Robot Self-consciousness By Inner Speech and Mental Images Generation  Antonio Chella (Engineering, University of Palermo, Palermo, Italy)   C2

Speaking to herself and generating mental images of her body in action are considered significant correlates of self-consciousness. The paper presents a cognitive architecture based on the Standard Model of Mind. The architecture is implemented on a Pepper robot able to generate inner speech and mental images involving the robot in action. The robot's inner speech is overt and audible, i.e., it is loud, and it reenters the robot by the auditory sensors via the robot microphones. Similarly, the robot's mental images are externally visible, i.e., they are depicted on a computer screen, and they reenter the robot by the visual sensors via the robot camera. Therefore, an external user is free to inspect the contents of the robot's self-consciousness. In detail, the generation of robot inner speech and mental images of itself are both sustained by a cognitive cycle involving a rich internal model made up by a 3D model of the robot and the external environment, and a linguistic memory of the robot implemented by a semantic network. For example, when the robot perceives an apple from the sensory input, it elicits a mental image of itself in front of the apple along with the related inner speech. The mental images and the inner speech reenter the robot by the visual and auditory sensors. An expectation generation system, implemented by suitable neural networks, generates expectations in forms of images and phrases of the possible actions the robot may perform, as grasping the apple or searching for similar fruits. The generated images and inner speech phrases are again reentered by the sensory input of the robot, thus repeating the cognitive cycle. A discussion of the importance of inner speech and mental images by the current theories of self-consciousness is provided. Finally, a reflection is offered on the impact of the proposed robot's self-consciousness framework considering the critical goal of transparency in AI systems.