Learning Compositional Tasks from Language Instructions

The ability to combine learned knowledge and skills solve novel tasks is a key aspect of generalization in humans that allows us understand perform described by language utterances. While progress has been made supervised learning settings, no work yet studied compositional reinforcement agent following natural instructions an embodied environment. We develop set photo-realistic simulated kitchen environment allow study the degree which behavioral policy captures systematicity studying its zero-shot performance on held out instructions. show our leverages additive action-value decomposition tandem with attention based subgoal prediction able exploit composition text generalize unseen tasks.