Our research aims at developing bio-inspired cognitive architectures for autonomous learning for robots. We integrate machine learning algorithms with facilitation mechanisms used by children, such as visual attention, to build computational frameworks for self-supervised learning. Indeed, several issues arise when applying established AI methodologies to data gathered in HRI frameworks. The problems are mostly caused by the scarcity of available data together with the diverse nature of the data stream and the inherent variability of HRI settings. To overcome this limitation, it is crucial to exploit the robot’s embodiment in the real world, i.e. the advantage of perceiving directly the world and continuously updating the sensorial data from humans and the environment. The open challenge is therefore to design frameworks aimed at autonomously and effectively managing the data captured by the robot during interactions with humans so that it can use this information to learn in an autonomous and self-supervised way. One way to let the robot autonomously label the data acquired is to rely on multimodal data (e.g., auditory and visual ones) where one modality helps to supervise the other. 
The robot’s ability to organize its own sensorial experience and to retrieve it in future contexts is of utmost importance for personalized, long-term HRI scenarios.

Ref.

• Gonzalez-Billandon, J., Grasse, L., Tata, M., Sciutti, A., & Rea, F. (2020). Self-supervised reinforcement learning for speaker localisation with the iCub humanoid robot. arXiv preprint arXiv:2011.06544.
• Gonzalez-Billandon, J., Belgiovine, G., Sciutti, A., Sandini, G., & Rea, F. (2021). Cognitive architecture aided by working-memory for self-supervised multi-modal humans recognition. arXiv preprint arXiv:2103.09072
• Gonzalez-Billandon, J., Sciutti, A., Sandini, G., & Rea, F. (2020, October). Towards a cognitive architecture for self-supervised transfer learning for objects detection with a Humanoid Robot. In 2020 Joint IEEE 10th International Conference on Development and Learning and Epigenetic Robotics (ICDL-EpiRob) (pp. 1-8). IEEE.
• Gonzalez-Billandon, J., Sciutti, A., Tata, M., Sandini, G., & Rea, F. (2020, May). Audiovisual cognitive architecture for autonomous learning of face localisation by a Humanoid Robot. In 2020 IEEE International Conference on Robotics and Automation (ICRA) (pp. 5979-5985). IEEE.
   

To address social adaptation ability, we equipped iCub with basic cognitive functionalities which would allow it to select the behaviors maximizing the pleasantness of interaction while being guided by an internal motivation system based on an autonomous decision-making process. We took a developmental stance and replicated the social abilities of a toddler, drawing inspiration from a typical interaction between a baby and its caregiver as a scenario to test the framework functionalities. The robot, therefore, acted like a young child, asking for the caretaker’s company or playing on its own. The social adaptive behavior was guided by an internal comfort state that the robot aims to keep at an optimal level, and which varies with the affective and interactive stimuli that it receives from its caretaker.
We also investigated and compared how iCub’s internal dynamics would be perceived by people, both in a condition when iCub did not personalize its behavior to the person, and in a condition where it was instead adaptive.

Ref.

• Tanevska, A., Rea, F., Sandini, G., Cañamero, L., & Sciutti, A. (2020). A Socially Adaptable Framework for Human-Robot Interaction. arXiv preprint arXiv:2003.11410.
• Tanevska, A., Rea, F., Sandini, G., Cañamero, L., & Sciutti, A. (2019, October). Eager to learn vs. quick to complain? how a socially adaptive robot architecture performs with different robot personalities. In 2019 IEEE International Conference on Systems, Man and Cybernetics (SMC) (pp. 365-371). IEEE.

Recent advances in reinforcement learning with social agents have allowed us to achieve human-level performance on some interaction tasks. However, most interactive scenarios do not have as end-goal performance alone; instead, the social impact of these agents when interacting with humans is as important and, in most cases, never explored properly.
We addressed the problem of including social aspects in the learning strategies of artificial agents in a competitive scenario.
We explored different reinforcement learning-based agents playing a competitive card game against human players. The card game in question is The Chef’s Hat Card Game, a 4-player round-based card game, where each person has a restaurant-context role. The game introduces a competitive multi-player scenario designed to be used both in human-human and human-robot interaction (HRI). As a reinforcement learning task, it provides a controllable action-perception cycle, where each player can only perform a restricted set of actions.
Based on the concept of competitive rivalry, our analysis aims to investigate if we can change the assessment of these agents from a human perspective.

Ref.

• Barros, P., Sciutti, A., Bloem, A. C., Hootsmans, I. M., Opheij, L. M., Toebosch, R. H., & Barakova, E. (2021, March). It's Food Fight! Designing the Chef's Hat Card Game for Affective-Aware HRI. In Companion of the 2021 ACM/IEEE International Conference on Human-Robot Interaction (pp. 524-528).
• Barros, P., Tanevska, A., Yalcin, O., & Sciutti, A. (2020). Incorporating Rivalry in Reinforcement Learning for a Competitive Game. arXiv preprint arXiv:2011.01337.
• Barros, P., Tanevska, A., Cruz, F., & Sciutti, A. (2020, October). Moody Learners-Explaining Competitive Behaviour of Reinforcement Learning Agents. In 2020 Joint IEEE 10th International Conference on Development and Learning and Epigenetic Robotics (ICDL-EpiRob) (pp. 1-8). IEEE.

The understanding of actions from multiple view-points is a crucial ability of humans. Evidences show how this skill is developed in children in the first months of life as well as in many other animal species. In particular, we focus on third person to first person view translation as primary and more common form of perspective translation in human robot interaction. This ability also plays a key role in imitation learning. For this reason, we not only try to endow robots with the ability of perspective translation , but also study how to exploit this mapping to be able to imitate actions through motor control. To address this problem, we work with the latest techniques of computer vision and generative machine learning.

Given the widespread adoption of robotics in the context of rehabilitation, it would be desirable that also robot tutors would exhibit a similar ability of understanding and adapting to the learner’s needs both from the physical and the affective perspectives. 
Effective tutoring during motor learning requires to provide the appropriate physical assistance to the learners, but at the same time to assess and adapt to their state, to avoid frustration. With the aim of endowing robot tutors with these abilities, we designed an experiment in which participants had to acquire a new motor ability - balancing an unstable inverted pendulum - with the support of a robot providing fixed physical assistance. 
Our findings highlight that both physical behavior and affective signals could be integrated by an autonomous robot to generate adaptive and individualized assistance, mindful both of the learning process and the partner’s affective state.

Ref.

• Belgiovine, G., Rea, F., Barros, P., Zenzeri, J., & Sciutti, A. (2020, November). Sensing the Partner: Toward Effective Robot Tutoring in Motor Skill Learning. In International Conference on Social Robotics (pp. 296-307). Springer, Cham.
• Belgiovine, G., Rea, F., Zenzeri, J., & Sciutti, A. (2020, July). A humanoid social agent embodying physical assistance enhances motor training experience. In 2020 29th IEEE International Conference on Robot and Human Interactive Communication (RO-MAN) (pp. 553-560). IEEE.