Unsupervised learning of visuomotor associations

Several scientists have proposed that perception in biological agents arises not solely from bottom-up processes but also from the interplay between motor commands and their sensory outcomes. While this idea has largely remained theoretical, the current work validates it through practical application in robotic control. The foundation for shape and space perception lies in visuomotor associations, which this thesis aims to learn in an unsupervised manner. Initially, an agent gathers motor and sensory data through random exploration, creating a high-dimensional data space encompassing all sensory and motor variables. During training, this data distribution is modeled using local principal component analyzers (local PCA) or kernel PCA. Subsequently, a recall mechanism is employed to complete partial patterns; for instance, given visual input about an object, the system generates a corresponding robot-arm posture for grasping it, akin to recall in recurrent neural networks. This method offers two key advantages: flexibility in input and output dimensions post-training, and resilience against variations in output patterns for the same input. Furthermore, the thesis illustrates that certain perceptual tasks can be addressed using forward models, which predict sensory input based on current sensory data and motor commands, allowing for mental simulation of movements. For example, a mobile robot can determine its position within a ci