M3IL: Multi-Modal Meta-Imitation Learning

Imitation Learning (IL) is anticipated to achieve intelligent robots since it allows the user teach various robot tasks easily．In particular, Few-Shot (FSIL) aims infer and adapt fast unseen with a small amount of data. Though FSIL requires few-shot data, high cost demonstrations in IL still critical problem. Especially when we want new task, need execute task for assignment every time. Inspired by fact that humans specify using language instructions without executing them, propose multi-modal setting this work. The model leverages image information training phase utilizes both or only testing phase. We also Multi-Modal Meta-Imitation M3IL, which can information. result M3IL outperforms baseline standard proposed settings. Our shows effectiveness importance setting.