A learning method for association between vision and ego-motion which is Capable of Adapting to Arbitrary Image Distortion

The present study considers how the function of vision can be constructed from the experiences of a robot. Specifically, a learning method is proposed for the association between vision and ego-motion. A novel concept of a time evolution operator on sensory inputs is introduced. The time variation of sensory inputs is obtained when the operator is applied to the current inputs. Association between the time evolution operator and the egomotion is learned. Once the learning has been performed, ego-motion can be inferred from the operator constructed from sensory inputs. The main feature of this method is that it does not use any knowledge of image construction obtained by optics and so is flexible with respect to arbitrary image distortion. This flexibility is confirmed by a computer simulation. The proposed method may be considered a method of recovery of structure and motion from motion that remains nearly unaffected by arbitrary image distortion.