Rendezvous (RVS) is a corpus for executing navigation instructions and resolving spatial descriptions based on maps. The task is to follow geospatial instructions given in colloquial language based on a dense urban map.
The details of the corpus and task are described in: Where Do We Go from Here? Multi-scale Allocentric Relational Inference from Natural Spatial Descriptions.
Using RVS data we propose a large-scale augmentation method for generating high-quality synthetic data for new environments using readily available geospatial data. Our method constructs a grounded knowledge-graph, capturing entity relationships. We compare data generated data with LLMs and CFG.
The details of the ugmentation experiments are described in: Into the Unknown: Generating Geospatial Descriptions for New Environments.
The navigation instructions can be found in this repository - here or can be downloaded from HuggingFace - here
The map-graph can be found here
bazel-bin/rvs/model/text/model_trainer --processed_data_dir OUTDIR --train_region Manhattan --dev_region Manhattan_dev --test_region Manhattan_dev --output_dir OUTDIR
conda create --name rvs -y
conda activate rvs
python3 -m pip install -r requirements.txt
apt install bazel
or follow the instructions on BAZEL website
source build_all.sh
source test_all.sh
source run_all.sh