Recent advancements in the Socially Assistive Robotics (SAR) has shown a vital potential for the use of social robots in mental health service. Our research (Refer figure) is a holistic approach to extend the benefits of SAR to achieve cognitive rehabilitation among individuals with Intellectual Disability (ID). The number of individuals with ID has increased worldwide, from 182 million in 1990 to 1540 million in 2013. In Spain, there are almost 300,000 individuals with ID. Due to the absence of any promising treatment, ID remains a life span condition and thus providing support to individuals with ID demands a huge physical, economical and emotional support by the worldwide community. Depending upon the cognitive stimulation interaction, the robot can work as a coach/instructor, as a companion or as a play partner, directing the execution of the stimulation activity for the user. A crucial step towards delivering an efficient cognitive stimulation by robots is to make them able to detect user’s feelings and to adjust the experience to fit them. Due to the limited ability in recognition and expression of their emotional state among individuals with ID, the analysis of physiological and behavioral correlation of emotions emerges as the most useful method for monitoring their emotions. However, affective state estimation among such type of users by robots has been challenged by the absence of any multimodal database for individuals with ID. We overcame this issue by introducing the MuDERI dataset. MuDERI is an annotated multimodal dataset of audiovisual recordings, RGB-D videos and physiological signals [Electrodermal activity (EDA) and Electroencephalogram (EEG)] of individuals with ID, recorded in a nearly real world settings. Stimulation activities in an appropriate setting outside the laboratory, usually requires movements on user’s side and hence, motion artifacts significantly affect the quality of the physiological signals, acquired in such contexts. Hence, this scenario demands methods for detection and correction of artifacts in real time with low computational cost. We proposed a computationally efficient wavelet-based method for denosing of EDA signal to assist the monitoring in real time. Now, wavelet and statistical techniques are being applied to extract meaningful features from the denoised physiological signals. Automatic emotional state classification and engagement prediction will be done from above features using assorted machine learning algorithms. This information will be used to plan, learn and adapt the robot behavior in real time during stimulation interventions. The benefits of this research can be easily extended to individuals struggling with a wide range of clinical concerns, including children with Autism Spectrum Disorder (ASD) and people with dementia. Employing SAR to empower cognitive stimulation can benefit cognition of these individuals, allowing increased autonomy and positively affecting their quality of life.