Publiée le 08/04/11 à 15h54

Licence Creative Commons CC-By-NC-ND

Robotic systems today have reached unprecedented levels of perceptual and motor abilities. One of the great challenges now is to equip these robots with cognitive systems that allow them to escape well defined industrial environments and face less constrained situations. In such contexts uncertainty can arise from at least two distinct sources – the changing and probabilistic structure of the environment itself, and periodic sensory and motor failures in the robot due to the increased challenges inherent in a less structured environment. Taking inspiration from biology, particularly from primate neurophysiology can help design cognitive systems architectures to enable robots to adapt to uncertainty and to have a satisfying performance, if not optimal, in very different situations.
The current research exploits the behavioral robustness of the primate prefrontal cortical mechanisms that have evolved to address such uncertainty. Based on known neurophysiology of the lateral prefrontal and anterior cingulate cortex (LPFC and ACC) we developed a neural network model that embodies and extend principals of reinforcement learning. The model relies on Reinforcement learning principles allowing an agent to associate reward values to particular actions and to improve its behavioral policy through trial-and-error. In addition, the model incorporates Meta-learning principles allowing the adjustment of meta- parameters of learning and action selection based on measures of the agent’s performance.
The model previously enabled us to perform model-based analysis on neurophysiological data and to identify meta-learning mechanisms in the prefrontal cortex of monkeys performing simple problem solving tasks. Here the model is used to reproduce monkey behavioral performance with a robot, and to enable the humanoid iCub to adapt to uncertainties introduced by the human during human-robot interaction. The combined results provide