The Touchdown Dataset

Touchdown is a corpus for executing navigation instructions and resolving spatial descriptions in visual real-world environments. The task is to follow instruction to a goal position and there find a hidden object, Touchdown the bear.

The details of the corpus and task are described in: Touchdown: Natural Language Navigation and Spatial Reasoning in Visual Street Environments. Howard Chen, Alane Suhr, Dipendra Misra, Noah Snavely, and Yoav Artzi.

Paper: https://arxiv.org/abs/1811.12354

A short video explaining the task and showing an example: https://youtu.be/uCcDbTZs3v4

The code for the original navigation expriments is available here: https://github.com/lil-lab/ciff

Data

This repository contains the Touchdown corpus. The navigation environment is made of a large number of panoramas. To download the panoramas, please use the StreetLearn environment. You can request access to the panoramic images by filling out the form in StreetLearn Dataset. More details are here.

Starting example

The example runs a random policy with dummy image features in the environment.

python3 navigator.py

Structure of directory

data/: includes JSON files train.json, dev.json, test.json. These are the data files for navigation and spatial description resolution (SDR) tasks.
graph/: includes .txt files for constructing the graph. nodes.txt contains all nodes in the graph. links.txt contains all edges in the graph.

Graph

The script graph_loader.py loads the graph with the following two files, and base_navigator.py uses it to initialize the graph.

nodes.txt: has 4 columns panoid, pano_yaw_angle, latitude, and longitude
links.txt: has 3 columns start_panoid, heading, and end_panoid

JSON files

The JSON files contain both data for the navigation task and the SDR task. All three files follow the same structure described as follows.

Route information

city: city name
route_id: unique route id
elapsed: time spent on writing instructions for this route
failure_stats: number of attempts the instrcution writer took to place Touchdown the bear at the final position/panorama
num_finished: number of runs for followers to find the bear
full_text: full instructions from navigation to Touchdown bear placement

Navigation task

navigation_text: instruction text for navigation
route_panoids: a list of panorama ids of the route from start to end
start_heading: start heading angle in degrees
end_heading: end heading angle in degrees

Spatial Description Resolution (SDR) task

td_location_text: instruction text for SDR
pre_pano, main_pano, post_pano: panorama ids, main_pano is the target position pano id where Touchdown is placed. per_pano and post_pano are the before and after target position panorama
pre_static_center, main_static_center, post_static_center: the click position {x: width_ratio, y: height_ratio} of where Touchdown is placed, {x: -1, y: -1} means Touchdown can't be found for the panorama

You can construct your Gaussian smoothed target from the *_center click positions or contact us for cached targets.

Experiments reproduction code

The Touchdown tasks are reproduced by Harsh et al (2020). For more details, please refer to this technical report and the VALAN codebase.

License

The Touchdown Dataset is licensed under a Creative Commons Attribution 4.0 International License.

You should have received a copy of the license along with this work. If not, see http://creativecommons.org/licenses/by/4.0/.

lil-lab / touchdown