The aim of the project is to develop an adaptive social robot that can effectively assist therapists during children therapies, by improving social interaction between the therapist and the child. The robot will thus help in addressing one important problem that therapists have when performing their job, which is the lack of focus and disengagement of children in the therapy. This very often results in additional work of the therapist to keep the children calm (e.g. giving them toys, showing videos, playing music, etc.), thus making the therapy less effective.

To address this problem we used Machine Learning techniques, in particular, we structured the project in two modules: Supervised and Reinforcement Learning modules.
In the Supervised Learning module, the agent analyses videos and audio clips acquired using its sensors (camera and microphone) to perform classification tasks, that is, assigning some values to the features characterizing the set of states of our Markov Decision Process. All these activities are related to the recognition of emotions experienced by the patient at each time interval.
In the Reinforcement Learning module, we have formulated the problem under a Reinforcement Learning point of view, that is, the goal of the agent is staying as much as possible in states in which the child is emotionally well and relates correctly to the object considered. This potentially results in not slowing down the therapy and engaging the physiotherapist in calming the patient. Therefore, we designed an MDP considering each state made up of some particular features.