What can I control? A framework for robot self-discovery Aaron Edsinger Computer Science and Artificial Intelligence Laboratory Massachusetts Institute of Technology Cambridge, Massachusetts 02139 US edsinger@csail.mit.edu Charles C. Kemp Computer Science and Artificial Intelligence Laboratory Massachusetts Institute of Technology Cambridge, Massachusetts 02139 US cckemp@csail.mit.edu Abstract In this paper we present a developmental progression for a humanoid robot that uses mutual information to discover controllable perceptual categories. Previously we have shown that the robot can discover a visual category that corresponds with its hand from less than 5 minutes of interaction with a hu- man. Here, we show how this discovery can be used to adapt the robot’s perceptual and motor systems such that the robot can sub- sequently discover its fingers. In this way, a robot can expand its control over the world so that newly mastered actions can open up new realms of influence, and new realms of influ- ence can lead to the mastery of new actions. 1. Introduction A robot can be more than a passive observer of the world as it learns and develops. It can exert influ- ence over its immediate surroundings through, for example, manipulation, vocalizations, or social inter- action. Ideally, a robot would incrementally discover what it can control, adapt its perceptual and motor systems to this discovery, and then refine its ability to influence these aspects of its world. Touchette and Lloyd (Touchette and Lloyd, 2004) have noted in their work on information-theoretic control that the degree to which the final state of a system, Φ, is controlled by the state of a controller, A, can be represented by their mutual information, I (Φ; A). Mutual information provides a general way to quantify the extent to which a system’s actions control observed characteristics of the world. Mutual information can be used to rank visual cat- egories by how well the robot can control them. The highest ranked categories should correspond to con- trollable parts of the robot’s body. Previously, we demonstrated that a robot can use this approach to discover a visual category that corresponds with its Figure 1: Domo, the humanoid robot used in this work, interacting with a person. hand from less than 5 minutes of natural interaction with a human (Kemp and Edsinger, 2006b). In this paper we extend and review this work. We present a developmental progression that is guided by the discovery of controllable perceptual cate- gories. This discovery allows us to expand the robot’s controllers and perceptual system, enabling further discovery. We present two stages of the developmen- tal progression: hand-discovery and finger-discovery. The robot first detects the controllabilty of a visual category corresponding to its hand through random motion of its arm. Next, we extend the perceptual and control system to include this discovery. The robot then discovers the controllability of a visual category corresponding to its finger through finger motion while fixating its hand. Results are presented using the robot shown in Figure 1. The next section of this paper discusses related work. In Section 3. we present the developmental progression and its application to hand and finger discovery. Section 4. describes implementation de- tails for the robot’s perceptual system. In Section 5. we describe quantitative results, and in Section 6. we conclude with a discussion of future work.