Teacher-Learner Interaction for Robot Active Learning

Robots are being adopted in an increasing number of new application areas, such as health care, logistics, and domestic services. Far from the structured environments of industrial settings, interaction between robots and humans will often become necessary and potentially beneficial. Simultaneously, the target audience of robots will grow to include users who lack the technical skills needed to program robots in the traditional manner. This dissertation proposes interactive learning methods based on Active Learning (AL) and Learning from Demonstration (LfD) that allow robots to learn by interacting with humans-in-the-loop. First, LfD, a learning paradigm that allows the user to program robots by providing examples of the desired behaviour, is adopted for the programming of in-contact tasks. To handle the variability of demonstrations collected through kinesthetic teaching, a probabilistic approach to the encoding of force profiles is proposed. Robot learning is then analysed as a collaborative task, showing how LfD often requires the human teacher to be an expert not only at the task in question but also at teaching a robot.To address the consequences of this rarely met requirement, AL, a learning paradigm that allows robots to participate actively in the teaching process by making queries to their teachers, is proposed for two applications: in combination with LfD for the learning of temporal task models, and as an aid for the tuning of robot programs in an End User Programming (EUP) scenario. The proposed AL approaches and their queries are designed taking into account the interaction between the robot and its human teacher. Investigating the interaction aspect of the proposed AL approaches revealed how their efficient queries may not always be optimal when human teachers are not considered ever-present, infallible sources of information. To investigate this issue, a AL approach that selects queries by taking into account the effort needed for the teacher to answer them is proposed and compared to traditional AL strategies, showing how these strategies impact the teacher's error rates, response times, and workload. Studying the interaction aspect further revealed the importance of transparency and the need for tools that expose robots' decisions to their users. A model-agnostic method that generates natural language explanations for robot policies is therefore presented, along with a study of the effect of its explanations on the user's understanding of the robot's policy. In summary, this dissertation investigates robot learning methods, emphasizing how their design should account for the interaction aspect of the training process.