Continual Robot Learning Using Self-Supervised Task Inference

Endowing robots with the human ability to learn a growing set of skills over course lifetime as opposed mastering single tasks is an open problem in robot learning. While multi-task learning approaches have been proposed address this problem, they pay little attention task inference. In order continually new tasks, first needs infer at hand without requiring predefined representations. paper, we propose self-supervised inference approach. Our approach learns action and intention embeddings from self-organization observed movement effect parts unlabeled demonstrations higher-level behavior embedding joint action-intention embeddings. We construct behavior-matching objective train novel Task Inference Network (TINet) map demonstration its nearest embedding, which use representation. A policy built on top TINet trained reinforcement optimize performance tasks. evaluate our fixed-set continual settings humanoid compare it different baselines. The results show that outperforms other baselines, difference being more pronounced challenging setting, can incomplete demonstrations. also shown generalize unseen based one-shot generalization experiments.