Learning from demonstration is an effective method for human users to\ninstruct desired robot behaviour. However, for most non-trivial tasks of\npractical interest, efficient learning from demonstration depends crucially on\ninductive bias in the chosen structure for rewards/costs and policies. We\naddress the case where this inductive bias comes from an exchange with a human\nuser. We propose a method in which a learning agent utilizes the information\nbottleneck layer of a high-parameter variational neural model, with auxiliary\nloss terms, in order to ground abstract concepts such as spatial relations. The\nconcepts are referred to in natural language instructions and are manifested in\nthe high-dimensional sensory input stream the agent receives from the world. We\nevaluate the properties of the latent space of the learned model in a\nphotorealistic synthetic environment and particularly focus on examining its\nusability for downstream tasks. Additionally, through a series of controlled\ntable-top manipulation experiments, we demonstrate that the learned manifold\ncan be used to ground demonstrations as symbolic plans, which can then be\nexecuted on a PR2 robot.\n