Coordination mechanisms in joint action

Coordinated actions of multiple agents that bring about a change in the environment are referred as ‘joint action’. Joint action is ubiquitous in our daily life - for example while carrying objects, playing sports, making music, or during rehabilitation. Many of them - for instance, walking in a narrow passage without bumping into one another - are not planned in advance. In particular, when two humans physically interact, sensory consequences and motor outcomes are not entirely predictable as they also depend on the other’s agent actions. It may be that agents predict partners’ actions and their outcomes as well. However, it is not clear whether and to what extent we actually do this. While sensorimotor mechanisms underlying individual actions have been widely investigated and many foundational principles have been identified, a clear picture of how joint action is controlled and develops over time is still missing. This work intends to investigate the perceptual and control mechanisms underlying joint action, with a particular focus on emergent coordination in physical interaction. We explore scenarios in which interacting agents need to negotiate a shared strategy through the haptic channel. To this end, we used a combination of computational and experimental approaches. We specifically focused on a few aspects: (i) the subjective experience of control during joint action, (ii) the translation of a computational model of joint action into an architecture that can be used to develop bio-inspired artificial agents, and (iii) learning and decision-making in joint action. We experimentally addressed physical human-human interaction using a dyadic robotic setup. It consists of a pair of haptic interfaces which can be programmed to implement different types of interaction. We used tasks where players were mechanically coupled and had to perform goal-directed movements in different interaction modalities - sensory modalities and tasks. When addressing joint action it is important to account for the more subjective experiences and feelings which may emerge and affect performances in joint action. To this end, we extended a previously proposed computational model of interaction to describe the sense of agency during an interaction. Sense of agency refers to the subjective experience of control iii we have over our actions and their outcomes. We provided a general framework which reconciles more deterministic accounts of action with the more subjective experiences of control. We proposed that the sense of agency emerges within the sensorimotor control loop and that experiences of self and joint agency coexist during interactions. These are weighted depending on the interactive scenario and actively affect human behavior. Investigating the mechanisms underlying joint action is crucial per se to understand humans behaviors underpinnings. However, these findings may serve the development of biomimetic artificial agents. We used the interaction model to devise an artificial partner architecture that is capable of estimating human action and establishing coordination. This architecture is a promising tool that may be used to further investigate joint action in a more controllable way but also to develop artificial agents which may be used in a variety of applicative scenarios - e.g. automation, and neuro-rehabilitation. When people share the same final goal but have incompatible sub-goals deciding to collaborate does not depend on the information available about the partner. To disentangle these findings we developed a computational model of interaction based on game theory and optimal Bayesian estimation. This can be used to make predictions and analyze experimental data, characterizing the interactive participants. We found that other than sensory reliability, initial attitude toward exploration leads dyads to establish coordination. We also investigated how priming interacting agents with identical or different cues affects interaction and the establishment of coordination in complex arm movements. We observed that people tend to synchronize their motions as the mechanical coupling is activated, however, agents tended to use their preferred initial solution which involves some degree of effort instead of negotiating a shared path. This provides insights into understanding to what extent interacting agents represent the partner and in which measure they use such information to select their own action. Overall, this thesis work attempted to investigate joint action by investigating different aspects of joint action. We not only focused on a mechanistic description of physical human- human interaction, but we also accounted for the more subjective experience of action and suggested an architecture that translates insights from the study of human-human interaction into the development of an application which can be used to control artificial agents which are capable of interacting with humans in a natural and intuitive way. We think that this approach is crucial when addressing a multifaceted and complicated phenomenon as joint action.