‘Autistic Robots’ for Embodied Emulation of Behaviors Typically Seen in Children with Different Autism Severities Kim Baraka 1,2(B ) , Francisco S. Melo 2 , and Manuela Veloso 3 1 Robotics Institute, Carnegie Mellon University, Pittsburgh, PA, USA kbaraka@andrew.cmu.edu 2 INESC-ID/Instituto Superior T´ ecnico, Universidade de Lisboa, Lisbon, Portugal fmelo@inesc-id.pt 3 Machine Learning Department, Carnegie Mellon University, Pittsburgh, PA, USA mmv@cs.cmu.edu Abstract. The goal of this work is to enable interactions of humans with a humanoid robot that can be customized to exhibit behaviors typically observed in children with Autism Spectrum Disorders (ASD) along dif- ferent severities. In a first step, we design robot behaviors as responses to three different stimulus families, inspired by activities used in the context of ASD diagnosis, based on the Autism Diagnosis Observation Schedule (ADOS-2). We implement a total of 16 (possibly blendable) robot behav- iors on a NAO humanoid robot according to different autism severities along 4 selected features from the ADOS-2. In a second step, we integrate those behaviors in a customizable autonomous agent with which humans can continuously interact through predefined stimuli. Robot customiza- tion is enabled through the specification of a feature vector modeling the behavioral responses of the robot, resulting in 256 unique customiza- tions. Our autonomous architecture enables the robot to automatically detect and respond to parameters of the interaction such as verbal and non-verbal stimuli, as well as sound location. In a third step, we evaluate our designed isolated behaviors in the autonomous system by running a study with three experts. This work paves the way towards potentially novel ways of training ASD therapists, interactive solutions for educating people about different forms of ASD, and novel tasks for ASD therapy with adaptive robots. 1 Introduction Children with Autism Spectrum Disorders (ASD) suffer from impaired commu- nication and social abilities, as well as possibly motor and cognitive skills. ASD manifests itself very differently across individuals, not only in severity but also in the areas of development that it affects [1]. Available diagnostic tools for ASD used by therapists provide us with a behavioral model of such individu- als. They do so by linking a taxonomy of typically observed behaviors to values on a set of features that have been identified to be relevant for characterizing c  Springer International Publishing AG 2017 A. Kheddar et al. (Eds.): ICSR 2017, LNAI 10652, pp. 105–114, 2017. https://doi.org/10.1007/978-3-319-70022-9_11