Learning to Stop: A Simple yet Effective Approach to Urban Vision-Language Navigation

Findings of the Association for Computational Linguistics: EMNLP 2020, pages 699–707 November 16 - 20, 2020. c⃝2020 Association for Computational Linguistics 699 Learning to Stop: A Simple yet Effective Approach to Urban Vision-Language Navigation Jiannan Xiang†, Xin Eric Wang‡, William Yang Wang§ †University of Science and Technology of China ‡University of California, Santa Cruz §University of California, Santa Barbara szxjn@mail.ustc.edu.cn, xwang366@ucsc.edu, william@cs.ucsb.edu Abstract Vision-and-Language Navigation (VLN) is a natural language grounding task where an agent learns to follow language instructions and navigate to specified destinations in real- world environments. A key challenge is to recognize and stop at the correct location, especially for complicated outdoor environ- ments. Existing methods treat the STOP ac- tion equally as other actions, which results in undesirable behaviors that the agent often fails to stop at the destination even though it might be on the right path. Therefore, we propose Learning to Stop (L2STOP), a sim- ple yet effective policy module that differen- tiates STOP and other actions. Our approach achieves the new state of the art on a challeng- ing urban VLN dataset TOUCHDOWN, outper- forming the baseline by 6.89% (absolute im- provement) on Success weighted by Edit Dis- tance (SED). 1 Introduction Vision-and-language navigation (VLN) aims at training an agent to navigate in real environments by following natural language instructions. Com- pared to indoor VLN (Anderson et al., 2018), nav- igation in urban environments (Chen et al., 2019) is particularly challenging, since urban environ- ments are often more diverse and complex. Several research studies (Mirowski et al., 2018; Li et al., 2019; Bruce et al., 2018) have been conducted to solve the problem. In this paper, we also focus on the urban VLN task. As shown in Fig. 1, given a natural language instruction, the agent perceives local visual scene and chooses actions at every time step, learning to match the instruction with the pro- duced trajectory and navigate to the destination. Existing VLN models (Wang et al., 2019; Tan et al., 2019; Ke et al., 2019; Ma et al., 2019b,a; Fried et al., 2018; Wang et al., 2018) seem to ne- glect the importance of the STOP action and treat Figure 1: Vision-and-language navigation task in an ur- ban environment. Our L2STOP agent chooses direc- tions at key points and leverages a stop indicator to pro- duce stop or non-stop signals. all actions equally. However, this can lead to unde- sirable behaviors, also noticed in Cirik et al. (2018); Blukis et al. (2018), that the agent fails to stop at the target although it might be on the right path, be- cause the STOP action is severely underestimated. We argue that the STOP action in the urban VLN tasks is crucially important and deserves special treatment. First, in contrast to errors on other ac- tions that are likely to be fixed later in the jour- ney, the price of stopping at a wrong location is higher, because producing STOP terminates the episode, and there will be no chance to fix a wrong …

Content truncated to 500 words