Enhancements in virtual robotics : Through simulation of sensors, events and 6pre-emptive' learning

Abstract: Virtual robotics can be used to dramatically improve the capabilities and performance of industrial robotic systems. Virtual robotics encapsulates graphical off-line programming systems and Computer Aided Robotics (CAR). However current virtual robotic tools suffer from a number of major limitations which severely restrict the ways in which they can be deployed and the performance advantages they offer to the industrial user. The research study focuses on simulation of sensors, programming of event based robotic systerns and demonstrates how intelligent robots can be trained adaptive behaviours in virtual environments. Contemporary graphical programming systems for robots can only be used to program limited sections of a robot program, since i) they do not support methods for the simulation of sensors and event detection; ii) they normally use a post-processor to translate programs from a general language to a controller specific language; iii) conternporary robots can not easily adapt to changes in their environments; and iv) robot programs created off-line must be calibrated to adjust to differences between the virtual and real robotic workcells.The thesis introduces a generic sensor model which can be used to model a variety of sensor types. This model allows virtual sensors to work as independent devices. It is demonstrated that using simulated sensors, event-based robot programs can be created and debugged entirely off-line. Off-line programming of event-based robotic systems demands methods for realistic handling of the communication between independent devices and process. The system must also possess the ability to manage and store information describing status and events in the environment. A blackboard architecture has been used in this research study to store environmental conditions and manage inter-process communication.Self-learning robots is a possible strategy to allow robots to adapt to environmental changes and to learn from their experience. If suitable learning regimes are developed robots can learn to detect changes between virtual and real environments thus minimising the need for calibration. Most learning is based on experience and this requires experimental data to be fed to the learning system. This thesis demonstrates that robot controllers using artificial neural networks for knowledge acquisition and storage can be 'pre-emptively learnt' in virtual robotic environments using virtual robots and simulated sensors. The controllers are able to generalise from the information acquired by the virtual sensors operating in the virtual environment. Arguably the biggest obstacle to the use of self learning robotic systems in real applications has been the need to train the 'real robots' extensively in the 'real environment'. 'Pre-emptive learning' removes this problem. Furthermore, it is therefore possible to develop and evaluate new learning regimes using virtual robotic systems. This approach provides an opportunity to create a variety of environments and conditions which would be impractical to create in a real environment (due to constraints of time, cost and availability).