Autonomous Learning of Appropriate Social Distance by a Mobile Robot
Abstract This thesis aims to design an appropriate human-following solution for a mobile robot. The research can be characterised as interactive model building for a Human Robot Interaction (HRI) scenario. It studies possible proposals for the robot system that learns to accomplish the task autonomously, based on the human preference about the positions and movements of the robot during the interaction. A multilayered feedforward network framework with backpropagation is the adopted learning strategy. The research breaks the task of following a human into three independent behaviours: social positioning, human avoidance and obstacle avoidance. Social positioning is the behaviour that moves the robot, via reasonable paths, to the most appropriate location to follow the human. Both the location and the paths reflect the preference of the human, which varies by individual. The main body of the research therefore proposes a using-while-learning system for this behaviour such that the robot can adapt to the human’s preference autonomously. This research investigated multilayered feedforward networks with backpropagation learning to fulfil the social learning task. This learning model is less used in HRI because a complete set of correct training data doesn’t exist as the human preference is initially unknown. The research proposes a novel method to generate the training data during the operation of learning and introduces the concept of adaptive and reactive learning. A novel training scheme that combines the two learning threads has been proposed, in which the learning is fast, robust and able to adapt to new features of the human preference online. The system enables the behaviour to be a real using-while-learning system as no pre-training of any form is needed to ensure the successful performance of the behaviour. Extensive simulations and interactive experiments with humans have also been conducted to prove the robustness of the system.