Papaers:

1. Rendezvous (RVS): A Novel Dataset for Spatial Allocentric Reasoning

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.

2. Into the Unknown: Generating Geospatial Descriptions for New Environments

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.

Data

The navigation instructions can be found in this repository - here or can be downloaded from HuggingFace - here

The map-graph can be found here

Run model

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

Installation:

Install conda environment

conda create --name rvs -y
conda activate rvs
python3 -m pip install -r requirements.txt

Install BAZEL

apt install bazel

or follow the instructions on BAZEL website

BAZEL: Build

source build_all.sh

BAZEL: Test

source test_all.sh

BAZEL: Run

source run_all.sh