Learning to grasp information with your own hands

Autonomous robots immersed in a complex world can seldom directly access relevant parts of the environment by only using their sensors. Indeed, finding relevant information for a task can require the execution of actions that explicitly aim at unveiling previously hidden information. Informativeness of an action depends strongly on the current environment and task beyond the architecture of the agent. An autonomous adaptive agent has to learn to exploit the epistemic (e.g., information-gathering) implications of actions that are not architecturally designed to acquire information (e.g. orientation of sensors). The selection of these actions cannot be hardwired as general-purpose information-gathering actions, because differently from sensor control actions they can have effects on the environment and can affect the task execution. In robotics information-gathering actions have been used in navigation [7]; in active vision [4]; and in manipulation [3]. In all these works the informative value of each action was known and exploited at design time while the problem of actively facing un-predicted state uncertainty has not received muc