Trust the Intent Recognition

In this project, we achieved our objective of introducing meaningful behavior between a humanoid neurorobot and a human actor, in which oxytocin (OT) release subserves social reinforcement (or reduction of inhibition), and thus with whom the system will sufficiently trust and cooperative. We designed a child-like virtual robot (“childbot”) in a projected (2.5D) simulation environment with rigid body physics (Webots; Cyberotics Ltd., Lausanne, Switzerland). In order to accomplish this, models of both neocortical regions and limbic structures (amygdala, hypothalamus) had to be implemented.

The “Trust the Intent” (TI) experiment has two conceptual phases: learning and challenge. During learning, both the neurorobot and the human actor hold a rod in one hand. A camera serves as the eye of the robot. The neurorobot brain is programmed to initiate a sequence of motions (vertical or horizontal) for 5 seconds each, for up to 30 seconds total. The human has the choice of mirroring the robot’s motion (“concordant movement”) or intentionally moving his/her rod perpendicular to the robot’s motion (“discordant”). The images of the human are processed by a Gabor filter representing V4 activity of moving edges. if the human performs consistently concordant action, the neurorobot learns to trust the human (oxytocin release into amygdala suppresses its tonic inhibition of cortical decision-making circuits); otherwise, the human is not trusted (default tonic amygdala inhibition to neocortex). During a challenge, the human actor slowly reaches for the another object on the table. If the human is trusted by the neurorobot brain, the neurorobot preemptively hands over the dumbbell to the trusted human; otherwise, it retracts the dumbbell towards itself. This trust is specific to the particular human face (the synaptic weights are preserved between sessions); if another human actor performs the scenario but performs discordant action, that person will subsequently be “distrusted” upon future challenges.