A Robotic Architecture for the Autonomous Open-Ended Learning of Interdependent Tasks

Autonomous acquisition of many different skills is neces- sary to foster behavioural versatility in artificial agents and robots. While the learning of multiple skills per se can be addressed through different machine learning techniques sequentially assigning a series of N tasks to the agent, autonomy implies that the agent itself has the capacity to select on which task to focus at each moment and to shift between them in a smart way. Intrinsic motivations (IMs) have been used in the field of machine learning and developmental robotics [1], [2] as a motivational signal for the autonomous selection of tasks (often called “goals”): the learning progress in accomplishing the tasks is used as a transient reward to select goals in which the system is making the most learning progress [3]. In real- world scenarios, tasks may require specific initial conditions to be performed or may be interdependent, so that to achieve a task the agent needs first to learn and accomplish other tasks. This latter case is of particular interest and it is still an open question from an autonomous open-ended learning perspective. We propose a reinforcement learning (RL) system for robot control that is capable of learning multiple interdependent tasks by treating the selection of tasks/goals as a Markov Decision Process (MDP) where the agent selects goals to maximise its overall competence.