We are at the start of a period of radical social change where robots will enter more and more into our daily lives (Bekey, 2008). As robots become more omnipresent, the problem of making robots more socially compatible with humans will become increasingly pressing (S. Thrun, 2004). We propose an integrative approach to social cognition, merging self- regulation of basic needs and sensorimotor learning, in biological and artificial agents. This project aims at a breakthrough by demonstrating a novel approach for understanding and modeling social behaviour and its implementation in robots.
State-of-the-art approaches to Social Robotics are based on two opposite views on social cognition (Miklósi & Gácsi, 2012). On the one hand, top-down representation-oriented approaches of social cognition state that agents interact based on a “theory-of-mind”, i.e., complex models of the intentions and beliefs of other agents (Carruthers & Smith, 1998). On the other hand, bottom-up behaviour-based approaches (E. A. Di Paolo, 2000; E. Di Paolo & De Jaegher, 2012; Steels, 2011) have emphasized the role of embodiment and reactive control, where even complex modes of social interaction are grounded in basic sensorimotor patterns enabling the dynamic coupling of agents.
None of these approaches have solved alone the issue of conceiving social robots interacting in a natural way with humans, meaning that both approaches are complementary and one cannot neglect neither one nor the other. INSOCO will attempt at bridging both approaches by addressing the question: how higher-level social cognitive representations are shaped from low-level reactive agent interactions (bottom-up) and how these structures, in turn, optimize social behaviour (top-down)?
Addressing this question will require embedding social agents with a cognitive architecture which is both (1) embodied and reactive to allow the first level of social behaviour complexity from the combination of primitive behaviours and (2) able to learn higher-level cognitive representations for top-down control. The Distributed Adaptive Control, a neurobiological valid architecture (Maffei, Santos-Pata, Marcos, Sánchez-Fibla, & Verschure, 2015; Verschure, Pennartz, & Pezzulo, 2014; Verschure, Voegtlin, & Douglas, 2003) follows this bottom-up top-down structure of generating behaviour and is well suited for the purposes of the project and has not fully been applied to a social context, yet.
INSOCO will adopt, test and validate the hypothesis that social behaviour emerges bottom-up from the individual regulation of inter-dependent physiological needs, coupling the agent’s learning processes for the co-acquisition of cognitive representations used in top-down control.
The INSOCO Project we will fulfill the following objectives aligned with the work plan to answer the proposed hypothesis: (O1) elaborate and model the necessary agent components to generate social behaviour: from self-regulation (extending the so-called Allostatic Control meta-regulation of Sanchez et al. 2010) to basic sensorimotor learning (sensorimotor associations built in the adaptive layer of DAC), (O2) integrate them in a social-enabled instantiation of the DAC cognitive architecture, (O3) prepare the technological components to implement them in real Robot-Robot (RR) and Human-Robot (HR) interaction scenarios and (O4) design a set of collaborative use cases and benchmarks to experimentally validate the hypothesis.
SPECS_lab: Director Paul Verschure
PRINCIPAL INVESTIGATOR: Marti Sanchez in collaboration with Clement Moulin
KEYWORDS: Social Robotics, Social sensorimotor contingencies, Cognitive architecture, Self-regulation, Sensorimotor Learning