Bildgestütztes Teach-In eines mobilen Manipulators in einer virtuellen Umgebung

In the last years the field of robotics moves away from applications in the industrial sector to the assistance of humans in everyday environments. To meet the challenges of the new application area robots are made up as a combination of a robotic arm and a mobile platform, so to enable the manipulation and transportation of objects based on visual information from cameras. These so-called mobile manipulators are still confined in laboratories since it is not a trivial task to map visual information on manipulator movements. Existing systems are able to grasp and transport objects, however they do not consider obstacles that can obstruct the manipulator movement. This work presents architecture for vision-based mobile manipulation in a partly known environment that considers obstacles for the robotic arm. Furthermore an approach is investigated for learning the desired vision-controlled movement of the manipulator from images acquired in a virtual environment. A pick-and-place scenario is used for evaluation of the realised system. In this scenario the mobile manipulator has to grasp an object from an optimal gripping direction while avoiding the obstacles around it and transport it to another location. The tasks main difficulty lies in the a priori unknown position of the objects, which have to be detected and localised through image processing. The vision-based control of the manipulator has to adapt the executed movement based on continuously acquired image data, so that collisions are avoided and the goal object is approached form the optimal gripping direction. The system should also be flexible in its planning and it try out alternative solutions when the robots current course of action is proven to be unsuccessful. The implemented architecture consists of three layers. The lowest, reactive layer contains specialised feedback loops, which adapt the manipulator movement to changes in the environment. These feedback loops, also known as reactive behaviours or skills, are either explicitly programmed or learned with neural networks. They use the information of the lead to manipulator to a goal position, keep the object in the viewing filed of the cameras or prevent collisions of the manipulator segments with obstacles. In order to enable a complex movement when reaching for an object while at the same time avoiding obstacles, the reactive behaviours should be combined so to contribute to the finale manipulator movement. Bayesian Belief Networks implement in the second, intermediate layer the corresponding behaviour coordination mechanism. They compute the behaviours’ applicabilities and determine thus the next manipulator movement based on the in real-time acquired camera data and the active behaviours. The hierarchical highest, deliberative layer creates plans and decomposes an assigned task in a sequence of subtasks, each of which can be successfully achieved by the coordination of a group of active behaviours. This layer determines also alternative actions in case the mobile manipulator fails to complete a subtask successfully. The components of the reactive and the intermediate layers are trained in a virtual environment with images acquired from virtual cameras. Two vision-based alternatives toe the standard position-based teach-in methods are implemented for assisting the training in the virtual environment. The algorithmic teach-in is used for training the goal-reaching behaviour whereas the stochastic teach-in is applied for the collision avoidance behaviour as well as for the coordination mechanism. The implemented system enables mobile manipulation in a partly unknown environment and considers obstacles when grasping a goal-object. The realised architecture makes the combination of reactive, vision-based behaviours possible; until today only the sequential execution of the reactive layer’s components was implemented for manipulators. Furthermore this work investigated the application of virtual environments for the teach-in of reactive, vision-based behaviours and of the behaviour coordination mechanism. The trained components come successfully into application on the real mobile manipulator.

Grantor dc:publisher

Publikationsserver der RWTH Aachen University

Year dc:date

2005

Author dc:creator

• Matsikis, Alexandros

Contributors dc:contributor

• Kraiss, Karl-Friedrich

• info:eu-repo/classification/ddc/620
• Erweiterte Realität <Informatik>
• Mensch-Maschine-System
• Greifmanipulator
• Teach-in-Methode
• Maschinelles Sehen
• Objekterkennung
• Ingenieurwissenschaften
• Robotik
• mobile Manipulation
• Teach-In
• bildgestützte Steuerung
• virtuelle Umgebung
• Verhaltensfusion

Statement dc:rights

• info:eu-repo/semantics/openAccess

Language dc:language

ger

Identifier

https://publications.rwth-aachen.de/search?p=id:%22RWTH-CONV-114952%22

OAI record oai:identifier

https://publications.rwth-aachen.de/oai2d?verb=GetRecord&metadataPrefix=oai_dc&identifier=oai%3Apublications.rwth-aachen.de%3A52752