Learning manipulative skills using an artificial intelligence approach.

The aim of this research was to design a non-linear controller based on an Artificial Neural Network and Reinforcement Learning algorithms implementation, which is able to perform an intelligent robotic assembly of mechanical components. Different information was applied and combined to develop a fully unsupervised, intelligent controller. In the author's design no class labelling or geometry feature pretraining takes place. Only force and torque signals together with the direction of insertion were supplied to the controller. A unique sandwich structure of the intelligent controller was proposed. It featured two major layers, a State Recognition module where the detection and localisation of the contact points were performed, and the Decision Making subsystem where the decision about the next action took place.All the algorithms were implemented and tested on simulated data before being applied to the real-life peg-in-hole insertion. The results are presented in the form of graphs and tables.Evaluation of the environmental uncertainty was accomplished. The signal from the force and torque sensor was acquired under controlled conditions. All the data was collected to establish the area and level of uncertainty (e.g. signal errors) the artificial controller would need to learn to cope with and compensate for.The empirical part of the thesis includes the investigation into the effects of different learning methods applied on the same geometry. The influence of action-selection methods on AI agent performance was analysed. The proposed controller was applied to a set of real life peg-and-hole experiments. Both circular and square peg geometries were used, and insertions into chamfered and non-chamfered holes were performed. Materials with different friction factors were used for mating parts.Fast and stable knowledge acquisition was clearly present in all the cases investigated. A significant reduction in contact force value during the initial stage of the learning process was recorded. The force was usually reduced to one tenth of the initial value. Some fluctuations were recorded but when the cylindrical peg was considered the value of contact forces never exceeded 0.5 N during the steady state.