Search

To endow robots with skills mastered by human experts has been a promising research domain in recent years with the massive applications of robots. This thesis introduces an advanced human-to-robot skill transfer method with a novel probabilistic framework. To validate this method, a dedicated human skill dataset built with humanoid motions from our everyday life are created for skill transfer experiments. Prior to proposing the innovative human-to-robot skill transfer framework, I conduct a comprehensive review within the robot skill learning domain. The 3 major problems, skill perception, skill representation and skill learning, and their solving methods are systematically considered in literature review. We start from the core model in learning motion trajectories, namely, the probabilistic mixture model, which are widely applied in skill learning from multiple human demonstrations. Curved Gaussian Mixture Model (CGMM) is proposed as a novel variant of the original probabilistic mixture model, Gaussian Mixture Model (GMM), whose component Gaussians have special curved principal axe. The existing curvatures enhance the performance in modelling non-linear data, such as motion trajectories in the time-state space. Relevant fuzzy object function based parameter estimation algorithms are proposed to learn the CGMM in batch and incrementally. Empirical results discussed in thesis proves the superiority of the CGMM over other state-of-the-arts in data fitting. Curved Gaussian Mixture Regression (CGMR) is a specialised regression method for the CGMM developed in this thesis. A human motion skill learned by the proposed CGMM and regressed with the CGMR in the human-to-robot skill transfer framework. The combination of CGMM and CGMR outperforms other type of motion trajectory encoding and retrieving algorithms. Additionally, in order to learn the skill model from a large amount of demonstration data sequentially, I propose a lifelong skill learning framework using the fuzzy incremental learning algorithm. Sawyer as a typical redundant industrial manipulator is applied to reproduce the learned humanoid skills in this thesis. Once multiple demonstrations from a human skill are perceived by the proposed virtual reality based perception system, a CGMM will be trained on the perceived data via the lifelong learning process accordingly. Then, a regressed trajectory with CGMR is taken as a reference in the Cartesian space to reproduce the joint movements of a robot. To get the joint movements which result in anthropomorphic motions in the Cartesian space, an inverse kinematics with morphologic constrains is proposed. At last, the skill perception, learning and reproducing are tested under the human-to-robot skill transfer framework with a Sawyer robot.

This thesis presents, critical compares and develops new methods to control and optimise locomotion for a range of systems. Jumping and running locomotion skills are examined in detail, and intelligent methods are discussed and adapted to optimise for correct form of motion, and performance outcomes. Existing control techniques are summarised and compared, including traditional analytical methods, central pattern generator oscillator systems, pattern generating neural networks, rule based systems and other specialist methods. Optimisation and learning methods presented in the literature are also summarised, and while several methods exist, modern global search methods were limited to genetic algorithms. This thesis applies particle swarm optimisation and quantum inspired evolutionary algorithms to vertical jump and walking optimisation, comparing their performance to a genetic algorithm. Improvements were developed for both binary and real-value variants of quantum inspired evolutionary algorithms, to benefit performance on the real-value problems involved in locomotion control. These improvements consisted of modifications to their rotation gate operators, including a novel scheme to reduce premature convergence in the binary methods, based on limiting the range of less significant bits. Methods were applied in simulated environments, although they can be adapted to real world robotic control, or for reference in optimising motion in humans. A discussion of the susceptibility of simulation runs to poor physical modelling was presented, as this was a significant problem during research. Results were generally mixed, to the extent that all tested methods may be usefully examined more in future work. The central pattern generators tested generated successful patterns more often than a recurrent neural network, and the results of the optimisation algorithms did not show sufficient advantage of one over the others.

This research investigates detection and classification of surface defects in extruded aluminium profiles in order to replace the traditional eye inspection which is still the method widely used today. Through an extensive literature review it is evident that extruded aluminium surface is not investigated properly, while similar industrial products such as copper strips or rolled steel have attracted more interest. An experimental machine vision system is used to capture images from surfaces of extruded aluminium profiles. Extensive feature selection is investigated and appropriate statistical features from a novel technique based on Gradient-Only Co-occurrence Matrices are used to detect and classify defects. The methodology created in this research, makes use of the Sobel edge detector to obtain the gradient magnitude of the image and is followed by the extraction of statistical texture measures from the gradient, after a transformation of the gradient values. Comparisons are made between the statistical features extracted from the original image (Gray-Level Co-occurrence Matrix) and those extracted from the gradient magnitude using a novel approach (Gradient-Only Co-occurrence Matrix). The features extracted from the image processing are classified by feed-forward artificial neural networks. Experiments were conducted for a three class and a four class case study, with the first consisting of Good Surface, Blisters and Scratches, and the second introducing Die Lines to the classes of the first case study. The artificial neural network training is tested using different combinations of statistical features with different topologies. Features are compared individually and grouped, showing better classification accuracy for the novel technique (98.9%) compared to research standard methodology of gray-level co-occurrence matrices (55.9%).

This dissertation presents and discusses the processes of investigation, implementation, testing, validation and evaluation of several computational intelligence-based systems for solving four large-scale real-world problems. In particular, two industrial problems from the pattern recognition and two from the process optimisation areas are studied and intelligent methods to address them are proposed, developed and tested using real-world data. The first problem investigated is the application of an intelligent visual inspection system for classification of texture images. Two major approaches, incorporating supervised and unsupervised (without a priori knowledge) learning techniques, are considered and neural network based classifiers are trained. The focus is kept on the application of unsupervised non-linear dimensionality reduction techniques in combination with unsupervised classification methods. A number of experiments and simulations are performed to evaluate the proposed approaches and the results are critically compared. Next, a classification problem for timely and reliable identification of emitters of radar signals is investigated. A large data set, containing a considerable amount of missing data is used. Several techniques for dealing with the incomplete data values are employed, including listwise deletion and multiple imputation. Methods incorporating neural network classifiers are studied and the proposed approaches are tested and validated over a number of simulations in the MATLAB environment. The third large-scale problem, presented in this work, addresses the need for optimisation of a thermodynamics first principle-based prediction model for simulation of a major purifying process, used in British Petroleum (BP) refineries. A technique incorporating genetic algorithms is applied for optimising a number of the model parameters and for closing up the gaps between the predicted and measured data. Several functions and a graphical user interface (GUI) tool are implemented in MATLAB to assist the analysis, optimisation, testing and validation of the investigated model. Significant overall improvement in its prediction capabilities is achieved. The final problem, covered in this research work, is the need to improve the convergence rate of a computationally very expensive aerodynamic optimisation process. It is addressed by exploring some physics-grounded heuristics and presenting a novel intelligent approach for automated shape optimisation. A set of basis functions (for spanning the design space) is derived in such a way that they facilitate the work of a time-consuming and expensive computational fluid dynamics (CFD) optimisation process. Two MATLAB-based GUI tools are developed to support the calculation, exploration, testing and validation of the studied approach. Experiments for optimising real aircraft geometry are run on supercomputers through an industrial partner (AIRBUS Operations Ltd). The initial results show very promising opportunities for improving the convergence rate of the slow optimisation process.

This thesis presents results from the developemnt, investigation, testing and evaluation of novel meta-heuristic techniques aiming to further improve the state-of-the-art of algorithms for local minima free Neural Network supervised learning. Several approaches for solving Global Optimisation problems that make use of novel meta-heuristic techniques, so called Low-discrepancy Sequences, and hybrid Evolutionary Algorithms are proposed here, investigated and critically discussed. Furthermore, the novel methods are tested on a number of multimodal mathematical function optimisation problems, as well as on a variety of Neural Network learning tasks, including real-world benchmark datasets. Comparison of the results from the investigated methods with such from standard Backpropagation, Evolutionary Algorithms, and other stochastic approaches (Simulated Annealing, Tabu Search, etc.) is conducted in order to demonstrate their competitiveness in terms of number of function evaluations, learning speed and Neural Network generalisation abilities. Finally, the investigated techniques are applied and tested on real-world problems for the intelligent recognition and classification of cork tiles. An Intelligent Computer Vision system is built. The system includes the following stages: image acquisition; image processing (feature extraction and statistical data processing); Neural Network architecture design; supervised learning utilising the proposed Global Optimisation techniques; and finally, extensive system evaluation. The presented examples and case studies demonstrate that the proposed techniques can be effectively applied for the optimisation of mathematical multimodal functions. The investigated methods are successful in local minima free Neural Network learning, and they can be used for solving real-world industrial problems.