REAL 2021 - Robot open-Ended Autonomous Learning: A Competition and Benchmark

Gianluca Baldassarre, Davide Montella, Vieri Giuliano Santucci, Emilio Cartoni

Abstract: Open-ended learning is a core research field of developmental robotics that aims to build learning machines and robots that can autonomously acquire knowledge and skills incrementally as infants and children do. Here we present a competition, in its third edition, that aims to create a standard benchmark within the field. The competition focuses on a simulated camera-arm robot that: (a) in a first ‘intrinsic phase’ acquires sensorimotor competence by autonomously interacting with objects through mechanisms such as curiosity, intrinsic motivations, self-generated goals, and reinforcement learning; (b) in a second ‘extrinsic phase’ is tested with tasks unknown in the intrinsic phase to measure the quality of knowledge and skills acquired in the first phase. Participants can install a simulated environment on their machines to handily develop and test their systems, then submit these to the online platform for evaluation and public ranking. As for infants, the benchmark requires the contextual solution of a number of challenges usually tackled in isolation, such as exploration, sparse rewards, object learning, generalisation, task/goal self-generation, and autonomous skill learning. The competition is contributing to foster the research in the field within different scientific communities, and to operationally define the multiple challenges posed by open-ended learning so as to compare how alternative systems can solve them.