Modular Continuous Learning Framework

Although multiple learning techniques exist to endow robots with different skills, open-ended learning is still an outstanding research problem in robotics. Open-ended learning would provide learning autonomy to robots such that they would not require human intervention to learn. This paper proposes a continuous learning framework consisting of a goal discovery module, a goal management module, and a learning module that can be used to implement open-ended learning in robotics. The framework is highly flexible, as it allows any clustering algorithm to be used for goal discovery and any reinforcement learning algorithm for goal learning. The experimental analysis conducted on a mobile robot supports the validity of the framework. Results show how the robot, when placed in a new environment, autonomously generates and learns new goals, thus forming a continuous learning framework capable of autonomously representing and learning skills in an open-ended way.