Drama, a connectionist model for robot learning: experiments on grounding communication through imitation in autonomous robots

The present dissertation addresses problems related to robot learning from demonstra¬ tion. It presents the building of a connectionist architecture, which provides the robot with the necessary cognitive and behavioural mechanisms for learning a synthetic lan¬ guage taught by an external teacher agent. This thesis considers three main issues: 1) learning of spatio-temporal invariance in a dynamic noisy environment, 2) symbol grounding of a robot's actions and perceptions, 3) development of a common symbolic representation of the world by heterogeneous agents. We build our approach on the assumption that grounding of symbolic communication creates constraints not only on the cognitive capabilities of the agent but also and especially on its behavioural capacities. Behavioural skills, such as imitation, which allow the agent to co-ordinate its actionn to that of the teacher agent, are required aside to general cognitive abilities of associativity, in order to constrain the agent's attention to making relevant perceptions, onto which it grounds the teacher agent's symbolic expression. In addition, the agent should be provided with the cognitive capacity for extracting spatial and temporal invariance in the continuous flow of its perceptions. Based on this requirement, we develop a connectionist architecture for learning time series. The model is a Dynamical Recurrent Associative Memory Architecture, called DRAMA. It is a fully connected recurrent neural network using Hebbian update rules. Learning is dynamic and unsupervised. The performance of the architecture is analysed theoretically, through numerical simulations and through physical and simulated robotic experiments. Training of the network is computationally fast and inexpensive, which allows its implementation for real time computation and on-line learning in a inexpensive hardware system. Robotic experiments are carried out with different learning tasks involving recognition of spatial and temporal invariance, namely landmark recognition and prediction of perception-action sequence in maze travelling. The architecture is applied to experiments on robot learning by imitation. A learner robot is taught by a teacher agent, a human instructor and another robot, a vocabulary to describe its perceptions and actions. The experiments are based on an imitative strategy, whereby the learner robot reproduces the teacher's actions. While imitating the teacher's movements, the learner robot makes similar proprio and exteroceptions to those of the teacher. The learner robot grounds the teacher's words onto the set of common perceptions they share. We carry out experiments in simulated and physical environments, using different robotic set-ups, increasing gradually the complexity of the task. In a first set of experiments, we study transmission of a vocabulary to designate actions and perception of a robot. Further, we carry out simulation studies, in which we investigate transmission and use of the vocabulary among a group of robotic agents. In a third set of experiments, we investigate learning sequences of the robot's perceptions, while wandering in a physically constrained environment. Finally, we present the implementation of DRAMA in Robota, a doll-like robot, which can imitate the arms and head movements of a human instructor. Through this imitative game, Robota is taught to perform and label dance patterns. Further, Robota is taught a basic language, including a lexicon and syntactical rules for the combination of words of the lexicon, to describe its actions and perception of touch onto its body.