On Learning and Generalizing Representations in a Dynamic Field Based Architecture for Human-Robot Interaction =Aprendizagem e generalização de representações numa arquitetura baseada em campos dinâmicos para interação humano-robô

摘要

Due to the increasing demand for adaptive robots able to assist humans in their everyday tasks, furnishing robots with learning abilities is one of the most important goals of current robotics research. The work reported in this thesis is focused on the integration of learning capacities in an existing Dynamic Field based control architecture developed for natural human-robot collaboration. Specifically, it addresses two important serial order problems that appear in the architecture at distinct but closely coupled levels of abstraction: 1) the learning of the sequential order of sub-goals that has to be followed to accomplish a certain task, and 2) the learning of representations of motor primitives that can be chained to achieve a certain sub-goal. A model based on the theoretical framework of Dynamic Neural Fields (DNFs) is developed that allows the robot to acquire a multi-order sequential plan of a task from demonstration by human tutors. The model is inspired by known processing principles of human serial order learning. Specifically, it implements the idea of two complementary learning systems. A fast system encodes the sequential order of a single demonstration. During periods of internal rehearsal, it acts as a teacher for a slow system that is responsible for extracting generalized task knowledge from memorized demonstrations of different users. The efficiency of the learning model is tested in a real world experiment in which the humanoid robot ARoS learns the plan of an assembly task by observing human tutors executing possible sequential orders of sub-goals. An extension of the basic model is also proposed and tested in a real-world experiment. It addresses the fundamental problem of a hierarchical encoding of complex sequential tasks. It is shown how verbal feedback by the tutor about a serial order error may lead to the autonomous development of a neural representation of a group of sub-goals forming a sub-task. The second serial order problem of learning goal-directed chains of motor primitives is addressed by combining the associative learning mechanism of the dynamic field model with self-organizing properties. Inspired by the basic idea of the Kohonen's map algorithm, it is shown how self-organizing principles can be exploited to develop field representations of motor primitives, like for instance, a specific grasping behaviour, from observed motion trajectories. Moreover, the integration of additional contextual cues (e.g. object properties) in the learning process may cause the splitting of an existing motor primitive representation into two new representations that are context sensitive. In model simulations, it is shown that the learning mechanisms for representing sequential task knowledge in the DNF model can be also applied to establish chains of motor primitives directed towards a final goal (e.g. reach-grasp-place). Such a chained organization has been discussed in the neurophysiological literature to support not only a fluent execution of known action sequences but also the cognitive capacity of inferring the goal of observed motor behaviour of another individual.