Representation Learning for Grounded Spatial Reasoning

Representation Learning for Grounded Spatial Reasoning Michael Janner, Karthik Narasimhan, and Regina Barzilay Computer Science and Artificial Intelligence Laboratory Massachusetts Institute of Technology {janner, karthikn, regina}@csail.mit.edu Abstract The interpretation of spatial references is highly contextual, requiring joint inference over both language and the environment. We consider the task of spatial reasoning in a sim- ulated environment, where an agent can act and receive rewards. The proposed model learns a representation of the world steered by instruction text. This design allows for pre- cise alignment of local neighborhoods with corresponding verbalizations, while also han- dling global references in the instructions. We train our model with reinforcement learning using a variant of generalized value itera- tion. The model outperforms state-of-the-art approaches on several metrics, yielding a 45% reduction in goal localization error.1 1 Introduction Understanding spatial references in natural language is essential for successful human-robot communica- tion and autonomous navigation. This problem is challenging because interpretation of spatial refer- ences is highly context-dependent. For instance, the instruction “Reach the cell above the westernmost rock” translates into different goal locations in the two environments shown in Figure 1. Therefore, to enable generalization to new, unseen worlds, the model must jointly reason over the instruction text and environment configuration. Moreover, the rich- ness and flexibility in verbalizing spatial references further complicates interpretation of such instruc- tions. 1Code and dataset are available at https://github. com/JannerM/spatial-reasoning Reach the cell above the westernmost rock Figure 1: Sample 2D worlds and an instruction de- scribing a goal location. The optimal path from a common start position, denoted by a white dashed line, varies considerably with changes in the map layout. In this paper, we explore the problem of spa- tial reasoning in the context of interactive worlds. Specifically, we assume access to a simulated envi- ronment, in which an agent can take actions to inter- act with the world and is rewarded for reaching the location specified by the language instruction. This feedback is the only source of supervision the model uses for interpreting spatial references. The key modeling task here is to induce a repre- sentation that closely ties environment observations and linguistic expressions. In prior work, this issue was addressed by learning representations for each modality and then combining them, for instance, with concatenation (Misra et al., 2017). While this approach captures high-level correspondences be- tween instructions and maps, it does not encode de- 49 Transactions of the Association for Computational Linguistics, vol. 6, pp. 49–61, 2018. Action Editor: Hal Daum´e III. Submission batch: 7/2017; Revision batch: 10/2017; Published 1/2018. c⃝2018 Association for Computational Linguistics. Distributed under a CC-BY 4.0 license. tailed, lower-level mappings between specific posi- tions on the map and their descriptions. As our ex- periments demonstrate, combining the language and environment representations in a spatially localized manner yields significant performance gains on the task. To this end, our model uses the instruction text to drive the learning of the environment representation. We start by converting the instruction text into a real- valued vector …

Content truncated to 500 words