Robot Policy Learning. After constructing and training the human policy generative network, we can assume that we have a reliable way of accounting for the human policy in the robot’s responses and reactions. The goal is not to make the robot purely reactionary, but socially aware instead. It is possible to have a model that takes into account the goals of the robot, as well as the presence and behaviours of other agents in the environment. The generative network serves as a human policy generator and describes how a human is likely to move in a given space based on their previous actions and the established social and cognitive rules. Therefore we can expect our robot to take human social norms into account. A separate deep-learning model for robot policy learning will be implemented. This model aims to optimize the robot’s behaviour within a social context. It employs multi-objective optimization techniques, where the robot’s policy generation takes the human’s policy into account before generating actions for the robot. This approach aims to align the robot’s behaviour with both task objectives and social norms. Deliverable 8.5 due June 2024 will report the finalised universal behaviour translator module based on DNN and the robot’s ability to learn which behaviours to deploy over time.
