Degree Name

Doctor of Philosophy


School of Electrical, Computer and Telecommunications Engineering


Learning from demonstration is an afficient way to attain a new skill. In the context of autonomous robots, using a demonstration to teach a robot accelerates the robot learning process significantly. It helps to identify feasible solutions as starting points for future exploration or to avoid actions that lead to failure. But the acquisition of pertinent observationa is predicated on first segmenting the data into meaningful sequences. These segments form the basis for learning models capable of recognising future actions and reconstructing the motion to control a robot. Furthermore, learning algorithms for generative models are generally not tuned to produce stable trajectories and suff er from parameter redundancy for high degree of freedom robots

This thesis addresses these issues by firstly investigating algorithms, based on dynamic programming and mixture models, for segmentation sensitivity and recognition accuracy on human motion capture data sets of repetitive and categorical motion classes. A stability analysis of the non-linear dynamical systems derived from the resultant mixture model representations aims to ensure that any trajectories converge to the intended target motion as observed in the demonstrations. Finally, these concepts are extended to humanoid robots by deploying a factor analyser for each mixture model component and coordinating the structure into a low dimensional representation of the demonstrated trajectories. This representation can be constructed as a correspondence map is learned between the demonstrator and robot for joint space actions.

Applying these algorithms for demonstrating movement skills to robot is a further step towards autonomous incremental robot learning.