Knowledge- and ambiguity-aware robot learning from corrective and evaluative feedback

Abstract In order to deploy robots that could be adapted by non-expert users, interactive imitation learning (IIL) methods must flexible regarding the interaction preferences of teacher and avoid assumptions perfect teachers (oracles), while considering they make mistakes influenced diverse human factors. this work, we propose an IIL method improves human–robot for imperfect in two directions. First, uncertainty estimation is included endow agents with a lack knowledge awareness (epistemic uncertainty) demonstration ambiguity (aleatoric uncertainty), such robot can request input when it deemed more necessary. Second, proposed enables train flexibility using corrective demonstrations, evaluative reinforcements, implicit positive feedback. The experimental results show improvement convergence respect other agent learns from highly ambiguous teachers. Additionally, user study, was found components improve teaching experience data efficiency process.