Learning to Navigate Unseen Environments: Back Translation with Environmental Dropout

Proceedings of NAACL-HLT 2019, pages 2610–2621 Minneapolis, Minnesota, June 2 - June 7, 2019. c⃝2019 Association for Computational Linguistics 2610 Learning to Navigate Unseen Environments: Back Translation with Environmental Dropout Hao Tan Licheng Yu UNC Chapel HillMohit Bansal {haotan, licheng, mbansal}@cs.unc.edu Abstract A grand goal in AI is to build a robot that can accurately navigate based on natural lan- guage instructions, which requires the agent to perceive the scene, understand and ground language, and act in the real-world environ- ment. One key challenge here is to learn to navigate in new environments that are un- seen during training. Most of the existing ap- proaches perform dramatically worse in un- seen environments as compared to seen ones. In this paper, we present a generalizable nav- igational agent. Our agent is trained in two stages. The first stage is training via mixed im- itation and reinforcement learning, combining the benefits from both off-policy and on-policy optimization. The second stage is fine-tuning via newly-introduced ‘unseen’ triplets (envi- ronment, path, instruction). To generate these unseen triplets, we propose a simple but effec- tive ‘environmental dropout’ method to mimic unseen environments, which overcomes the problem of limited seen environment variabil- ity. Next, we apply semi-supervised learn- ing (via back-translation) on these dropped- out environments to generate new paths and instructions. Empirically, we show that our agent is substantially better at generalizabil- ity when fine-tuned with these triplets, outper- forming the state-of-art approaches by a large margin on the private unseen test set of the Room-to-Room task, and achieving the top rank on the leaderboard.1 1 Introduction One of the important goals in AI is to develop a robot/agent that can understand instructions from humans and perform actions in complex environ- ments. In order to do so, such a robot is re- quired to perceive the surrounding scene, under- stand our spoken language, and act in a real-world 1Our code, data, and models publicly available at: https://github.com/airsplay/R2R-EnvDrop … Walk past the piano through an archway directly in front. Go through the hallway when you see the window door. Turn right to the hanged pictures on your right … … !" !"#$ Instruction Bird-view Actions Figure 1: Room-to-Room Task. The agent is given an instruction, then starts its navigation from some staring viewpoint inside the given environment. At time t, the agent selects one view (highlighted as red dotted bounding boxes) from a set of its surrounding panoramic views to step into, as an action at. house. Recent years have witnessed various types of embodied action based NLP tasks being pro- posed (Correa et al., 2010; Walters et al., 2007; Hayashi et al., 2007; Zhu et al., 2017b; Das et al., 2018; Anderson et al., 2018b). In this paper, we address the task of instruction- guided navigation, where the agent seeks a route from a start viewpoint to an end viewpoint based on a given natural language instruction in a given environment, as shown in Fig. 1. The naviga- tion simulator we use is the recent …

Content truncated to 500 words