Analysing autonomous open-ended learning of skills with different interdependent subgoals in robots

Open-ended learning is a relevant approach allowing the design of robots able to autonomously acquire goals and motor skills useful for solving users' problems. An important challenge in this field involves the autonomous learning of interdependent tasks, where learning one skill requires the achievement of environment states (goals) representing preconditions for the skill that is being learned. Here we enhance and compare two robotic architectures, based on previously proposed works, to study which features favour the learning of goals in the presence of different types of interdependencies. The architectures are tested with a Baxter robot solving a series of tasks, consisting in learning to turn on some button-lights while respecting increasingly complex relations between them. The results show that dealing with goal interdependencies at the high level of the architectures is advantageous with longer goal chains; instead, dealing with the interdependencies at the lower motor-skill level is advantageous when exploration can cause conditions precluding the accomplishment of desired goals.

Automated summary

This study compared two robotic architectures to study the features that favor learning goals in the presence of different types of interdependencies. Results showed that dealing with high-level goal interdependencies is advantageous for longer goal chains, while dealing with low-level motor-skill interdependencies is beneficial for avoiding situations where desired goals cannot be achieved during exploration.