Sensorimotor Robot Policy Training using Reinforcement Learning

Sammanfattning: Robots are becoming more ubiquitous in our society and taking over many tasks that were previously considered as human hallmarks. Many of these tasks, e.g., autonomously driving a car, collaborating with humans in dynamic and changing working conditions and performing household chores, require human-level intelligence to perceive the world and to act appropriately. In this thesis, we pursue a different approach compared to classical methods that often construct a robot controller based on the perception-then-action paradigm. We devise robotic action-selection policies by considering action-selection and perception processes as being intertwined, emphasizing that perception comes prior to action and action is key to perception. The main hypothesis is that complex robotic behaviors come as the result of mastering sensorimotor contingencies (SMCs), i.e., regularities between motor actions and associated changes in sensory observations, where SMCs can be seen as building blocks to skillful behaviors. We elaborate and investigate this hypothesis by deliberate design of frameworks which enable policy training merely based on data experienced by a robot,without intervention of human experts for analytical modelings or calibrations. In such circumstances, action policies can be obtained by reinforcement learning (RL) paradigm by making exploratory action decisions and reinforcing patterns of SMCs that lead to reward events for a given task. However, the dimensionality of sensorimotor spaces, complex dynamics of physical tasks, sparseness of reward events, limited amount of data from real-robot experiments, ambiguities of crediting past decisions and safety issues, which arise from exploratory actions of a physical robot, pose challenges to obtain a policy based on data-driven methods alone. In this thesis, we introduce our contributions to deal with the aforementioned issues by devising learning frameworks which endow a robot with the ability to integrate sensorimotor data to obtain action-selection policies. The effectiveness of the proposed frameworks is demonstrated by evaluating the methods on a number of real robotic tasks and illustrating the suitability of the methods to acquire different skills, to make sequential action-decisions in high-dimensional sensorimotor spaces, with limited data and sparse rewards.