Acquiring Human-Robot Interaction skills with Transfer Learning Techniques

Abstract

International audienceRecently, an increasing interest in the research commu nity is how to enable robots to interact and communicate with humans, which is a major step towards integrating robots in our daily life. Within the large domain of Human- Robot Interaction (HRI), this research area is called Social robotics. Communicating with humans is a complex process, involving several modalities such as speech, facial expressions, gaze, head movements, hand gestures, etc. Social robotics focuses on modeling how humans use these communication channels in order to smoothly interact with each other, and how to endow a robot with similar skills. Modeling such complex multimodal in- teractions has proven to be a challenge.Machine learning is invading this domain. Recent research performed on interactive data (i.e. challenging the problem of generating co-verbal gestures of one participant given the co-verbal gestures of the other and verbal activities of both participants) showed a great potential for machine learning in learning complex HRI behaviors. However, the behavioural models are task- and situation-specific: if anything changes in the learning problem (the input distribution or shape, the interaction scenario, etc), we need to learn everything again from scratch. This means that policies learnt in one environment might not work – or work sub-optimally – in other environments and conditions.A growing interest of research in machine learning is on how to transfer knowledge learned from one task (source task) to a new task (target task) . The idea is to capture the similarities between the source and the target tasks, and exploit them in order to learn the tar- get task faster. The shared knowledge can be implicit, like finding latent variables for the input distributions of both tasks [Pan et al. 2011]. It can also be explicit, like representing robotic knowledge as skills.Our challenge is thus to develop methods to transfer the knowledge model the robot learns on one interactive situation with humans to new tasks/situations - notably to new interactive tasks -, in order to enable rapid learning and adaptation to these new tasks