Institute for
Robotics and Process Control

Deutsch   English

Finished Dissertations

Bin-Picking - New Approaches for a Classical Problem

(Dirk Buchholz, 2015)
(External) This book is devoted to one of the most famous examples of automation handling tasks - the "bin-picking" problem. To pick up objects, scrambled in a box is an easy task for humans, but its automation is very complex. In this book three different approaches to solve the bin-picking problem are described, showing how modern sensors can be used for efficient bin-picking as well as how classic sensor concepts can be applied for novel bin-picking techniques. 3D point clouds are firstly used as basis, employing the known Random Sample Matching algorithm paired with a very efficient depth map based collision avoidance mechanism resulting in a very robust bin-picking approach. Reducing the complexity of the sensor data, all computations are then done on depth maps. This allows the use of 2D image analysis techniques to fulfill the tasks and results in real time data analysis. Combined with force/torque and acceleration sensors, a near time optimal bin-picking system emerges. Lastly, surface normal maps are employed as a basis for pose estimation. In contrast to known approaches, the normal maps are not used for 3D data computation but directly for the object localization problem, enabling the application of a new class of sensors for bin-picking.

Primary reviewer: Prof. Dr.-Ing. F. M. Wahl
Secondary reviewer: Prof. Dr. D. Henrich

On Hierarchical Models for Visual Recognition and Learning of Objects, Scenes, and Activities

(Jens Spehr, 2014)
(External) In many computer vision applications, objects have to be learned and recognized in images or image sequences. This book presents new probabilistic hierarchical models that allow an efficient representation of multiple objects of different categories, scales, rotations, and views.
The idea is to exploit similarities between objects and object parts in order to share calculations and avoid redundant information. Furthermore inference approaches for fast and robust detection are presented. These new approaches combine the idea of compositional and similarity hierarchies and overcome limitations of previous methods. Besides classical object recognition the book shows the use for detection of human poses in a project for gait analysis. The use of activity detection is presented for the design of environments for ageing, to identify activities and behavior patterns in smart homes.
In a presented project for parking spot detection using an intelligent vehicle, the proposed approaches are used to hierarchically model the environment of the vehicle for an efficient and robust interpretation of the scene in real-time.

Primary reviewer: Prof. Dr.-Ing. F. M. Wahl
Secondary reviewer: Prof. Dr. J. M. Buhmann

On-Line Trajectory Generation in Robotic Systems

(Torsten Kröger, 2009)
(External) This PhD thesis focuses on sensor integration in robotics, in particular in robotic manipulation control systems. We consider a mechanical system with multiple degrees of freedom equipped with one or more sensors delivering digital and/or analog sensor signals. There is no question that sensor integration and sensor-based control belong to the dominating domains for the future advancement of robotic systems. Although there has been much research on this objective, there is still one important question that has not been answered yet: If we consider a robot in an arbitrary state of motion, how can we calculate a trajectory, if we want the robot to react instantaneously to unforeseen sensor events?

The core part derives a class of algorithms that generate motion trajectories for robotics systems on-line, that is, within one control cycle (typically one millisecond or less). Such an algorithm is executed in parallel to low-level motion controllers, and systems using it are able to react instantaneously to unforeseen (sensor) events. In order to answer the above question, the algorithm enables switchings from sensor-guided robot motion control (e.g., force/torque or visual servo control) to trajectory-following motion control and vice versa, which is an important feature for the practical realization of sensor-based robot motion control systems. The proposed on-line trajectory generation algorithm acts as an open-loop pose controller at an intermediate control layer that functions as one element of the important bridge between low-level robot motion control and higher-level (sensor-based) motion planning. Furthermore, it enables robotic systems to perform a kind of robotic reflex.

As the first derivation step, the algorithm is developed for systems with one actuator only, and subsequently it becomes extended to be applicable in robotic systems with multiple degrees of freedom. The generated trajectories are time-optimal and synchronized for all degrees of freedom, such that all degrees of freedom reach their target at the same time instant.

Using the proposed on-line trajectory generation algorithm as one control submodule in a hybrid switched-control system simplifies the execution of sensor-guided and sensor-guarded motions. As the algorithm is able to take over control at any time instant and in any state of motion, safe and continuous motions can be guaranteed - even if sensors fail. An additional benefit is that unforeseen (sensor-dependent) switchings of reference frames and/or control spaces becomes possible.

The proposed concept is of a very basic nature and, thus, addresses various fields of robotics, in which sensor integration plays a fundamental role, for example, in service robotics, manipulation control systems, mobile robotics and manipulation, and robotic surgery. Samples and use-cases accompany the thesis in order to provide a comprehensible insight into this interesting and relevant field of robotics.

Primary reviewer: Prof. Dr.-Ing. F. M. Wahl
Secondary reviewer: Prof. Dr. ir. H. Bruyninckx

Automatisierte Programmierung von Robotern für Montageaufgaben

(Ulrike Thomas, 2008)
This thesis proposes new concepts and methods for flexible and efficient programming of robots. The suggested concepts are based on the idea that programs for robots can be generated automatically from CAD data. These concepts have been prototyped to evaluate their performance.

Therefore, a new system has been developed to plan assembly sequences automatically. With this system one can generate and evaluate assembly sequences with respect to various optimization criteria. Planning is based on the assembly-by-disassembly strategy, which applies appropriate graph cut algorithms to identify potential assembly sequences. A new efficient method is suggested for solving the seperability problem accurately. Furthermore, a new probabilistic algorithm is developed to compute disassembly paths in 6d configuration spaces. For evaluation of the suggested methods, the new assembly planning system was applied to several industrial assemblies.

Today, industrial robotic systems employ several sensors for assembly. Applying force torque sensors results in a more flexible and robust assembly by robots. Thus, skill primitives have been developed to provide an open programming interface for using hybrid position, velocity and force control. With skill primitive nets it is possible to program complex robot tasks more efficiently. In this thesis new methods have been developed to generate the here defined skill primitive nets automatically from CAD data. Therefore, triangle meshes of assembly parts are segmented into topological elements, which are the basis for the computation of contact formation graphs. These contact formation graphs are used for the automated generation of skill primitive nets. This concept was evaluated using several industrial assembly tasks.

An alternative method for automated robot programming is the employment of force torque maps. This thesis proposes a solution to generate force torque maps from CAD data automatically. Using particle filters, typical variants of "peg-in-hole" tasks can be executed successfully by robots despite the fact of position uncertainties.

Finally, the solutions developed are integrated into a complete process chain, ranging from the specification of assembly tasks via planning of assembly sequences to the execution of planned assembly tasks using serial and parallel robots.

Primary reviewer: Prof. Dr.-Ing. F. M. Wahl
Secondary reviewer: Prof. Dr. rer. nat. D. Henrich

Sensor-Based Surgical Robotics: Contributions to Robot Assisted Fracture Reduction

(Ralf Westphal, 2007)
Nowadays, the treatment of choice for femoral shaft fractures is the minimally invasive technique of intramedullary nailing. Apart from its advantages, however, the technique also has a number of known shortcomings like a frequent occurrence of malaligned reductions, which may have a significant impact on the functional biomechanics and the rehabilitation process of the patient. The high X-ray exposure, especially for the operating team, is a second point, why it may be desirable to support this surgical procedure by sophisticated technical tools. An interdisciplinary research project between the Hannover Medical School and the Technical University of Braunschweig is investigating the applicability of robotized surgical procedures in this context.

The present thesis originates from this project and shows the potential of supporting surgeons by means of robotic assistant systems, which can perform the reduction process of broken bone fragments. First, the development of a telemanipulator system is presented, with which the reduction can be performed based on 2D and 3D imaging data. This system is evaluated using exposed bones as well as human specimens. Based on the experiences from the telemanipulator system, a new concept for a (semi-)automated robotized fracture reduction procedure is developed. First, the automated reduction computes a reduction trajectory, which minimizes additional traumatization of the patient's soft tissue and subsequently moves the robot along that trajectory utilizing skill primitives incorporating sensor guarded and sensor guided motions.

Furthermore, this work presents automated image analysis methods, enabling the use of a robot as a precise drill guidance tool. Apart from fracture reduction, the insertion of the intramedullary nail and the distal locking of the nail are further challenging operation tasks, which remarkably benefit from the integration of surgical navigation, computer assisted surgery, and robotics. In this context, a concept to integrate Petri-nets and skill primitives within the well-known "Model-View-Controller"-pattern is presented in order to achieve a reliable workflow, which is not limited to strictly sequential execution tasks.

It can be shown that the methods developed in this thesis can improve the precision of these surgical operations and at the same time reduce the X-ray exposure to the patient and the operating team.

Primary reviewer: Prof. Dr.-Ing. F. M. Wahl
Secondary reviewer: Prof. Dr.-Ing. H. Wörn

Realistische haptische Simulation materialabtragender chirurgischer Eingriffe

(Andreas Petersik, 2007)
Since several decades, flight simulators are extensively used for pilot training in the aviation industry. Using such systems, pilots regularly have to practice difficult and dangerous situations. The thus gained skills enable them to cope with real emergency situations much more safely.
In the area of surgery, the use of simulators for learning complex and risky interventions receives increasing attention. The review of the state of the art shows that current research is concentrating on simulators for minimally invasive surgery. The described approaches are mostly mimicking the visual appearance of the surgical scene, less frequently also the haptic impression of a tool tip interacting with soft tissue. Approaches which are capable to simulate material removal of solid structures, e. g. the ablation of bone with a rotating burr, exist only in their infancy.
The goal of this work was therefore to develop algorithms and systems for the haptic rendering of material removal procedures, in which an extended tool like a burr interacts with complex bony structures. The main algorithmic problems consist in the detection of collisions and the calculation of the resulting forces. As an additional challenge, both a high spatial and a high temporal resolution with a haptic update rate of more than 1000~Hz are required for a realistic impression. Furthermore, methods must be developed which allow for a realistic haptic impression in spite of the very limited forces of today's commercially available force feedback devices.
Most of the existing algorithms for collision detection are based on surface models as they are used in classical computer graphics. With such approaches, the computational costs increase with the number of triangles in the scene. Collisions with high-resolution models as they are required in this work cannot be calculated by those approaches on standard PC hardware with the required performance.
For this reason, the presented work proposes a new approach for collision detection between an extended tool and a volumetric anatomical model. In contrast to other volumetric models, the model used here is based on a technique which allows the calculation of the surfaces of segmented objects with subvoxel resolution and thus leads to very detailed isosurfaces. For collision detection, a known approach is extended which represents the tool as a discretized array of surface sample points and search vectors. By checking the surface sample points for collisions and searching for the object's surface along the search vectors, the direction of the restoring force can be approximated. The problem of calculating the magnitude of the force is solved with a vector projection method yielding high precision and speed. For the first time the approach of a haptic proxy, an intermediate virtual object which cannot penetrate other objects, is used in a volumetric model. This leads to a realistic haptic impression even for very thin objects.
For simulating realistic forces during material removal, two different approaches are presented. The first one uses the obvious approach to calculate a drilling force based on the material distribution around the tool. However, since commercial force feedback devices are not capable of rendering the high resulting forces, the calculated forces must be scaled down, which leads to lower forces and thus a high material removal rate which is not realistic. Therefore, a geometrical approach was developed with a haptic proxy that remains on the surface of the object under modification. By allowing the proxy to dive into the object by an adjustable amount, the progression speed can be controlled. Typical effects of different tools are simulated by adding e. g. vibrations depending on speed and type of the tool.
The developed algorithms were implemented in simulation systems for petrous bone surgery and dentistry, respectively. As confirmed by numerous users from different medical fields, these systems provide a realistic haptic feedback. Usefulness of the described applications is validated in two independent studies (Zirkle et al., Laryngoscope, Volume 117 (2), February 2007, Sternberg et al., Int. J. Oral Maxillofac. Surg., Volume 36, Issue 5, 2007). With the presented methods, it is thus possible for the first time to simulate bone removal procedures with realistic haptic feedback.

Erstgutachter: Prof. Dr.-Ing. F. M. Wahl
Zweitgutachter: Prof. Dr. Karl Heinz Höhne

The 3d-Puzzle-Problem - Efficient Methods for pairwise Matching of Three-Dimensional Fragments

(Simon Winkelbach, 2006)
The reconstruction of three-dimensional fragmented objects (3d-puzzle-problem) is a highly relevant task with many applications. The field of application comprises archaeology, surgery, bioinformatics and robotics. Examples are the reconstruction of broken archaeological artefacts, human bone fracture reduction in surgery, protein-docking, and the assemblage of industrial components.

This work considers the whole processing chain, starting from data acquisition with different sensors, the general registration of surfaces, up to special requirements for matching fragments in different applications. In this context, two novel and efficient pairwise matching approaches will be introduced, which are highly robust against measurement inaccuracies, material deterioration and noise. In their basic configuration, both methods search for a relative pose, where the surface contact between all fragments is as high as possible. The first approach is based on a randomized generation of likely pose hypotheses and an efficient forecasting of the contact area. The second approach is based on a deterministic coarse-to-fine strategy without any random variables.

Furthermore, this work discusses how a~priori knowledge of the broken objects (like shape priors, mirror symmetries and symmetry axes) can be used to increase the efficiency, accuracy and robustness. Particularly, it shows how to use a~priori knowledge to reconstruct broken femurs (thigh bones) and pelvis fractures, which is an important building block for computer-assisted fracture reduction in surgery.

In addition to the 3d-puzzle-problem, an automatic matching of surfaces has applications in many other important computer vision related fields. It will be shown that the developed approaches are also applicable for 3d object recognition and pose estimation, as well as for registration of range data.

Primary reviewer: Prof. Dr.-Ing. F. M. Wahl
Secondary reviewer: Prof. Dr.-Ing. H. Burkhardt

Robotersteuerungsarchitektur auf der Basis von Aktionsprimitiven

(Bernd Finkemeyer, 2004)
Today's life has been simplified by robots in an extensive way. Heavy work is done by robots and many applications are not imaginable without them. But the usage of robots is still unflexibel. In many cases, uncertain environments and unforeseen events present significant problems.

To solve these problems, the usage of sensors is essential. Although many adequat sensors and corresponding control algorithms are available, their usage in commercial applications is restricted. The following two questions arise: Why is the mentioned situation still state of the art? Where are the problems in using sensors? Four requirements are formulated to answer these questions. They must be fulfilled for an effective sensor usage. In this thesis, it is analyzed how state of the art robot control systems comply with these requirements.
The acquired knowledge is used to specify the skill primitive. It defines a minimale movement of a robot, which can be sensor guided and sensor guarded at same time. To execute complex robot tasks, skill primitives are combined to skill primitive nets. The skill primitive provides a sophisticated interface to the control system. In contrast to others, it allows not only the parameterization of new desired control values. But it also defines and optimizes the controller for the current task. This thesis introduces the concept of the adaptiv selection matrix. It allows switching between controllers in a knee-jerk way, which is necessary, if the chosen sensor and controller are not able to handle the situation in this instant. With the introduced skill primitive concept, a number of control concepts from literature can be used and easily applied in practice.

The realization of a control system, which uses all advantages of the skill primitive concept, requires a powerful and modular software architecture. It must be possible to distribute the software in a computer network. As current software and middleware concepts do not meet the control requirements, the middleware MiRPA (Middleware for Robotic and Process Control Applications) becomes specified and its real-time ability gets verified.

The complete skill primitiv implementation is based on MiRPA. All software modules and their interactions are modelled in UML (Unified Modelling Language). Extensibility and flexibility are very important properties for the software design.

Some robot tasks like object placing or cell phone battery assembly show the practical use of the skill primitive concept.

Primary reviewer: Prof. Dr.-Ing. F. M. Wahl
Secondary reviewer: Prof. Dr.-Ing. A. Knoll

Beiträge zur Planung, Dekomposition und Ausführung von automatisch generierten Roboteraufgaben

(Heiko Mosemann, 2000)
The goal in robot assembly is to construct a mechanical product consisting of several parts which can be assembled by robots. A high level assembly planning system generates sequence plans specifying the order in which parts are to be assembled to form the desired product and computes the trajectories to bring the parts together. In addition to such sequence plans, task plans for actually performing each assembly operation must be generated. The high level assembly sequence plan is the basis for such lower level plans, and taken together they ensure the efficient and flexible assembly of a mechanical product. The main problem in assembly planning is to find a good sequence plan; the issues to be considered are very complex. One basic idea in most of the work on assembly planning is the assembly by disassembly strategy.

The high level assembly planning system introduced in this thesis uses an assembly hierarchy, an extended cycle finder and physical reasoning to simplify the search for an optimal plan. The assembly planning system has a modular structure and an open architecture. It covers all modern aspects of high level assembly planning and introduces new approaches for physical evaluation of sequence plans like the stability analysis. Assembly stability is a very important constraint for assembly sequence generation and evaluation. The analysis of (sub)assembly stability avoids mating parts which tend to disassemble under the influence of gravity. Furthermore, the number of reorientations which gives a good idea on the value of an assembly sequence depends on the stability of the parts to be assembled. This thesis introduces algorithms to calculate the set of potentially stable orientations of an (sub)assembly considering static friction under uniform gravity. Furthermore, the assembly planning system introduced in this thesis takes the CAD descriptions of assembly components and high level assembly specifications using symbolic spatial relationships as input. The generated assembly sequences are stored in an AND/OR-Graph which constitutes a compact representation of all feasible assembly sequences.

This thesis presents a method to decompose an assembly sequence into skill primitives. Therefore, we analyze the corresponding hyperarc of the underlying AND/OR-graph. Features like the local depart space, symbolic spatial relations, and the tool classify the type of assembly sequence (peg in hole, placements, alignments, etc.). Skill primitives are robot movements or commands for grippers and tools. The Unified Modeling Language is used to model the assembly sequences and skill primitives. A robot control system uses the skill primitives as input to select the desired control scheme (position, force or hybrid). The execution of automatically generated assembly plans by robots is one of the key technologies of modern and flexible manufacturing. During the execution of assembly sequences the robot comes into contact with the environment. Since there are positional and geometrical uncertainties from object representation, robot motion, and sensor information, compliance is used typically to prevent excessive contact forces. The contact forces and the resulting torques provide information about the contact geometry which is used to guide the assembly operation. This thesis presents a new and fast algorithm to identify assembly process states considering static friction under uniform gravity. This identification of assembly process states enables a robot to select and modify its motion strategies adequately according to the state. This thesis gives a symbolic representation of contact states and analyzes the static properties of assemblies at each contact state by using the theory of polyhedral convex cones.

Primary reviewer: Prof. Dr.-Ing. F. M. Wahl
Secondary reviewer: Prof. Dr.-Ing. A. Knoll

Probabilistische Erkennung von 3d Freiformobjekten mit Bayesschen Netzen

(Björn Krebs, 2000)
In this work a new approach to 3d object recognition for free form objects is introduced. The statistical behavior of features and objects is modeled in terms of Bayesian networks. A consistent representation of statistical and geometrical a priori knowledge allows to evaluate the uncertainty within the recognition process. The significance of features to discriminate the set of object, hypotheses and their dependencies to the whole database of objects guarantees the selection of the most probable object hypotheses. The integration of different types of features can be achieved consistently within the framework of a Bayesian net representation. A solution to the segmentation problem is presented on the feature level by geometrical relations and on the object level by a fuzzy ICP matching algorithm. A weighted least square minimization of matched correspondences allows the robust computation of the object locations even in the case of noise.

Reliable features for free form objects like corners, edges and rims are introduced. Efficient algorithms for the extraction of these features forms the basis of robust 3d CAD object recognition systems. By extracting features in range data and CAD data in the same way an automatic construction of Bayesian nets is achieved. The recognition system adapts to the current complexity of the scene by integrating the recognition results from the evidence propagation into a decision reasoning.

The proposed methods are basic concepts for 3d object recognition. The consistent representation in a theoretical and mathematical framework guarantees a sound foundation which makes their application for other purposes of uncertain data processing available, too. The introduction into a new approach to 3d probabilistic recognition is the major goal of this work. The proposed methods are integrated into a robust and flexible 3d object recognition system and evaluated in practical experiments.

Primary reviewer: Prof. Dr.-Ing. F. M. Wahl
Secondary reviewer: Prof. Dr.-Ing. H. Niemann

Bewegungsplanung für mobile Roboter in dynamischen Umgebungen auf Basis automatisch erzeugter statistischer Daten

(Eckhard Kruse, 1998)
In advanced applications, mobile robots have to cope with the presence of both static and dynamic obstacles (e.g. humans). Because motions of dynamic obstacles usually cannot be predicted completely, it is impracticable to plan robot paths which are definitely collision-free. Consequently, most current approaches found in literature resort to reactive techniques. Without any global strategy for reaching the goal position, these approaches may yield rather inefficient robot behavior; planning is done locally and in rather short terms in order to satisfy constraints regarding computation time. For better performance, pre-planning and reactive planning have to be combined. The more a priori knowledge is available during pre-planning, the less efforts have to be spent on inefficient reactive maneuvers. This perspective is largely ignored by current planning approaches. As precise prediction of obstacle motion is not possible, implicitly they presume that motions are completely arbitrary, without any regularity.

This is the starting point of this work; a new concept for motion planning is presented: The prevailing behaviour of dynamic obstacles is described with statistical data. Motion planning benefits from this additional knowledge. Accordingly, the approach is called statistical motion planning. Both preplanning and reactive planning can be enhanced (however, reactive planning is only sketched within this work).

This work begins with an introduction giving an overview of current research regarding motion planning. Then, the theoretical foundation of statistical motion planning is examined in detail. A simplified model yields precise mathematical results. The collision probability, i.e. the probability that the robot encounters an obstacle during its motions, is introduced as important criterion to rate robot motions.

In order to apply the concepts to more realistic problems, the theoretical model is extended. The main objective is to develop a method for planning robot paths with minimum collision probability. An according approach, which works efficiently under practical conditions, is presented. Due to the mathematical models, heuristics and approximations can be eschewed almost completely.

The theoretical model and the motion planning approach presume statistical data about obstacle motions to be given. A sensing system which observes the workspace is presented. It is capable of gathering statistical data automatically in a real environment.

The statistical motion planning approach and the system which acquires statistical data have been implemented and tested. Experiments are presented in order to illustrate the differences to conventional approaches and to judge the statistically planned paths. It turns out that the theoretical concepts can be put into practice with success. Statistical motion planning offers improvements and new solutions to motion planning in dynamic environments.

Primary reviewer: Prof. Dr.-Ing. F. M. Wahl
Secondary reviewer: Prof. Dr.-Ing. G. Hommel

Modellbasierte automatisierte Greifplanung

(Frank Röhrdanz, 1997)
Grasping has evolved from a somewhat marginal topic to an important field in robotics research. This increasing interest in grasping is partly due to the increasing importance of flexible assembly in industrial automation.

The thesis describes the model based grasp planning system AutoGRASP for automatically grasping objects in a robot's workspace. In contrast to existing grasp planning systems various constraints are taken into account required for a successful execution of a grasp operation. The computations performed by AutoGRASP are split into offline and online computations, with as much a priori knowledge as possible used in the offline phase.

During the offline phase a geometric grasp planning is performed using the concept of symbolic grasps. Symbolic grasps are generated by filter operations performing a kind of shape matching between the geometry of the gripper and the objects to be grasped. To reduce computational costs, representative gripper orientations are determined for each symbolic grasp. The new concept of representative gripper orientations guarantees, that the gripper's palm can achieve form closure with the objects to be grasped. Thus, higher stability is achieved to resist dynamic disturbance forces arising during the motion of the robot. For each representative gripper orientation collision free approach trajectories and grasp frames are calculated in a local xy-configuration space respective to the objects. The resulting sets of grasp frames define grasp classes that arc evaluated taking into account several evaluation criteria. For the generation of regrasp sequences, placement classes of objects are generated and evaluated. Placement classes describe stable object placements on a horizontal plane. For the evaluation of placement classes a new placement function is proposed yielding obvious evaluation results. During the offline phase grasp and placement classes are used for the generation and evaluation of compatible regrasp operations. The thesis introduces the regrasp graph for the representation of the space of compatible regrasp operations.

During the online phase all necessary external grasp constraints are taken into account. External grasp constraints result while a grasp operation is executed in a robot's workcell. An important external grasp constraint concerns the stability of the scene to be manipulated. In this thesis an efficient algorithm is described which analyses the effects of a grasp operation on scene stability. The new contribution is the proposed general algorithm which can be applied in other related areas of robotic applications as well. If a specified grasp operation cannot be achieved with a single grasp directly, a regrasp sequence automatically is determined. The regrasp problem is solved by an evaluated breadth-first search in the regrasp graph. In contrast to existing regrasp planning systems the initial and goal nodes of the search are determined during the online phase only. A regrasp sequence is determined taking into account the length of the regrasp sequence and the quality of the applied grasp and placement classes.

Primary reviewer: Prof. Dr.-Ing. F. M. Wahl
Secondary reviewer: Prof. Dr.-Ing. R. Dillmann

Reibungsanalyse und Identifikation von Dynamikparametern bei direktangetriebenen und getriebebehafteten Robotern

(Martin Prüfer, 1996)
Robots are widely used in todays applications. The tasks performed by robots range from assembly and welding to automated measurements in quality assurance. A variety of industrial robots has been derived to meet these different needs of different applications. Full utilization of the adapted robots require powerful and more sophisticated control units. Commercially available robot control units neglect this fact and mostly are based on PID control.

Numerous publications cover robot control in theory, but reports on real experimental results are rare. This experimental verification of control concepts has to cope with different practical problems (e.g. measurement noise, real time aspects, security considerations) and therefore, it is time consuming and expensive. Nevertheless, this step is necessary, because many problems don't occur in simulation but will arise during real experiment. This gives a deeper insight into more details and will be the basis for general statements.

This work discusses aspects of parameter identification for direct driven and geared robots. For both types of robots the results of real experiments are presented. Different control strategies will also be discussed and compared with each other. Another important result of this work is the analysis and precompensation of physical effects (friction, ripple).

The thesis is structured as follows: The introduction is followed by a short review of the basics of control, robot kinematics and dynamics, as far as necessary in this context. The third chapter gives the motivation for the construction of an experimental robot arm. In addition, the new developed multiprocessor control unit for industrial robots is introduced. Chapter four is devoted to physical effects. Here the main item is the temperature dependent friction, which could be identified and successfully precompensated. In chapter five the identification of robot dynamic parameters is discussed. The new contribution is the comparison of energy based integral versus differential model. Chapter six compares different control strategies and shows the practical use of knowledge, derived from parameter identification. The thesis is concluded in chapter seven, where also ideas for possible extensions and future work will be given.

Primary reviewer: Prof. Dr.-Ing. F. M. Wahl
Secondary reviewer: Prof. Dr.-Ing. J. U. Varchmin

Global Observation System for Control and Surveillance of Driverless Transport Systems in Production Environments

(Claudio Laloni, 1995)
Driverless transportation systems are increasingly used in automation technology in the field of production engineering in particular. One reason is that these kind of systems offer the possibility to build a flow of materials with modifiable connection structure between the different working cells and working areas respectively. The transportation systems used today require a fixed transport network inside of which they can move. As a consequence both start- and destination point must be reachable directly from the network. If a vehicle meets an obstacle either it must be removed or an alternative path through the transport network must be found. Guide wires or identification labels are installed along the net. The flexibility provided by such a transport network can not be compared with the one of manually controlled vehicles like forklifts because on the one hand no transports can take place between two arbitrary points and on the other hand the system can not react to modifications of the workspace.
The department of robotics and process control of the technical university of Braunschweig develops the transport system MONAMOVE (MOnitoring and NAvigation for MObile VEhicles) in order to provide the mentioned flexibility nevertheless.
In this work concepts and methods for the global observation system have to be acquired and validated. A layer model forms the basis of the entire system. It consists of a sensor layer, an analysis layer and a coordination layer. In a first step the functions of the system are substantiated and the resulting system requirements are derived. Subsequently the sensor layer is described in detail. Amongst others it is discussed which kind of sensor technology is applicable. Furthermore both problems concerning the sensor calibration process and aspects of the improvement of the sensor data quality are considered. The next section is about the analysis layer which interprets the given sensor data and uses specialized modules in order to detect modifications of the drivable area. At last the coordination layer is illustrated. Its function is to coordinate the application of the different modules of the analysis layer and it guarantees the essential temporal continuous surveillance of the workspace. This work ends with the evaluation of the experimental results concerning the prototype-like implemented system layers.

Primary reviewer: Prof. Dr.-Ing. F. M. Wahl
Secondary reviewer: Prof. Dr. Erwin Paulus

Ein aktives 3D-Robotersensorsystem auf der Grundlage eines verallgemeinerten Ansatzes zur Erstellung modellbasierter Objekterkennungsverfahren

(Thomas Stahs, 1994)
Building powerful systems for the recognition and pose estimation of three-dimensional, industrial objects is still a very challenging research goal in the field of robotics and computer vision. This is the basic motivation for the present thesis as well, but in contrast to most existing techniques the approach presented here is not restricted to a single application or a special type of objects. To achieve this, the model based object recognition problem is considered in a first step on a more abstract level leading to a generalized conception for the realization of such systems. This conception is based on a new recognition strategy widely independent of particular object features and thus applicable to a large class of recognition problems. For a fast and controlled adaptation of this general approach to a special application it is embedded into a new methodology for the sytematized construction and evaluation of such systems. As a basic feature, this framework for the realization of model based object recognition systems strongly emphasizes an integrated view towards a new system. This is an essential precondition for the mutual adjustment of its components and thus for an efficacious utilization of geometrical model information. In addition to this integrated view, the abstraction of particular object features facilitates a comprehensive analysis of the model based recognition problem and leads to new, general design criteria for all involed system components. The practical realization of special recognition systems according to this framework is finally supported by a so-called specification- and evaluation-catalogue summarizing all problem dependent decisions and all criteria for a complete system evaluation respectively.

The subsequent presentation of the 3D-robot-sensor-system for the recognition and pose estimation of industrial parts demonstrates the use of this new framework for a typical application. Due to the key role of the sensor for the recognition performance of every system, a first step introduces a new and very powerful active sensor concept. The 3D-robot-sensor developed for this purpose is based on the well-known Coded-Light-Approach and allows the fast generation of range images in a robot work cell from problem dependent viewing positions. Subsequent steps present the remaining system components starting with special analysis tools for local range image processing, a systemspecific modelling concept for the description of objects and finally the recognition process itself adapted for this special application. Of course, all of these components are designed, mutually adjusted and evaluated according to the criteria of the specification- and evaluation-catalogue mentioned above. In order to reveal all strengths and weaknesses of this special system, numerous experimental results illustrate the evaluation of its components as well as their mutual adjustment down to a quantitative level. Additional experiments demonstrate the behaviour of the overall system for scenes with industrial parts of increasing complexity. Compared to existing approaches the improved recognition performance of this new system is mainly achieved by the active range sensor concept combined with the extensive utilization of geometrical object features according to the recognition strategy proposed in the basic framework.

For the future, this new framework shows a feasible way to overcome the prototype character of today's recognition systems. Instead of this, building a model based recognition system should be regarded as a conceptually well-founded and controlable software-engineering-task. The present thesis is intended as a first step in this direction.

Primary reviewer: Prof. Dr.-Ing. F. M. Wahl
Secondary reviewer: Prof. Dr.-Ing. H. Burkhardt

Automatic Driverless Transportation Systems in Dynamic Environments

(Ralf Gutsche, 1993)
Today driverless transportation systems are often used in the production process within factories which normally have a fixed predefined route network. Hence a path between a start-and destination pose must lead along this network. In order to apply such a system in an efficient manner it is necessary to assure that the path is a collision free one during the work. Nevertheless if a part of the path should be blocked by an obstacle a reaction and rescheduling respectively can not take place until the vehicle entered the appropriate section of the path. For example, if the obstacle is located in a narrow corridor such that no local collision avoiding path is available, in this case the vehicle has to travel back a large part of the already moved section in the opposite direction in order to reach its destination on an alternative route.
The goal of this work is to compensate the mentioned disadvantages by introducing the so-called MONAMOVE system (Monitoring and Navigation system for Mobile Vehicles) for foresighted path planning in a dynamic environment . The basic ideas of this system are to be described and new concepts for path planning have to be developed. Fundamentally the whole system consists of four basic components: the global observation system, a world model with user interface, a navigator and a pilot. Thereby the navigator is the foresighted part of MONAMOVE. It calculates a possible path for the robot by taking into account static and dynamic obstacles. In contrast to that the pilot is responsible for the reactive planning. This includes the capability to avoid suddenly emerging obstacles. In this work a path planning method is considered at first which produces right hand oriented paths. Thus vehicles moving along the same direction in a corridor form a convoy which makes it easier for the system to follow a collision avoidance strategy since other dynamic obstacles like human beings often act in the same way. In a further step no one these explicit rules are considered. The paths are planned by means of statistical data of dynamic obstacles. In the end multi-vehicle-coordination is addressed based on geometric crossroads which a calculated automatically by the navigator.

Primary reviewer: Prof. Dr.-Ing. F. M. Wahl
Secondary reviewer: Prof.Dr. Ing. J.Hesselbach

Zur Berechnung geschlossener Lösungen des inversen kinematischen Problems

(Harald Rieseler, 1991)
Sorry, currently in German language only...
Der Entwurf und der Einsatz von Industrierobotern werfen eine Reihe fundamentaler Fragen auf. Von besonderem Interesse ist die Frage nach der geschlossenen Lösung des inversen kinematischen Problems, d.h. nach einer Abbildung einer vorgegebenen kartesischen Position und Orientierung eines Werkzeugs am Ende eines Roboterarms in die zur Realisierung dieser Lage erforderlichen Gelenkstellungen des Roboters. Dabei zeigt sich, daß die mathematischen Beziehungen i.allg. eine so hohe Kompliziertheit erreichen, daß eine manuelle Lösungsherleitung nahezu unmöglich wird.

In der vorliegenden Arbeit wird ein Konzept zur automatisierten, symbolischen Herleitung geschlossener Lösungsformeln dieses inversen kinematischen Problems (IKP) für nicht-redundante Roboter vorgestellt. Den Kern dieses Konzepts bildet ein Satz von Prototypgleichungen zur Herleitung von Lösungen für immer wieder auftretende Gleichungskombinationen. Dieses Prototypenkonzept geht einher mit einer Reduktion des komplexen Suchraums trigonometrischer Gleichungen durch eine vor dem Lösungsprozeß stattfindende Extraktion lösungsrelevanter Gleichungsmerkmale.

Es wird gezeigt, daß dieses Konzept geeignet ist, für mächtige Roboterklassen, die über Gleichungen mit einem maximalen Grad von 4 lösbar sind, eine geschlossene Lösung innerhalb von 1-2 Minuten zu ermitteln. Eine dieser Klassen umfaßt alle Roboter, die drei sich ständig in zueinander parallelen Ebenen bewegende Gelenke (= ebene Gelenkgruppe) besitzen. Die Lösbarkeit dieser Klasse über Gleichungen maximal 4. Grades ist ein theoretisches Ergebnis dieser Arbeit, das gleichsam als Einstieg in die Lösungsphilosophie des vorgeschlagenen Invertierungssystems erbracht wird. Des weiteren werden für diese Klasse kinematische Kriterien ermittelt, die auf quadratische Lösungen des IKP führen; eines der entscheidenden Kriterien im Roboterdesign. Ergänzend dazu wird gezeigt, daß alle resultierenden Roboterarme durch das vorgestellte Konzept analytisch invertierbar sind.

Eine Vielzahl von Tests einer ersten Prolog-Implementierung des Gesamtsystems belegen, daß weite Teile der industriell relevanten Roboter automatisch invertiert werden können. Darüber hinaus wird deutlich, daß nur einige quadratisch lösbare Roboter existieren, die von dieser Testimplementierung nicht invertiert werden können, da sie Teilstrukturen enthalten, deren Bearbeitung noch nicht implementiert ist.

Die Arbeit schließt mit einem Einblick in denkbare Anwendungen eines Invertierungssystems in den Bereichen der Robotersimulation und der Invertierung redundanter Roboterstrukturen.

Primary reviewer: Prof. Dr.-Ing. F. M. Wahl
Secondary reviewer: Prof. Dr. F. Krückeberg

It took 0.15s to generate this page.