Extended-Self Recognition for Autonomous Agent Based on Controllability and Predictability

We propose an autonomous agent with a notion of self, aiming to naturally avoid irreparable failures of its own body, even in environments where uncertainty is not prede-signed. The self is not limited to the agent's own body; instead, the boundary between the self and others changes through recognition by the agent. In this study, the self and others are defined in terms of controllability and predictability, which are important factors of uncertainty. These definitions allow the self to be extended, which enables the self to include not only the agent's own body but also objects in the environment. The region of self that the agent recognizes is represented as a self-recognition block, which serves to quantify controllability and predictability. By making the self-recognition block a virtual observable within the agent's recognition, we enable the agent to explicitly handle controllability and predictability, which are initially hidden. We evaluated the proposed autonomous agent in a task of grasping and storing a target object whose controllability and predictability were unknown before being grasped by a robot. The results showed that the storage trajectory drifted 3.5 cm farther away from the storage box as a phenomenon corresponding to the proprioceptive drift in the rubber hand illusion. Furthermore, the target object was successfully stored without hazardous collisions between the grasped object and the storage box that could lead to the robot's irreparable failure, confirming the potential of the proposed agent to handle uncertain environments.