The Joint Programme of Work

Research Objectives

The goal of the VIRGO network is to coordinate European research and postgraduate training activities that address the development of intelligent robotic systems able to navigate in (partially) unknown and possibly dynamic environments. This goal will be achieved through a framework which enhances RTD activities in European laboratories, already established in the aforementioned scientific area. Specifically, in pursuing this goal, alternative environment representations based on visual information will be studied and methods to process these representations and use them to control the motion of the mechanical parts of a robot will be developed. Although VIRGO will focus on the use of visual information, other sensors will also be considered in order to simplify and facilitate certain tasks. Environment learning issues will also be considered within VIRGO, aiming at autonomous acquisition of the discriminating environment features.

Methodological Approach and Work Plan

VIRGO will focus on the use of visual information in robot navigation, that is information acquired from a sequence of images. This information will, however, be combined with information provided by other sensors (e.g. sonar, IR, tactile) to simplify certain tasks. In acquiring and processing the visual information, our approach is based on the Purposive and Qualitative vision paradigm, which offers many advantages for the particular task.

In order to approach the coordination of the VIRGO Network in a tractable way, it is broken down into tasks (T1-T8) with specific (sub)goals. The systematic and in-depth execution of these tasks will provide solutions and insights to the problems involved. Emphasis will be given to the integration of methodologies, since the goal of autonomous navigation requires more than solutions to specific subproblems.

Tasks

• T1: Analysis of current methods for robot navigation.
• T2: Visual competences.
• T3: Landmark recognition.
• T4: Visual memory organization.
• T5: Sensor Fusion.
• T6: Strategic/Tactical Planner.
• T7: Learning methodologies.
• T8: Transfer and Exploitation.

Participation in the Tasks of VIRGO

	FORTH	AUC	DIST	TUG	KTH	BONN	INRIA	GMD	DIKU	UNIZH
T1	All Partners
T2
T3
T4
T5
T6
T7
T8	All Partners

Brief technical description of tasks

• T1: Analysis of current methods for robot navigation.
  Participation: All Leader: KTH
  A systematic analysis of current methods for robot navigation will be performed to identify problem areas and inadequacies. Although this project focuses on vision-based navigation, this analysis will also consider non-visual methods in order to provide a clear understanding of current techniques and approaches.
• T2: Visual competences.
  Participation: FORTH, AUC, DIKU, INRIA, KTH, TUG Leader: FORTH
  A number of visual cues have a definite impact on robot motion. Therefore, the system should possess a number of competences, that is abilities to capture aspects of the whole visual information. In this task, the following competences will be studied: egomotion estimation, independent motion detection, depth estimation, time to contact with points/curves/areas and even with sheer gray-levels, tracking and visual homing. Our general approach in this task will be ``qualitative'', since the robot has only partial knowledge of the parameters of its vision system. Moreover, we will consider how and what invariant information can be recovered from the images and efficiently used in mobile robotics applications. The above approach will be based on the assumption of an active observer, possessing full control of the image acquisition process (active vision paradigm).
• T3: Landmark recognition.
  Participation: INRIA, FORTH, GMD, AUC, KTH, BONN Leader: INRIA
  Robot navigation will be primarily based on landmarks, that is characteristic aspects of the environment that can be used to describe it. Each landmark is described by a set of features; following the recent trends in computer vision and in order to achieve robust recognition, qualitative descriptors will be studied rather than quantitative. Candidate descriptors are ordinal depth, qualitative surface descriptors, shape, texture, color, etc. Landmarks will initially be identified manually; however, in coordination with T7, automated landmark identification (learning) will also be studied. Moreover, since landmark recognition cannot be studied independent of memory considerations, T3 will run in close coordination with T4.
• T4: Visual memory organization.
  Participation: AUC, FORTH, DIST, UNIZH, GMD, INRIA Leader: AUC
  The system's knowledge of the environment is specified by the contents of a visual memory (VM) which contains a large number of patterns corresponding to landmark descriptions known to the system. Patterns can also encode other information, e.g. the direction of the robot motion when this pattern (landmark) is encountered, hints for the other landmarks, the position of the corresponding segment of the environment with respect to a specific reference system, etc. VM contains descriptions of the local environment in reference to coordinate systems defined with respect to the landmarks. Therefore, it substitutes the 3-D geometric representation of space (traditionally used) with a set of selected space views with known topological relationships.
• T5: Sensor Fusion.
  Participation: TUG, DIKU, BONN, KTH, UNIZH Leader: TUG
  Vision alone provides useful information for space perception and navigation. However, certain tasks can be simplified and approached more effectively by employing other sensors. The case of an object approaching the robot sideways can be easily accounted for by using a ring of sonars around the robot and contact with objects is trivially detected with tactile sensors. The information provided by other sensors will be integrated with the visual information, complementing it or even replacing it (in case of emergency, a non-visual sensor might take over the robot control temporarily). Fusion of peripheral and central vision will also be considered under this task.
• T6: Strategic/Tactical Planner.
  Participation: GMD, AUC, DIST, FORTH Leader: GMD
  Based on representations of the environment, motion planning is performed at two levels. At the higher level, a strategic planner is responsible for specifying a sequence of intermediate goals, the accomplishment of which results in the final goal. The intermediate goals correspond to landmarks that have been chosen by the system through a learning process. The strategic planner uses approximate descriptions of the environment and, therefore, its plans are not always achievable. A tactical planner at a lower level, navigates the robot between two intermediate positions that are specified by the strategic planner. The tactical planner considers the current status of the environment (e.g. obstacles) and uses appropriate visual competences (T2) to accurately perform its task.
• T7: Learning methodologies.
  Participation: DIST, GMD, FORTH, BONN, UNIZH, TUG Leader: DIST
  Above, a teacher has been assumed to identify the discriminating characteristics of the environment (landmarks). This task will aim at investigating issues related to spatial learning (``iconic'' navigation), that is the possibility of a robot to be ``shown'' a path (e.g. using a joystick or following -tracking- a person) and to be able to repeat the same path (or paths with similar characteristics) on the basis of information stored during the ``teaching'' phase and on the current images. In particular, we intend to investigate how to (a) ``condense'' the iconic information (take images at certain times), (b) model and define the sensor - motor interaction, (c) handle ``memorized'' images, (d) cope with changes in the environment after the teaching phase, and (e) combine the iconic approach with explicitly placed visual landmarks.
• T8: Transfer and Exploitation.
  Participation: All Leader: FORTH
  Evaluation of results from the point of view of possible industrial exploitation. This task will also involve industrial sponsors of VIRGO, who will also be encouraged to interact with VIRGO participants during all phases of the project programme.

Schedule and Milestones

The duration of VIRGO will be 36 months. The time-schedule of the project is shown in the following chart.

Milestones:

• End of Month 12. Joint report on state-of-the-art approaches in vision-based robot navigation.
• End of Month 18 (Mid-Term). Joint progress report on T2 and planned VIRGO enhancements to specific local demonstrators. Specifications of integration requirements for such enhancements to be effective.
• End of Month 24. Joint report on the work conducted in visual competences, landmark recognition and visual memory organization.
• End of Project. Joint report on the work conducted in the VIRGO Project. VIRGO enhancements to local demonstrators.

The above reports will also form the basis for internal discussions and relevant joint publications.

Research Effort of the Participants

The likely research effort that each Participant will contribute to the joint programme of work is summarized in the following table:

Involvement of Industry

VIRGO has already achieved the endorsement of industrial sponsors. Five companies have shown strong interest in the goals and objectives of the VIRGO Network; these are ALCATEL Espace (France), ITMI APTOR (France), AITEK s.r.l. (Italy), Robosoft (France), and TELEROBOT (Italy). Moreover, the Intelligent Systems Group from Systems Technology Research in Daimler Benz AG (Germany) is highly interested in actively participating in VIRGO with its own budget and resources. These companies, as well as others that will be contacted in the early phases of VIRGO, will be invited to contribute to the early discussions and the definitions of specific goals within the project's tasks. Moreover, companies which develop robotic platforms and vision equipment will be invited to contribute to network workshops.

The industrial involvement outlined above will help focus the network research efforts towards issues entailed in technology applications in the autonomous navigation area.