be handled in a robust manner.
The object recognition system can be trained to recognize objects using a single, low-cost camera. Its main
strength lies in its ability to provide reliable recognition in
realistic environments where lighting can change dramatically. The following are the two steps involved in using
the Evolution Robotics’ object recognition algorithm:
Training: Training is accomplished by capturing one or
more images of an object from various viewpoints. For
planar objects, only the front and rear views are necessary.
For 3-D objects, several views, covering all facets of the
object, are necessary.
Recognition: The algorithm extracts up to 1,000 local,
unique features for each object. A small subset of those
features and their interrelation identifies the object. The
object’s name and its full pose up to scale with respect to
the camera are the results of recognition.
The object recognition algorithm is based on extracting
salient features from an image of an object [6]. Each
feature is uniquely described by the texture of a small
window of pixels around it. The model of an object consists
of the coordinates of all these features along with each
feature’s texture description. When the algorithm attempts
to recognize objects in a new image, it first finds features in
the new image. It then tries to associate features in the new
image with all the features in the database of models. This
matching is based on the similarity of the feature texture.
If many features in the new image have good matches
to the same database model, that potential object match is
refined. The refinement process involves the computation
of an affine transform between the new image and the
database model, so that the relative position of the features
is preserved through the transformation. The algorithm
outputs all object matches for which the optimized affine
transform results in a small root-mean-square pixel error
between the features found in the new image, and the
corresponding affine-transformed features of the original
model.
Figure 1 shows the main characteristics of the recognition algorithm. The first row displays the two objects
to be recognized and the other rows presents recognition
results under different conditions. The main characteristics
of the object recognition algorithm are summarized in the
following list:
a) Invariant to rotation and affine transformations:
The object recognition system recognizes objects even
if they are rotated upside down (rotation invariance) or
placed at an angle with respect to the optical axis (affine
invariance). See second and third row of Figure 1.
b) Invariant to changes in scale: Objects can be recognized at different distances from the camera, depending
on the size of the objects and the camera resolution. The
recognition works reliably from distances of several meters.
Fig. 1. Examples of object recognition under a variety of conditions. The
first row presents the objects to be recognized and other rows displays
recognition results.
See second and third row of Figure 1.
c) Invariant to changes in lighting: The object recognition system can handle changes in illumination ranging
from natural to artificial indoor lighting. The system is insensitive to artifacts caused by reflections or back-lighting.
See fourth row of Figure 1.
d) Invariant to occlusions: The object recognition
system can reliably recognize objects that are partially
blocked by other objects, and objects that are partially in
the camera’s view. The amount of occlusions allowed is
typically between 50% and 90% depending on the object
and the camera quality. See fifth row of Figure 1.
e) Reliable recognition: The object recognition system has 80-95% success rate in uncontrolled settings. A
95-100% recognition rate is achieved in controlled settings.
The recognition speed is a logarithmic function of the
number of objects in the database; i.e., the recognition
speed is proportional to log (N), where N is the number
of objects in the database. The object library can store
hundreds or even thousands of objects without a significant
increase in computational requirements. The recognition
frame rate is proportional to CPU power and image resolution. For example, the recognition algorithm runs at 5
frames per second (fps) at an image resolution of 320x240
on an 850MHZ Pentium III processor and 3 fps at 80x66
on a 100MHz MIPS-based 32-bit processor.
Reducing the image resolution decreases the image
quality and, ultimately, the recognition rate. However, the
object recognition system allows for graceful degradation
with decreasing image quality. Each object model requires
about 40KB of memory.
IV.
V SLAM
Evolution Robotics’ Visual Simultaneous Localization
and Mapping (vSLAMTM ) system is robust and requires
no initial map. It is the core navigation functionality
of Evolution Robotics Software Platform (ERSPTM ), and
enables simultaneous localization and mapping using a
low-cost camera and wheel encoders as the only sensors.
Because a web camera is used instead of a SICK laser
range sensor or other expensive range measuring device,
the vSLAM technology offers a dramatic cost reduction.
With vSLAM, a service robot can operate in a previously
unknown environment to build a map and to localize itself
in this map.
Relying on visual measurements rather than range measurements when building maps allows vSLAM to deal with
cluttered spaces easily. Range-based SLAM techniques,
involving laser scanners and similar devices, often fail in
cluttered environments. The vSLAM algorithm handles dynamic changes in the environment well, including lighting
changes, moving objects and/or people, simply because it
possesses such a wealth of information about the created
landmarks.
Figure 2 illustrates the inputs and outputs of vSLAM.
The inputs to the algorithm are dead reckoning information (for example, wheel encoder odometry) and images
acquired by a camera. The main components of vSLAM
are Visual Localization, SLAM, and Landmark Database.
Fig. 2.
Block diagram of vSLAM.
In the Visual Localization module the object recognition
system described in Section III is used to define new
and recognize old natural landmarks. The landmarks are
used for localization of the robot. In particular, when
a previously defined landmark is recognized, a relative
measurement of the robot pose is computed with respect to
the pose of the robot when the landmark was first defined.
Some related work is described in [6], [8], [11]
In the SLAM module, the relative measurements are
fused with odometry data to update the robot pose and
the poses of all the landmarks. For related work, see for
example [3], [5], [7], [9], [10]. The SLAM algorithm
is developed to track multiple hypothesis, but also to
maximize the robustness to dramatic disturbances such as
kidnapping, which is the scenario where the robot is lifted
and moved to a new location without notification.
In the Landmark Database module information is stored
and updated on the landmarks in the vSLAM map. For
example, the unique identifiers are recorded, as well as
estimates of the landmark poses and uncertainty measures
of the estimates.
When entering a new area or environment, the robot can
either load an existing map of that area, or immediately
start creating a new map with a new set of landmarks.
During its operation, the robot will attempt to create
new visual landmarks. A visual landmark is a collection of
unique features extracted from an image. Each landmark is
associated with an estimate of where the robot was located
(the x, y, and heading coordinates) when the landmark was
created. As the robot traverses the environment, it continues
creating new landmarks and correcting its odometry as
necessary, using these new landmarks and a corresponding
map.
vSLAM is constantly improving the robot’s knowledge
about its environment. For example, new landmarks are
automatically created if the robot is mobile for too long
without recognizing any known landmarks in its database.
This enables the robot to constantly refine its understanding
of its environment and increase its confidence in where it
is located.
In order to understand the limitations of vSLAM, it is
necessary to understand the different steps that take place
in the vSLAM algorithm. These are:
1) Traverse the environment, collect odometry, and acquire an image.
2) Process the image to extract visual features.
3) If the number of visual features are sufficiently large
and distinct, then
a) Determine if the features correspond to a previously stored landmark.
b) If the features do correspond to a previously
stored landmark, then:
i) Compute a visual measurement of where the
robot is located now, relative to where it was
located when the landmark was created.
ii) Based on the visual measurement, the current map, and the collected odometry; improve the localization of the robot.
iii) Based on visual measurement, localization,
and collected odometry; refine the estimated
location of the landmarks..
c) If the features do not correspond to a previously
stored landmark, then try to create a new landmark. In order to create a new landmark, some
of the same visual features must be detected in
three consecutive image frames.
i) Assign a unique landmark number to the
features, and add this landmark to the previously created landmarks.
ii) Initialize the landmark pose based on odometry information.
4) Return to step 1.
Note that, if the environment does not contain a sufficient
amount of visual features, then steps 3a through 3c will
never be executed. Hence, no landmark will be created
and no visual measurements will be computed. vSLAM
will still produce pose estimates, but the estimates will
be exactly what is obtained from wheel odometry. This
scenario is rarely experienced, but is most likely to happen
in environments that are free from furniture, decorations,
and texture.
The vSLAM algorithm is typically good at detecting and
correcting for kidnapping. Note that dramatic slippage and
collisions may be considered as kidnapping events and
must be handled by any reliable navigation technology.
vSLAM recovers quickly from kidnapping once the map
has become dense with landmarks. Early in the mapping
procedure, is it very important to avoid kidnapping and
disturbances in general.
Figure 3 shows the result of vSLAM after the robot has
traveled around in a typical two-bedroom apartment. The
robot was driven around along a reference path (this path is
unknown to the SLAM algorithm). The vSLAM algorithm
builds a map consisting of landmarks marked with blue
circles in the figure. The corrected robot path, which uses
a combination of visual features and odometry, provides a
robust and accurate position determination for the robot as
seen by the red path in the figure.
Fig. 3.
Example result of SLAM using vSLAM. Red path (darker
gray): vSLAM estimate of robot trajectory. Green path (lighter gray):
odometry estimate of robot trajectory. Blue circles: vSLAM landmarks
created during operations.
The green path (odometry only) is obviously incorrect,
since, according to this path, the robot is traversing through
walls and furniture. The red path (the vSLAM corrected
path), on the other hand, is consistently following the
reference path.
Based on experiments in various typical home environments on different floor surfaces (carpet and hardwood
floor), and using different image resolution (Low: 320 x
280, High: 640 x 480) the following robot localization
accuracy was achieved:
Floor
surface
Carpet
Carpet
HW
HW
Image
resolution
Low
High
Low
High
Median
error
(cm)
13.6
12.9
12.2
11.5
Std dev
error
(cm)
9.4
10
7.5
8.3
Median
error
(deg)
4.79
5.97
5.28
3.47
Std dev
error
(deg)
7.99
7.03
6.62
5.41
V. E VOLUTION ROBOTICS S OFTWARE A RCHITECTURE
The Evolution Robotics Software Platform provides basic components and tools for rapid development and proto-
typing of robotics applications. The software architecture,
which is the infrastructure and one of the main components
of ERSP, addresses the issue of short product development
cycles (time-to-market). The design of ERSP is modular
and allows applications to use only the required portions
of the architecture.
Figure 4 shows a diagram of the software structure; and
the relationship among the software, operating system, and
applications.
Fig. 4.
ERSP structure and relation to application development.
There are five main blocks in the diagram – three of
them, Applications, OS & Drivers, and 3rd Party Software
correspond to external components to the ERSP. The other
two blocks correspond to subsets of ERSP: the core technology libraries (left-hand-side block) and the implementation libraries (center block). The core technology libraries
consist of the following components:
• Architecture: The Architecture component provides a
set of Application Programmer’s Interfaces (APIs) for
integration of all the software components with each
other, and with the robot hardware. The infrastructure
consists of APIs to interact with the hardware, to build
task-achieving modules that can make decisions and
control the robot, to orchestrate the coordination and
execution of these modules, and to control access to
system resources. The three primary components of
the Architecture are the Hardware Abstraction Layer
(HAL), the Behavior Execution Layer (BEL), and the
Task Execution Layer (TEL).
• Vision: The Vision APIs correspond to the object
recognition algorithm described in Section III.
• Navigation: The Navigation APIs provide mechanisms for controlling movement of the robot. These
APIs provide access to modules for mapping, localization (described in Section IV), exploration, path
planning, obstacle avoidance, occupancy grid mapping, target following, and teleoperation control.
• Interaction: The Interaction APIs support building
user interfaces for applications with graphical user
interfaces, voice recognition, and speech synthesis.
The architecture is composed of three layers and corresponds to a mixed architecture in which the two first
layers follow a behavior-based philosophy [1], [2] and the
third layer incorporates a deliberative stage for planning
and sequencing [4]. The implementation libraries use the
interfaces defined by the architecture to implement the
three layers of the architecture for particular hardware and
purposes.
The first layer, HAL, provides interfaces to the hardware
devices and low-level operating system (OS) dependencies.
This assures portability of Architecture and application
programs to other robots and computing environments.
At the lowest level, HAL interfaces with device drivers,
which communicate with the hardware devices through
a communication bus. The description of the resources,
devices, busses, their specifications, and the corresponding
drivers are managed through configuration files based on
Extensible Markup Language (XML).
The second layer, BEL, provides infrastructure for development of modular robotic competencies, known as
behaviors, for achieving tasks with a tight feedback loop
such as following a trajectory, tracking a person, avoiding
an object, etc. Behaviors are the basic, reusable building
blocks on which robotic applications are built. BEL also
provides techniques for coordination of the activities of
behaviors, for conflict resolution, and for resource scheduling. Behaviors are organized into a behavior network to
achieve more complex goals, such as avoiding obstacles
while following a target. Behavior networks are executed
synchronously at a rate set by the user.
BEL defines a common and uniform interface for all
behaviors and the basic protocols for interaction among
the behaviors, as well as the order of execution for each
behavior in a behavior network. The user has the flexibility
of selecting the communication ports and the data transference protocols between behaviors, as well as deciding the
internal computation in the behavior. A typical behavior
network would contain a number of sensory behaviors and
a number of actuation behaviors. The sensory behaviors’
outputs are fed as inputs to the actuation behaviors, which
then control the robot according to the sensory data. By
analyzing the flow of behavior inputs and outputs, the
Behavior Manager automatically orders the execution of
behaviors in the network so that sensory behaviors execute
before actuation behaviors. The coordination of behaviors
is transparent to the user.
Finally, the third layer, TEL, provides infrastructure
for developing goal-oriented tasks along with mechanisms
for the coordination of task executions. Tasks can run in
sequence or in parallel. Execution of tasks can also be
triggered by user-defined events. Events are conditions or
predicates defined on values of variables within BEL or
TEL. Complex events can be defined by logical expressions
of basic events.
While behaviors are highly reactive, and are appropriate
for creating robust control loops, tasks are a way to express higher-level execution knowledge and coordinate the
actions of behaviors. Tasks can run asynchronously using
event triggers or synchronously with other tasks. Timecritical modules such as obstacle avoidance are typically
implemented in the BEL, while tasks implement behaviors
that are not required to run at a fixed execution rate. Tight
feedback loops for controlling the actions of the robot
according to perceptual stimuli (presence of obstacles,
detection of a person, etc.) are typically implemented in
the BEL. Behaviors tend to be synchronous and highly
data driven. TEL is more appropriate to deal with complex control flow which depends on context and certain
conditions that can arise asynchronously. Tasks tend to be
asynchronous and highly event-driven.
VI. P RODUCT C ONCEPTS
The product conceptualization is the step that takes a
given technology and designs an application for it. This
section describes a few product concepts enabled by the
technologies described in this paper. The described concepts are:
• The use of object recognition for command and control of robots
• The use of vSLAM for efficient vacuuming by robotic
vacuum cleaners
An important capability of a service robot is its ability
to interact with people in a natural, friendly, and robust
manner. Traditionally, speech-based interaction has been
the natural choice in a variety of designs. However, the
shortcomings of speech recognition in terms of recognition
robustness and accuracy with respect to user location
renders this technology to be unreliable for robotic applications. Therefore, we propose the use of vision-based
object recognition to communicate a command to the robot.
Commands are issued by showing predefined cards to the
robot. The described robustness of the algorithm makes
it an excellent technology for reliable recognition of the
cards.
The idea of commanding a robot with cards was initially prototyped with Evolution Robotics ER2TM robots.
However, the idea is now productized in the latest version
of Sony’s AIBOTM robotic dog, the ERS-7. This robot is
equipped with a set of LEDs to provide visual feedback to
the user. Thus, each time the robot recognizes a cards, a
visual display is shown as well as a sound is played.
A second important capability of a service robot is
its ability to navigate in an unknown environment. For
example, an autonomous vacuum cleaner must be able
to estimate its own position in order to carry out an
efficient sweeping of the floor. Due to the complexity
of the problem, most autonomous vacuum cleaners today
avoid dealing with the localization problem. Instead, they
typically operate under some random walk strategy. This
is necessarily an inefficient way of doing coverage since
many spots will be covered multiple times before the last
spot is covered. Also, there is no mechanism to guarantee
that all of the environment has been covered at all.
The usage of vSLAM to do mapping and localization
in support of efficient sweeping has been prototyped on
various robot platforms at Evolution Robotics, and has
demonstrated to dramatically improve coverage and efficiency. Experiments have shown that vSLAM offers an
increased degree of autonomy and “intelligence” to vacuum
cleaners and service robots in general.
VII. C ONCLUSIONS
We have described three core technologies developed
by Evolution Robotics that we constitute key components
in the next generation of service robots. They succeed in
satisfying the typical requirements on service robotics: lowcost. short time-to-market, reliability to perform in real
world settings.
The basic requirement is that these technologies have to
be low cost to be employed in service robots. At the same
time, these technologies must enable the products to have a
short time-to-market and provide the products “intelligent”
features.
The proposed technologies have been developed at
Evolution Robotics. The cost requirement is achieved by
leveraging on low-cost hardware components like cameras
and IR sensors.
• ER Vision: The object recognition system gives the
robot an ability to interact with and make inference
from its environment.
• vSLAM: The vision-based navigation system makes it
possible for a robot to simultaneously build a map and
localize itself in the map. In particular, the vSLAM
technology does not require an initial map, but builds
a map from scratch.
• ERSP: The proposed software architecture addresses
the issue of of short product development cycles.
The Evolution Robotics Software Platform provides
basic components and tools for rapid development and
prototyping of robotics applications. The design of
ERSP is modular and allows applications to use only
the required portions of the architecture. It is designed
in a way that allows the developer to reconfigure the
hardware without rewriting more than a very small
amount of code.
R EFERENCES
[1] Arkin, R. C., Behavior-Based Robotics, MIT Press, 1998
[2] Brooks, R. A., A Robust Layered Control System for a Mobile Robot,
IEEE Journal of Robotics and Automation, March 1986.
[3] Fox, D., Burgard, W., and Thrun, S. Markov localization for mobile
robots in dynamic environments. Journal of Artificial Intelligence
Research 11, pp. 391-427, 1999.
[4] Gat, E., On Three-Layer Architectures, in D. Kortenkamp et al. eds.,
AI and Mobile Robots. AAAI Press, 1998.
[5] Gutmann, J.-S. and Konolige, K., Incremental Mapping of Large
Cyclic Environments, Proceedings of the IEEE International Conference on Robotics and Automation (ICRA), San Francisco, CA,
2000.
[6] Lowe, D. Local feature view clustering for 3D object recognition.
Proc. of the 2001 IEEE/RSJ, Int. Conf. on Computer Vision and
Pattern Recognition, Kauai, Hawaii, USA, December 2001.
[7] Roumeliotis, S.I. and Bekey, G.A. Bayesian estimation and Kalman
filtering: A unified framework for mobile robot localization. Proceedings of the IEEE International Conference on Robotics and
Automation (ICRA), pages 2985–2992, San Francisco, CA, 2000.
[8] Se, S., Lowe, D., and Little, J. Local and global localization for
mobile robots using visual landmarks. Proc. of the 2001 IEEE/RSJ,
Int. Conf. on Intelligent Robots and Systems, Maui, Hawaii, USA,
October 2001.
[9] Thrun, S., Fox., D., and Burgard, W. A probabilitic approach to
concurrent mapping and localization for mobile robots. Machine
Learning, 31:29-53, 1998.
[10] Thrun, S. Mapping: A Survey. Technical Report, CMU-CS-02-111,
Carnegie Mellon University, Pittsburgh, PA, USA, February, 2000.
[11] Wolf, J., Burgard, W., and Burkhardt, H. Robust vision-based
localization for mobile robots using an image retrievel system based
on invariant features Proceedings of the 2002 IEEE Int. Conf. on
Robotics and Automation, Washington, DC, USA, May, 2002.