This thesis describes the design, development and implementation of a robot mounted active stereo vision system for adaptive robot arm guidance. This provides a very flexible and intelligent system that is able to react to uncertainty in a manufacturing environment. It is capable of tracking and determining the 3D position of an object so that the robot can move towards, and intercept, it. Such a system has particular applications in remotely controlled robot arms, typically working in hostile environments. The stereo vision system is designed on mechatronic principles and is modular, light-weight and uses state-of-the-art dc servo-motor technology. Based on visual information, it controls camera vergence and focus independently while making use of the flexibility of the robot for positioning. Calibration and modelling techniques have been developed to determine the geometry of the stereo vision system so that the 3D position of objects can be estimated from the 2D camera information. 3D position estimates are obtained by stereo triangulation. A method for obtaining a quantitative measure of the confidence of the 3D position estimate is presented which is a useful built-in error checking mechanism to reject false or poor 3D matches. A predictive gaze controller has been incorporated into the stereo head control system. This anticipates the relative 3D motion of the object to alleviate the effect of computational delays and ensures a smooth trajectory. Validation experiments have been undertaken with a Puma 562 industrial robot to show the functional integration of the camera system with the robot controller. The vision system is capable of tracking moving objects and the information this provides is used to update command information to the controller. The vision system has been shown to be in full control of the robot during a tracking and intercept duty cycle.

Off-line programming techniques have been developed to improve productivity of advanced robot systems. In the manufacturing industry, they enable robot workcells to be simulated and robot programs to be generated without interrupting the production process. However, due to the poor structural integrity of present-day robots, discrepancies exist between the CAD model used in the off-line programming system and the real robot. These discrepancies severely limit the effectiveness of off-line programming techniques, since manual operator intervention is required to modify the off-line generated trajectories required for the robot to operate in the manufacturing cell. The work presented here addresses the problems associated with some of these discrepancies to enhance off-line programming systems, by concentrating on two aspects of robot kinematic performance. Identification of the 'actual' kinematics of the manipulator using a new calibration methodology enables the static positioning accuracy of the device to be improved. Validation experiments have been performed using the new kinematic calibration methodology. Identification of the 'actual' kinematics of a Puma-560 industrial robot has shown that this robot's average positioning error can be reduced by approximately 93%. By providing the workcell designer with a new performance index known as the Condition Vector as an indication of the variation in robot kinematic performance throughout its workspace, workcells can be arranged and robot postures selected based on desired robot characteristics for prescribed tasks. Validation experiments for the Condition Vector, undertaken on two industrial robots, were not conclusive but provided promising results.

This thesis is devoted to the study and extension of screw image space and its application to kinematic restraint and robotic grasping. In the study of restraint and robotic grasping, problems arise in relation to how to map the conditions and requirements of restraint in a special representational space, and how to map an object in the same space, so that all analysis and synthesis can be performed and further developed in such a space. Meanwhile, it is desirable that such a space would allow us to establish a relationship between equilibrium and geometry of a restraint circumstance, and also a relationship between geometry and algebra. The study is to introduce the restraint mapping in a newly extended screw image space. With the help of a new look at the properties of screws and screw systems, the screw image space is extended with its relevant spaces, the relationship among them is clarified together with a set of definitions. The screw image space is further completed with the study of its entities including hyperplanes, simplexes and polytopes, and further with the partition of the space. A framework is thus established and associated to restraint mapping and the extension of screw image space, and a set of theories is developed to study the entities in screw image space and to apply them to the restraint mapping. The study is based on the linear dependence of screws with detailed algebraic reasoning, which puts forward new properties of zero pitch screw combinations and theory of linear dependence of reciprocal screw systems together with algebraic and geometric reasoning. The study is successfully applied to kinematic restraint and robotic grasping with a set of theorems and methodologies, not only by mapping the restraint of an object, but also by mapping a set of restraint screws along the surface of an object. The graspability of an arbitrary object can thus be determined, the planning and optimisation can be carried out in the screw image space, together with three new invariant quality measures. An optimal grasp is hence achieved with isotropic resistance to arbitrary externally applied forces. The mapping and entities in screw image space are further weighted to account for the stiffness of contacts in dealing with frictional restraint, and planning is thus based on a stiffness weighted mapping. The planning and optimisation are further given in the concept of normal related restraint, and are achieved in the relevant screw image spaces. An augmented space is then established with the introduction of an affine condition. The relationship between the affine solution and the volumetric ratios of sub-simplexes to an n-simplex reconciles the quality measures with the optimisation. With the further introduction of elastic compatibility, frictional grasps are decomposed, and the force equation of equilibrium is augmented. The approach makes it possible to plan grasps in screw image space and to solve them in augmented space. The approach is further used to predict the failure of a specific case of grasping, and gives a satisfactory result, when compared with an experimental result. The final phase of the study is applied to the unknown grasping of unknown objects. By aggregating contact normals and their position vectors of an unknown object by means of newly developed tactile fingertip detectors, and by mapping them into screw image space, the description of an unknown object is completed. The planning and optimisation can thus operate in screw image space, giving a sufficient prediction of a grasp to be applied on the object. Examples and case studies are given through the thesis. Experiments are quoted to demonstrate the implementation of the new system of theories and methodologies in this Thesis. Further, a set of strategies and their methodology is given and incorporated into a package in C++ with a general application to the study of restraint in screw image space. The thesis ends with a concluding chapter reviewing the contents of the thesis and the main achievements of the study, proposing suggestions for further work.

The thesis describes a distributed software environment designed for the development, evaluation and comparison of new techniques in knowledge based control of robot assembly work cells. It has characteristics which fulfil deficiencies within previous systems and contains within it new techniques in task specification, distributed control[1,2], object recognition[3,4] and path planning[5]. The control of the resources within the cell is based upon an extension of the facilities of a classical blackboard architecture to include plan execution. Unlike previous schemes, these additions allow Plethora to reason about the intent of an action, the current state of the cell and asynchronous events within a single framework. It is this seamless operation and extended representational adequacy that allows Plethora to explore new techniques dealing with the uncertainty inherent in a flexible work cell. The task is specified in domain terms and interpreted to produce a partially ordered set of goals. This new technique is based upon a two-stage ordering process using constructional constraints and necessary collision avoidance. Two new methods, one for object identification and the other for path planning, have also been developed using the system. These have two advantages, efficiency and the ability to operate on data from a vision system or Plethora's geometric modeller. Both methods can be completed within the critical times typical of an assembly work cell. Finally, results of an experiment using the system on a laboratory work cell illustrate how it encompasses previous techniques and can be used to develop new techniques not possible with earlier architectures.

No description available.

No description available.

No description available.

No description available.

Hazardous weather conditions significantly limit the operational capability of civil helicopters. This limitation arises from the crew's inability to determine the location of obstacles in the environment by sight. In order to assist the crew in these circumstances a range of equipment and sensors may be installed in the helicopter. However, with multiple sensors on board, the problem of efficiently assimilating the large amount of imagery and data available generates a significant workload. A reduction of the workload may be achieved by the automation of this assimilation (sensor fusion) and the design of a system to guide the pilot along obstacle free paths. In order to provide the guidance to avoid obstacles a system must have knowledge about the obstacles' possible positions and likely future positions relative the system's own aircraft. Since the information being provided by the sensors will not be perfect, (i.e. it will have some uncertainty associated with it), and since the process model, which must be used to predict any future positions, will also be uncertain, the required positions must be estimated. As the dynamics of moving obstacles will be a priori unknown, it will be necessary to learn process models for them. The dynamics of the obstacles cannot be guaranteed to be linear, therefore these process models must be capable of reflecting this non-linear behaviour. The uncertain information produced by the various sensors will be related to required knowledge about the obstacles by a sensor model, however this relationship need not be linear, and may even have to be learned. Currently used estimation techniques (e.g. the ordinary extended Kalman filter) are inadequate for estimating the uncertainty involved in the obstacles' positions for the highly non-linear processes under consideration. Neural network approaches to non-linear estimation have recently allowed process and sensor models to be learned (sometimes implicitly), however these approaches have been quite ad hoc in their implementation and have been even more negligent in the estimation of uncertainty. The main contributions of this research are the design of non-linear estimators which may use process and sensor models that result from learning processes, and the use of the output of these estimators to determine guidance for obstacle free paths through the environment in 3 dimensions.

No description available.