• Authors: Jialu Li, Hao Tan, and Mohit Bansal (UNC Chapel Hill)

• Paper

Vision-and-Language Navigation (VLN) tasks require an agent to navigate through the environment based on language instructions. In this paper, we aim to solve two key challenges in this task: utilizing multilingual instructions for improved instruction-path grounding and navigating through new environments that are unseen during training. To address these challenges, first, our agent learns a shared and visually-aligned cross-lingual language representation for the three languages (English, Hindi and Telugu) in the Room-Across-Room dataset. Our language representation learning is guided by text pairs that are aligned by visual information. Second, our agent learns an environment-agnostic visual representation by maximizing the similarity between semantically-aligned image pairs (with constraints on object-matching) from different environments. Our environment agnostic visual representation can mitigate the environment bias induced by low-level visual information. Empirically, %we show that our agent gets large improvements in all metrics when generalizing to unseen environments with the cross-lingual language representation and the environment-agnostic visual representation, and achieving the new state-of-the-art for multilingual agents on the Room-Across-Room dataset. on the Room-Across-Room dataset, we show that our multi-lingual agent gets large improvements in all metrics over the strong baseline model %(over state-of-the-art baselines) when generalizing to unseen environments with the cross-lingual language representation and the environment-agnostic visual representation. Furthermore, we show that our learned language and visual representations can be successfully transferred to the Room-to-Room and Cooperative Vision-and-Dialogue Navigation task, and present detailed qualitative and quantitative generalization and grounding analysis.

1. Follow instructions here to install Matterport3D simulators.

2. Install requirements:

pip install -r python_requirements.txt

3. Download data:

wget -r https://nlp.cs.unc.edu/data/clear

• Image features: CLIP-ViT-B-32-views.tsv for CLIP-ViT/32 features, and ResNet-152-imagenet.tsv for ResNet features.

• Training Data: rxr_SPLIT_guide_multi.jsonl, where SPLIT is train, val_seen, or val_unseen

• Data for detected objects in each view: objects_27.json

For simplicity, we also provide the learned language and visual representation:

• Language Representation: encoder_best_val_unseen_loss

• Visual Representation: visual_resnet_best_val_unseen_loss for ResNet features, and visual_clip_best_val_unseen_loss for CLIP features

bash run/lang.bash 0

0 is the id of GPU.

bash run/visual.bash 0

0 is the id of GPU.

• objects_constraints: path to the file containing objects in each view.

bash run/agent.bash 0

0 is the id of GPU.

• dataset: RxR, R2R, CVDN

• load_encoder: path to learned language representation

• load_visual: path to learned visual representation

• feature_size: 2048 to use ResNet features, 512 to use CLIP-ViT/32 features

If you find this work useful, please consider citing:

@inproceedings{li2022clear,
  title     = {CLEAR: Improving Vision-Language Navigation with Cross-Lingual, Environment-Agnostic Representations},
  author    = {Jialu Li, Hao Tan, Mohit Bansal},
  booktitle = {Findings of NAACL},
  year      = {2022}
}

We thank the developers of EnvDrop, HAMT, CLIP for their public code release.