Emergence of structured interactions: from a theoretical model to pragmatic robotics Arnaud Revel, Pierre Andry To cite this version: Arnaud Revel, Pierre Andry. Emergence of structured interactions: from a theoretical model to pragmatic robotics. Neural Networks, Elsevier, 2009, 22 (2), pp.116-125. hal-00426214 HAL Id: hal-00426214 https://hal.archives-ouvertes.fr/hal-00426214 Submitted on 2 Dec 2009 HAL is a multi-disciplinary open access L’archive ouverte pluridisciplinaire HAL, est archive for the deposit and dissemination of sci- destinée au dépôt et à la diffusion de documents entific research documents, whether they are pub- scientifiques de niveau recherche, publiés ou non, lished or not. The documents may come from émanant des établissements d’enseignement et de teaching and research institutions in France or recherche français ou étrangers, des laboratoires abroad, or from public or private research centers. publics ou privés. Emergence of structured interactions: from a theoretical model to pragmatic robotics A.Revel ETIS Lab, CNRS UMR 8051, ENSEA P.Andry ETIS Lab, CNRS UMR 8051, Univ Cergy-Pontoise Abstract In this article, we present two neural architectures for the control of socially interacting robots. Beginning with a theoretical model of interaction inspired by developmental psychology, biology and physics, we present two sub- cases of the model that can be interpreted as “turn-taking” and “synchrony” at the behavioral level. These neural architectures are both detailed and tested in simulation. A robotic experiment is even presented for the “turn-taking” case. We then discuss the interest of such behaviors for the development of further social abilities in robots. Key words: Robotics, Social interaction, Neural models, Development 1. Introduction to the interaction, embed fundamental dynamical properties, and make emerge communicative behav- With the democratisation of robotics, an impor- iors such as “ability to synchronize” and “alternate tant effort is made to make those “technological” the exchange”. To support this approach, we take machines more attractive for human beings. For in- inspiration from different scientific fields that have stance, a major focus of interest is put on the expres- to deal with interaction: dynamical systems, devel- siveness and the aspect of the robots [11,27,34]: the opmental psychology, neurobiology. question is then how to design “cute” robots? Nev- In this paper, we will show how “synchrony” and “al- ertheless, if the aim of these approaches is to facili- ternation” can emerge from a simple neural model tate human-machine communication, they often ne- of interaction. Next, the the underlying dynamical glect to consider the dynamics of the interaction be- parameters of the model will be applied to 2 robotic tween two agents. In our opinion, intuitive commu- controllers in order to increase their interactional nication (verbal or not) refers to the ability “to take abilities. turn” into the interaction or to be “synchronized” with the other: in summary, to adapt its own dy- namics to the other’s behavior via the integration of the global dynamical exchange. In consequence, our aim is to design robotic controllers that can adapt Email addresses: [email protected] (A.Revel), [email protected] (P.Andry). Preprint submitted to Elsevier 1 May 2008 2. Neural model of agents interaction only a learning function but also a communication one [36,12]. They have suggested that neo-natal 2.1. Inspiration imitation if a pre-linguistic mode of communication that does not require any interaction protocol but 2.1.1. Developmental psychology can be the basis of subsequent higher level commu- Developmental psychology aims at understanding nication and social abilities. More precisely, Nadel has emphasized the fact that synchrony and “turn- how the sensory, motor and cognitive capabilities of the individual evolve from birth. In the frame of taking” are fundamental for communication [37]. Several recent experiments seem to validate this longitudinal studies applied to populations with se- point of view. Oullier, for instance [41], has proposed lected developmental ages, the researches allow to detect and observe the emergence, or the disappear- a simple experimental paradigm in which pairs of participants facing each other are required to ac- ance of behaviors allowing to formulate hypothesis and models of the underlying sensori-motor and cog- tively produce actions, while provided (or not) with nitive mechanisms. In this discipline, a lot of issues the vision of similar actions being performed by someone else. Results reveal that spontaneous phase are concerned with the progressive rise of communi- cation among babies and young children. synchrony (i.e. in-phase coordinated behavior) be- tween two people emerges as soon as they exchange It has been described that human communication visual information, even if they are not explicitly in- can adopt several modalities [47]: language, paralan- guage and kinesic. Paralanguage are the non-verbal structed to coordinate with each other. In a similar experiment, [26] tries to iden- voice and sounds which we can be emitted. Kinesic stands for the “body-language” (facial expressions, tify whether an interpersonal motor co-ordination gaze, gestures, postures, head and body movements, emerges between two participants when they inten- tionally tried to not co-ordinate their movements haptics and proxemics. Obviously, nonverbal com- munication is anterior to language in development. between each other. The goal of the first two sit- uations was for participants to intentionally co- As our goal is to tackle intuitive communication, ordinate or not co-ordinate their movements be- and as our robots are not verbal, we are essentially interested in that latter kind of communication. tween each other. The results revealed in the “not co-ordinate” condition the emergence of an unin- At a behavioral point of view, it has been shown that in our daily activities the social context is for- tended co-ordination in the frequency domain. What matted by physical interactions. One example is, for is however interesting to notice is that the results also reveal the presence of individual intrinsic mo- instance, the emergence of rhythmic applause in a crowd [38]. Another interesting example is the fact tor properties (motor signature) in the 2 conditions. These results indicate that, when there was infor- that we can adopt a similar posture [9] when inter- mation sharing, participants could not avoid (unin- acting with another. More precociously, adults and neonate participate in a social interaction via head tentionally) coordinating with someone. This is consistent with the results found by movements [43] what seems to suggest that very early in development, humans are equipped to deal [55] in EEG. In a specially designed dual- with social interaction. electroencephalogram system, pairs of participants executing self-paced rhythmic finger movements In particular, Neo-natal imitation (from [32,33,29]), has been widely studied for its with and without vision of each other’s actions were continuously monitored and recorded. After analy- presumed implication in communication. In neo- sis, a pair of oscillatory components (phi1 and phi2) natal imitation the observation by the baby of facial expressions performed by the experimenter gives located above right centro-parietal cortex distin- guished effective from ineffective coordination: in- rise to basic imitation (tongue protrusion, eye blink- ing, vocalizations). As experiments has been done crease of phi1 favored independent behavior and in- with babies aged of a few minutes old, this kind of crease of phi2 favored coordinated behavior. The au- thor’s hypothesis is that the phi complex rejects the imitation is not supposed to be learned. This phe- nomenon asks an important question concerning influence of the other on a persons ongoing behav- ior, with phi1 expressing the inhibition of the human the function of such a low-level imitative behavior. mirror neuron system and phi2 its enhancement. Answering this question, a few developmental psy- chologists have emphasized that imitation has not In this section, we have shown that synchroniza- tion behaviors exist very early in human and per- 2 sists in time. This can suggest that they could have systems [45,42,53]. This has lead to a better under- a early and perennial effect on the structure of inter- standing and modeling of the mechanisms involved personal non-verbal communication. during mutual interaction. As far as we are concerned, this discipline pro- vides a theoretical framework for the study of inter- 2.1.2. Synchrony in biological systems acting/coupled social systems. We have seen that behavioral synchronisation ex- Taking inspiration from non-linear dynamics we ists in social contexts. The question is then: are they propose that basic dynamical social properties may individual mechanisms that can help synchrony be- emerge from the dynamical perception/action inter- tween humans. action of two identical oscillatory systems in inter- In fact, many biological activities in human are action. The main idea is that the value of the oscil- naturally synchronized. For instance, it has been lators embedded in each robot models an internal shown that heartbeat, respiration and locomotion in propensity to “interact” with the other robot. humans walking or running on a treadmill are syn- The oscillator is supposed to be connected both chronized [40]. to sensations and actions: perception modulates this The role of circadian rhythm has also been proved oscillator while this latter modifies the actions (in- to have effect on the oscillation of hormone secre- hibition, modulation, activation...). tion, core body temperature and lymphocyte num- ber [3,35,48]. It has also been shown that we are well equipped 2.2. Formal model of the oscillator to satisfy auto-synchrony (for instance with fingers [22]). What is interesting to note is that the result The oscillator we use is made of 2 neurons (u and is different in function of speed of motion: we are in v) inhibiting each other proportionally to the pa- phase for slow motion, and anti-phase for increasing rameter β.