Demiris, John.

2019-02-15T14:18:01Z

2019-02-15T14:18:01Z

1999

Imitation mechanisms in artificial and biological agents are of great interest mainly for two reasons: from the engineering point of view, they allow the agent to efficiently utilise the knowledge of other agents in its social environment in order to quickly learn how to perform new tasks; from the scientific point of view, these mechanisms are in¬ triguing since they require the integration of information from the visual, memory, and motor systems. This thesis presents a dual-route architecture for movement imitation and considers its plausibility as a computational model of primate movement imitation mechanisms.

The developed architecture consists of two routes, termed passive and active. The active route tightly couples behaviour perception and generation: in order to perceive a demonstrated behaviour, the motor behaviours already in the imitator's repertoire are utilised. While the demonstration is unfolding, these behaviours are executed on internal forward models, and predictions are generated with respect to what the next state of the demonstrator will be. Behaviours are reinforced based on the accuracy of these predictions. Imitation amounts to selecting the behaviour that performed best, and re-enacting that behaviour. If none of the existing behaviours performs adequately, control is passed to the passive route, which extracts the representative postures that describe the demonstrated behaviour, and imitates it by sequentially going through the extracted postures. Demonstrated behaviours imitated through the passive route form the basis for acquiring new behaviours, which are added to the repertoire available to the active route. A stereo vision robotic head, and a dynamically simulated 13 DoF articulated robot are utilised in order to implement this architecture, illustrate its behavioural characteristics, and investigate its capabilities and limitations. The experiments show the architecture being capable of imitating and learning a variety of head and arm movements, while they highlight its inability to perceive a behaviour that is in the imitator's repertoire, if the behaviour is demonstrated with execution parameters (for example, speed) unattainable by the imitator.

This thesis also proposes this architecture as a computational model of primate move¬ ment imitation mechanisms. The behavioural characteristics of the architecture are compared with biological data available on monkey and human imitation mechanisms. The behaviour of the active route correlates favourably with brain activation data, both at the neuronal level (monkey's F5 'mirror neurons'), and at the systems level (human PET and MEP data that demonstrate activation of motor areas during ac¬ tion observation and imagination). The limitations of the architecture that surfaced during the computational experiments lead to testable predictions regarding the beha¬ viour of mirror neurons. The passive route is a computational implementation of an intermodal-matching mechanism, that has been hypothesised to underlie early infant movement imitation (the AIM hypothesis). Destroying the passive route leads to the architecture being unable to imitate any novel behaviours, but retaining its ability to imitate known ones. This characteristic correlates favourably with the symptoms dis¬ played by humans suffering from visuo-imitative apraxia. Finally, dealing with novel vs. known behaviours through separate routes correlates favourably with human brain activation (PET) data which show that the pattern of activation differs according to whether the observed action is meaningful or not to the observer.

http://hdl.handle.net/1842/33651

The University of Edinburgh

Annexe Thesis Digitisation Project 2019 Block 22

Already catalogued

Movement imitation mechanisms in robots and humans

Thesis or Dissertation

Doctoral

PhD Doctor of Philosophy