Developmental learning of internal models for robotics

摘要

Robots that operate in human environments can learn motor skills asocially, from self-exploration, or socially, from imitating their peers. A robot capable of doing both can be more adaptive and autonomous. Learning by imitation, however, requires the ability to understand the actions of others in terms of your own motor system: this information can come from a robot's own exploration. This thesis investigates the minimal requirements for a robotic system that learns from both self-exploration and imitation of others. Through self-exploration and computer vision techniques, a robot can develop forward models: internal models of its own motor system that enable it to predict the consequences of its actions. Multiple forward models are learnt that give the robot a distributed, causal representation of its motor system. It is demonstrated how a controlled increase in the complexity of these forward models speeds up the robot's learning. The robot can determine the uncertainty of its forward models, enabling it to explore so as to improve the accuracy of its predictions. Paying attention to the forward models according to how their uncertainty is changing leads to a development in the robot's exploration: its interventions focus on increasingly difficult situations, adapting to the complexity of its motor system. A robot can invert forward models, creating inverse models, in order to estimate the actions that will achieve a desired goal. Switching to social learning, the robot uses these inverse models to imitate both a demonstrator's gestures and the underlying goals of their movement.