This paper introduces an automatic system for the prediction of extraversion during the rst minutes interaction between humans and and a humanoid robot. In such interactions the behavioural response of people depends by their personality traits and by their attitude towards robots. A set of non-verbal features is proposed to characterize such behavioural responses. Results obtained using such features on a dataset of adults interacting with the iCub robot show the e ectiveness of this approach.
A crucial challenge for social robots is to adapt their behaviours to the
personality of their interacting partners. They should be able to deal with physical
features, preferences, social behaviours and psicololgical traits that make unique
the personality and the behaviour of each human being[14]. To achieve this,
social robots should accordingly create and maintain a model of known partners,
as well as unknown acquaintances, updating it according to new experiences and
new information[
Several studies focus on the automated online estimation of personality traits of human partners[6], on their in uence on the exchange of verbal and non-verbal signals, as well as on the mechanisms underlying the production of behaviors, emotions and thoughts during interaction with social robots[13][7]. In particular these issues have been investigated in the project EDHHI[5]1, focused on studying social interactions between humans and the humanoid robot iCub.
Part of the dataset collected in EDHHI has been exploited in this study to investigate the possiblity of predicting the personality trait of extraversion from a set of non-verbal features extracted during a rst, short interaction with the robot (i.e., the rst minutes). The focus on the extraversion proposed in this 1 http://www.loria.fr/ sivaldi/edhhi.htm study has been motivated by its greater impact on social behaviours respect to other traits during the interactions[15]. The particular context of a social interaction with a robot can induce novelty and anxiety e ects in people. The presented study uses an estimate of the negative attitude towards the robot[9] displayed by people to incorporate such e ects. 2
robot greeted the participant by genlty waving the hands, looking at him, holding a colored toy in its right hand. Standing in front of the robot, the participant was free to act as he likes: talking to the robot, touching it, and so on. As the participant did not receive any indication by the experimenter, if he wanted to, he could start interacting more actively with iCub, asking questions, picking and giving back the small toy, and so on (see Fig. 2).
Due to the absence of constraints, the design of the experiment is focused on induce spontaneous reactions in the human partners.
Participants : 39 healthy adults without any prior experience with robots volunteered to participate to the experiments (11 male and 28 female, aged 37.8y 15.2y).
Each participant received an ID number to preserve the anonymity of the study and signed an informed consent form to partake in the study Fig. 2. iCub interacting with two participants. and granted us the use of their recorded data and videos.
Data Collection : The dataset from Project EDHHI includes the video stream collected by a Kinect RGB-D sensor (v.1, 30fps) placed above the head of the robot in such a way to retrieve the body and face of the human interacting with the robot. The dataset used in this work includes 39 videos (one for each participants) of the rst minutes of their interaction with iCub, synchronized with the robot events logged by the Wizard-Of-Oz application used to control the robot. The average duration of the videos was 110.1s (SD=63.9s). 3
According to the state of art of psychology on personality traits, the extraversion dimension encompasses speci c facets as sociability, energy, assertiveness and excitement-seeking. The features adopted in this work address those facets[12].
F1) Histogram of Quantity of Motion (h-QoM): The quantity of motion is an approximation of the energy of the movement and it is computed as the area (i.e. the number of pixel) of a Silhouette Motion Image[2] normalised over the area of the silhouette. The histogram included 64 bins to guarantee a consistent dynamics.
F2-3) Histograms of Synchrony and dominance (h-Sync, h-dom): Synchrony and dominance are calculated according to the Event Synchronisation technique[11]. Events considered are: a subset of the iCub actions and the fullbody energy peaks of the participant. Event Synchronisation provides a couple of measures counting: i. the synchrony, as the number of actions occourring quasisimultaneously with respect to the global number of actions occurred through the overall interaction; ii. the dominance, as how many times an action of one of the two interactants comes before the corresponding action performed by the other one. Histograms of such measures have been calculated from sliding windows.
F4) Standard deviation of human-robot distance (STD-d): This feature is computed as the average of the pixels' values of the silhouette extracted from the depth image of the Kinect. 4
The extracted features from the interaction with iCub are used to create a model of the personality of each participant to the EDHHI project.
The NEO-PIR questionnaire is a powerful instrument to evaluate the personality traits of people, however in the studied scenario the model should take in account that people's behaviour could vary accordingly to their attitude towards the robots. This \contextual" information depends on self-believes and previous experiences of the participants with robotics and can strongly a ect their behaviour during the experiment. The scores of the NARS questionnaire have been combined with the NEO-PIR scores to take in account this phenomenon. In particular a Principal Component Analysis has been carried out on a scores vector including: NEO-PIR, NARS-S1, NARS-S2, and N ARS S3 scores. The analysis shown that only the rst principal component was meaningful, with an eigenvalue greater than 1 (PCA's eigenvalues: 2.17, 0.85, 0.56 and 0.32; PCA's component load: 0.32, -0.56, -0.57 and 0.51), revealing that the personality can be captured by its score. The distribution of the values of this rst component is computed and its median allowed the de nition of the two classes for the machine learning process.
The dataset resulted in a 39 (instances) x 72 (features) matrix. A Logistic Regression Classi er (LRC) [10] with penalty parameter C = 1 and L2 norm L2 has been then adopted, using a 10-run, 10-fold, strati ed, cross-validation. Table 1 shows the performances obtained when di erent subsets of features feed used. Notably, the classi cation results using dominance and synchrony information overtake the chance level.
The results obtained are consistent with previous results on prediction of extraversion in human-human interaction using non-verbal features (e.g., [15]). 2. Camurri, A., et al.: Interactive systems design: A kansei-based approach. In: Proc.
Conf. New interf. musical expression. pp. 1{8 (2002) 3. Costa, P., McCrae, R., Rolland, J.: NEO-PI{R. Inventaire de Personnalite revise.
Editions du Centre de Psychologie Appliquee, Paris. (1998) 4. Fumagalli, M., Ivaldi, S., et al.: Force feedback exploiting tactile and proximal force/torque sensing. theory and implementation on the humanoid robot icub.
Autonomous Robots (4), 381{398 (2012) 5. Ivaldi, S., et al.: Towards engagement models that consider individual factors in hri: on the relation of extroversion and negative attitude towards robots to gaze and speech during a human-robot assembly task. arXiv:1508.04603 [cs.RO] pp. 1{24 (2015) 6. McCrae, R.: The ve-factor model of personality. In: P.Corr, Mathhews, G. (eds.)
Handbook of Personality Psychology, pp. 148{161 (2009) 7. Meerbeek, B., Saerbeck, M., Bartneck, C.: Iterative design process for robots with personality. Kerstin Dautenhahn, editeur, AISB2009 Symposium on New Frontiers in Human-Robot Interaction pp. 94{101 (2009) 8. Natale, L., et al.: The icub platform: a tool for studying intrinsically motivated learning. In: B., G.and M., M. (ed.) Intr. Motiv. Learning etc. Springer (2013) 9. Nomura, T., et al.: Experimental investigation into in uence of negative attitudes toward robots on human-robot interaction. AI & SOCIETY 20(2), 138{150 (2006) 10. Pampel, F.: Logistic Regression: a primer. Sage Publications (2000) 11. Quiroga, R.Q., et al.: Event synchronization: a simple and fast method to measure synchronicity and time delay patterns. Phys Rev E 66(4) 12. Rahbar, F., Anzalone, S., Varni, G., Zibetti, E., Ivaldi, S., Chetouani, M.: Predicting extraversion from non-verbal features during a face-to-face human-robot interaction. In: International Conference on Social Robotics. Sprinegr (2015) 13. Tapus, A., et al.: User-robot personality matching and assistive robot behavior adaptation for post-stroke rehabilitation therapy. Int. Serv. Rob. 1(2) (2008) 14. Vinciarelli, A., Mohammadi, G.: A survey of personality computing. IEEE Transactions on A ective Computing 5(3), 273{291 (2014) 15. Zen, G., et al.: Space speaks: towards socially and personality aware visual surveillance. In: 1st ACM Int. Worksh. on Multimodal Perv. Video Anal. pp. 37{42 (2010)