Learning robot tasks with loops from experiences to enhance robot adaptability

Learning robot task models with loops helps to increase both the applicability and the compactness of task knowledge. In the framework of Experience-Based Planning Domains (EBPDs), previously formalized by the authors, an approach was developed for learning and exploiting high-level robot task models (the so-called activity schemata) with loops. The paper focuses on the development of: (i) a method-Contiguous Non-overlapping Longest Common Subsequence (CNLCS)-based on the Longest Common Prefix (LCP) array for detecting loops of actions in a robot experience; and (ii) an abstract planner to instanciate a learned task model with loops for solving particular instances of the same task with varying numbers of objects. Demonstrations of this system in both real and simulated environments prove its potentialities.