id	author	title	date	pages	extension	mime	words	sentences	flesch	summary	cache	txt
work_ejjref45ynetbnipdfgfc5kmoe	Michael Janner	Representation Learning for Grounded Spatial Reasoning	2018	14	.pdf	application/pdf	6640	571	60	Figure 1: Sample 2D worlds and an instruction describing a goal location. instruction representation enables the model to sustain high performance when handling both local and Text instructions Prior work has investigated human usage of different types of referring expressions to describe spatial relations (Levinson, 2003; Generalization over both environment configurations and text instructions requires a model that the local structure and global spatial attributes inherent to natural language instructions. To that end, our model combines the textual instructions with the map in a spatially localized manner, as opposed to prior work which joins goal representations and environment observations via simpler Figure 5: Reward achieved by our model and the two baselines on the training environments during reinforcement learning on both local and global instructions. Table 2: Performance of models trained via reinforcement learning on a held-out set of environments and Figure 6: Value map predictions for two environments paired with two instructions each.	./cache/work_ejjref45ynetbnipdfgfc5kmoe.pdf	./txt/work_ejjref45ynetbnipdfgfc5kmoe.txt