Goal-conditioned self-supervised learning from play for collaborative robots

Abstract

Collaborative robots are expected to work alongside humans to achieve shared goals. Deep reinforcement learning offers high generalization ability, but its exploration process may create physical risks for human partners. In contrast, deep imitation learning is efficient but may have limited generalization ability since it relies on demonstration data. To address these limitations, this study proposes another deep learning-based framework that utilizes ``play data'' collected through teleoperation of a robot based on an operator's curiosity, during interaction with a human partner. The framework consists of a model that infers a latent representation of achievable goals and a model that generates actions based on the inferred latent goal representation. An additional mechanism optimizes the latent goal representation based on human behavior during interaction through the prediction error minimization mechanism. Experimental results on human-robot collaboration tasks demonstrate the effectiveness of the proposed framework.