Allowing humans to interactively train artificial agents to understandlanguage instructions is desirable for both practical and scientific reasons,but given the poor data efficiency of the current learning methods, this goalmay require substantial research efforts. Here, we introduce the BabyAIresearch platform to support investigations towards including humans in theloop for grounded language learning. The BabyAI platform comprises anextensible suite of 19 levels of increasing difficulty. The levels graduallylead the agent towards acquiring a combinatorially rich synthetic languagewhich is a proper subset of English. The platform also provides a heuristicexpert agent for the purpose of simulating a human teacher. We report baselineresults and estimate the amount of human involvement that would be required totrain a neural network-based agent on some of the BabyAI levels. We put forwardstrong evidence that current deep learning methods are not yet sufficientlysample efficient when it comes to learning a language with compositionalproperties.