Sensory-motor learning in humans and other animals is largely an unsupervised, adaptive, and active process. Exteroceptive and proprioceptive modalities are crucial for motor action generation. Generated actions then alter the perception of the outside world. This forms an action-perception loop that can be modeled by the active efficient coding framework. In this thesis, we propose self-calibrating models based on the active efficient coding framework to autonomously learn the sensorimotor control for tracking and localization.

First, humans learn multisensory eye-hand coordination starting from infancy without supervision. In the past work, visuomotor mapping was often learned using either fiducial markers or pre-defined feature descriptors. We propose an architecture to combine visua...[ Read more ]

Sensory-motor learning in humans and other animals is largely an unsupervised, adaptive, and active process. Exteroceptive and proprioceptive modalities are crucial for motor action generation. Generated actions then alter the perception of the outside world. This forms an action-perception loop that can be modeled by the active efficient coding framework. In this thesis, we propose self-calibrating models based on the active efficient coding framework to autonomously learn the sensorimotor control for tracking and localization.

First, humans learn multisensory eye-hand coordination starting from infancy without supervision. In the past work, visuomotor mapping was often learned using either fiducial markers or pre-defined feature descriptors. We propose an architecture to combine visual and proprioceptive cues to learn an eye-hand coordination task autonomously. Using the iCub simulator we demonstrate that the robots can self-calibrate themselves to track the end effector. The visual sensory representation and visuomotor coupling emerge during learning while the robot babbles its arm like an infant does in early development. The evolved eye control policy has characteristics that are qualitatively similar to human oculomotor plant responses.

Second, the brain frequently computes the sensory consequence of motor actions. For instance, the primary visual cortex was found to encode the mismatch between the actual and predicted visual flow in the retina. Inspired partly by these experiments, we propose a predictive multisensory integration architecture to learn the robot end-effector tracking task. We demonstrate that the use of prediction in multisensory integration enables the agent to incorporate the information from proprioceptive and visual cues better. The robot develops the ability to perform smooth pursuit-like eye movements to track its hand, both in the presence and absence of visual input, and to track exteroceptive visual motions.

Third, we get inspiration from echolocating bats, specifically from the big brown bat (Eptesicus fuscus). Echolocating bats autonomously learn sonar-based localization of targets. In past work, developmental models of bat echolocation often rely on the explicit labels of the target direction in azimuth and elevation. We propose an architecture to autonomously learn echolocation without any direction labels. The model self-calibrates to control the head directly towards the sound source to localize it. The big brown bat also waggles its head to change the relative elevation of the pinna tips during localization tasks. We show that head waggles applied in the form of random head rolls improve the localization performance and the head rotation speed to better localize the target.