Learning by demonstration is a natural approach that can be used to transfer knowledge from humans to robots, similar to how people teach each other through examples. To date, numerous approaches have been developed for learning by demonstration, focusing on two main aspects of the learning problem: the teaching of "motor skills" and the transfer of high-level knowledge of "tasks". This thesis focused on the topic of high-level task learning. Currently, approaches that address this problem utilize the assumption that the task to be learned consists of a precise sequencing of steps that need to be executed. The methods, therefore, aim to accurately reproduce an exact copy of the provided demonstration. However, these methods may suffer from overspecializations or misinterpretations of the task to be learned, if there is any variation in the training example. This variation could be due to the fact that demonstrations can be affected by noise or even by significant changes in task structure from one training example to another. This thesis presents an approach to addressing this challenge through generalization, from a small number of demonstrations of the same task. The aim is to extract a task representation that encodes the essential information from all the training examples, in the presence of small or even large variation in the training examples. The proposed solution consists of two main components: a representation that enables the learner to store the generalized representation of the task and the learning algorithm that allows the construction of a generalized task representation. The approach is validated in simulation and on a physical robot working in the real world, showing the ability to generalize to a wide range of scenarios that may typically occur in teaching by demonstration.
展开▼
