Curiosity driven exploration of sensory-motor mappings

Abstract

Adaptive mechanisms present numerous benefits to artificial systems interacting with the environment. For instance learning sensory-motor mappings eliminates the necessity of (re)calibration process for active vision systems when the aspects of the system or the environment is changed. The amount of time spent on the adaptation process may be reduced if an efficient strategy is followed. We propose such a mechanism which uses intrinsic motivation for sensory-motor learning. We tested our framework in a series of experiments where the environment the system learned a mapping for was systematically changed. Our curiosity-driven framework yielded distinct exploration patterns where distorted areas were concentrated immediately after a change was applied.