This repository holds the developement code for generating symbolic planning domains over large 3D scene graphs from the iGibson database. Building upon the PDDLGym Planners, we provide access to a diversity of satisficing and optimal classical planners to test on our task planning domains, or custom environments created with our task samplers.
A note to current users: This repository is subject to major revisions by January 2022.
If you use Taskography in follow-up works, consider using the following publication:
@inproceedings{agia2021taskography,
title={Taskography: Evaluating robot task planning over large 3D scene graphs},
author={Agia, Christopher and Jatavallabhula, {Krishna Murthy} and Khodeir, Mohamed and Miksik, Ondrej and Vineet, Vibhav and Mukadam, Mustafa and Paull, Liam and Shkurti, Florian},
booktitle={5th Annual Conference on Robot Learning},
year={2021}
}
Start by navigating to your projects folder, then enter the following commands to create a python virtual environment and install the necessary packages.
cd path/to/projects/folder
python3 -m venv taskography; source taskography/bin/activate
# install pddlgym as editable
git clone https://github.com/taskography/pddlgym.git
pip install -e pddlgym
# install code dependecies
git clone --recurse-submodules https://github.com/taskography/3dscenegraph-dev
pip install -r 3dscenegraph-dec/scenegraph/requirements.txt
Download the tiny (35 scenes) and medium (105 scenes) splits from the iGibson 3D Scene Graph Dataset and unpack it into the directory of choice. Unzip the files as shown below:
cd path/to/data/folder
unzip 3DSceneGraph_tiny.zip; mv data tiny
unzip 3DSceneGraph_medium.zip; mv 3DSceneGraph_medium medium
The cloned PDDLGym repository should contain all domains evaluated in the paper. Our results can be reproduced by executing the following provided bash scripts:
./run_optimal_planners_1.sh
./run_optimal_planners_2.sh
./run_optimal_planners_3.sh
./run_optimal_planners_4.sh
./run_satisficing_planners_1.sh
./run_satisficing_planners_2.sh
./run_satisficing_planners_3.sh
./run_satisficing_planners_4.sh
However, if you wish to run any of the supported planners on a particular domain, scenegraph/plan.py the appropriate entry point. Its usage, not including all optional arguments, is as follows:
python plan.py --domain-name <pddlgym_domain_name> --planner <planner_name> [--log-dir <log-dir>] [--expid <expid>] [--timeout <timeout>]
Our hand-crafted PDDL domain files can be found in scenegraph/domain. The task categories corresponding to each domain file is:
- Rearrangement(k) = taskographyv2.pddl
- Courier(n, k) = taskographyv3.pddl
- Lifted Rearrangement(k) = taskographyv4.pddl
- Lifted Courier(n, k) = taskographyv5.pddl
The provided task samplers leverage these domain files to create PDDL problem files as per the inputted preferences. As an example, the following command creates a Rearrangment(k=10) dataset over the iGibson 3D Scene Graph tiny split with 100 training tasks generated over 10 training scenes and 5 test tasks for each of the 25 remaining test scenes:
python --data-root /path/to/data/folder/ --data-split tiny --domain domain/taskographyv2.pddl --output-dir path/to/pddlgm/repo/pddl/ --train-scenes 10 --train-samples 10 --test-samples 5 --task-length 10 [--seed <seed>]
Additional arguments may be required for the Courier domains (taskographyv3.pddl and taskographyv5.pddl) to specify the bagslot capacity for the agent.
This repository includes code from several open source repositories. We would like to acknowledge the authors for there efforts that made this project achievable.
- PDDLGym: pddlgym, Silver & Chitnis
- PDDLGym Planners: pddlgym_planners
- Tarski: tarski
- adl2strips: adl2strips